Real-time high-resolution system for predicting Earth system models using satellite-borne consistent models

Abstract

This satellite system provides high-resolution real-time enhancement of low-resolution Earth System Model (ESM) simulation data, eliminating ground processing delays and providing immediate access to detailed information necessary for weather forecasting and disaster response. [Solution] The present invention provides a satellite system that directly mounts a generative machine learning algorithm using a consistency model (CM) onto a satellite to enhance low-resolution ESM data to high resolution in a single step. The system comprises a satellite-mounted high-performance computing module, a real-time observation sensor array, an onboard data processing unit, a CM implementation module, a dynamic scale adaptation system, an uncertainty quantification module, a direct distribution communication system, and a data assimilation engine. By integrating observation and high-resolution processing on the same platform, it achieves real-time capabilities and enables direct data distribution even to areas with limited communication infrastructure.

Description

Technical Field

[0001] The present invention relates to an Earth observation satellite system, and in particular to a satellite system that directly mounts generative machine learning using a Consistency Model (CM) on a satellite to upscale low-resolution Earth System Model (ESM) simulation data to high resolution in real time. More specifically, the present invention relates to a satellite system that integrates observation data acquisition and super-resolution processing on the satellite, eliminates the data transfer delay associated with ground processing, and thereby supports rapid and accurate decision-making in time-critical applications such as weather forecasting, disaster response, agricultural management, and water resource management. The present invention is particularly intended for applications in fields such as extreme weather event prediction and early warning, climate change impact assessment, agricultural productivity improvement, and water resource management optimization.

Background Art

[0002] High-resolution and accurate Earth System Model (ESM) simulation is indispensable for assessing the ecological and socio-economic impacts of climate change. In particular, climate variables such as precipitation have a significant impact on vegetation, crop yields, infrastructure, economy and the like. However, conventional ESMs have high computational costs, making it difficult to run simulations at sufficiently high resolutions. Current ESMs are typically operated at a spatial resolution of about 10 to 100 km, which is too coarse to assess the local impacts of climate change. This leads to large uncertainties, especially in the prediction of extreme precipitation events and the assessment of local climate change impacts.

[0003] To address this problem, machine learning-based downscaling (improving spatial resolution) techniques have been proposed. Generative deep learning methods such as Normalizing Flows, Generative Adversarial Networks (GANs), and Diffusion Models have shown promising results. However, these methods have challenges such as high computational cost, unstable training, and the need for retraining for specific ESMs. In particular, GANs suffer from training instability and mode collapse, and Normalizing Flows tend to produce images of low quality. Furthermore, these methods are often trained on specific ESMs, requiring retraining for new or updated ESMs, making them unsuitable for processing large ESM ensembles.

[0004] The study by Hess et al. (2025) proposes a method for upscaling low-resolution ESM simulations to high resolution using a generative machine learning algorithm called a consistent model (CM). This method is computationally more efficient than conventional stochastic differential equation (SDE)-based diffusion models and has the advantage of generating high-resolution data in a single step. It is also independent of specific ESMs and can upscale any ESM simulation to high resolution without retraining. Furthermore, it has been shown that uncertainty can be quantified through stochastic downscaling and that it can be appropriately generalized to future climate states without explicitly incorporating physical constraints.

[0005] CM is a type of diffusion model, but unlike conventional SDE diffusion models, it learns a self-consistent function f(x(t), t) = x(tmin) and satisfies the relationship f(x(t),t) = f(x(t'),t')∀t,t'∈[tmin,tmax], enabling generation in a single step. This reduces the network evaluation required by conventional SDE diffusion models from hundreds to thousands of times to just one, significantly improving computational efficiency. Furthermore, CM offers high controllability in the sampling process, and by adjusting the spatial scale to be preserved, it can improve small-scale spatial patterns while preserving large-scale spatial patterns of ESMs.

[0006] However, all existing approaches, including the method by Hess et al., assume that they will be executed on a terrestrial computer system, which presents the following challenges: 1. Data Transfer Delay: The process of transmitting satellite observation data to the ground, processing it, and then redistributing it involves considerable time delays. This delay is a significant problem, especially in time-constrained applications such as weather forecasting and disaster response. For example, in the case of flood prediction due to heavy rainfall, conventional approaches result in delays of several hours to half a day between observation and the distribution of prediction information, drastically reducing the time available for responses such as evacuation and floodgate operation. 2. Bandwidth constraints: Transmitting high-resolution observation data to the ground requires a large bandwidth, and processing data from multiple satellites simultaneously places a heavy load on the communication infrastructure. For example, transmitting high-resolution global precipitation observation data to the ground requires a bandwidth of several hundred Mbps to several Gbps, which may exceed the capabilities of current satellite communication systems. 3. Limitations on global access: In areas with limited communication infrastructure, access to high-resolution climate data is difficult, and the communication infrastructure itself may be damaged, especially during disasters. This creates a "digital divide," where the most vulnerable and most critical regions are left behind from high-quality climate information. 4. Data Assimilation Delay: Because the data assimilation process, which integrates observational data with ESM forecasts, takes place on the ground, it takes time for the latest observational information to be reflected in the forecast. This reduces the accuracy of forecasts, especially for rapidly developing weather systems (e.g., rapid development of tropical cyclones or sudden convective precipitation). 5. Energy Consumption: Transmitting large amounts of data to the ground and processing it on large-scale ground-based computer systems results in significant overall energy consumption. In particular, the combined energy required to transmit high-resolution observational data to the ground and the energy needed to process it in large-scale ground-based data centers could result in substantial greenhouse gas emissions. 6. Concentration of Computing Resources: In the ground-based processing approach, computing resources are concentrated in specific ground data centers, which can lead to processing bottlenecks. This can result in processing delays and reduced quality, especially when processing large amounts of data from multiple satellites simultaneously or when receiving requests from multiple users at the same time. 7. Operational Complexity: The complex process chain of satellite observation, data transfer, ground processing, and redistribution increases operational complexity and creates numerous points of failure. Failures or delays at each step can impact overall performance and reduce system reliability and availability.

[0007] To address these challenges, a new approach is needed that integrates observation and high-resolution processing on satellites to provide high-quality data in real time. In particular, the high computational efficiency and single-step generation capability of CMs are ideal characteristics for performing high-resolution processing within the limited computing resources of satellites. [Prior art documents] [Non-patent literature]

[0008] [Non-Patent Document 1] Hess, P., Aich, M., Pan, B. et al. Fast, scale adaptive and uncertainty aware downscaling of Earth system model fields with generative machine learning. Nat Mach Intell 7, 363 373 (2025). https: / / doi.org / 10.1038 / s42256 025 00980 5 [Overview of the Initiative] [Problems that the invention aims to solve]

[0009] The present invention aims to solve the following problems:

[0010] 1. Providing a satellite system that eliminates data transfer delays associated with ground processing and enables high-resolution ESM data in real time: Conventional ground processing approaches involve delays of several hours to several days in the process of transmitting satellite observation data to the ground, processing it, and then redistributing it. This invention aims to reduce this delay to several minutes to several tens of minutes, thereby significantly improving the effectiveness of early warning systems, particularly for extreme weather events such as heavy rainfall and floods. Specifically, the goal is to reduce the time from observation to warning issuance to less than half of the conventional time, thereby significantly increasing the time available for responses such as evacuation and floodgate operation. For example, in the case of precipitation forecasting during a typhoon approach,

[0011] 2. Providing a satellite system that integrates observational data and high-resolution processing on the same platform to improve the immediacy and accuracy of data assimilation: In conventional approaches, the data assimilation process, which integrates observational data and ESM forecasts, is performed on the ground, resulting in a delay in reflecting the latest observational information in the forecast. This invention aims to minimize the time gap between observation and forecast by performing real-time data assimilation on satellite, thereby improving forecast accuracy. Specifically, it aims to increase the frequency of data assimilation from the conventional every 6 hours to every 30 minutes to every hour, particularly to improve the forecast accuracy of rapidly developing weather systems. For example, it aims to improve the forecast accuracy of rapid development (rapid intensification) of tropical cyclones by 30 to 50%.

[0012] 3. Construction of a satellite system that provides real-time, high-resolution data streams for monitoring and predicting global weather and climate phenomena: Conventional approaches to providing global high-resolution climate data involve significant delays. This invention aims to enable global monitoring and prediction of weather and climate phenomena by providing a real-time, high-resolution data stream covering the entire globe using a constellation of multiple satellites. Specifically, the goal is to provide data with a spatial resolution of 0.75° × 0.9375° (approximately 75 to 100 km) at an update frequency of every 3 to 6 hours globally, and every 30 minutes to 1 hour in specific regions. This will enable real-time monitoring of the development and movement of large-scale weather systems (e.g., El Niño-Southern Oscillation, Madden-Julian Oscillation, blocking high-pressure systems, etc.) and prediction of their impacts.

[0013] 4. Realizing a satellite system that provides high-resolution climate data in real time, necessary for disaster response and emergency decision-making: Conventional approaches make it difficult to rapidly provide high-resolution climate data in the event of a disaster. This invention aims to provide high-resolution climate data necessary for disaster response and emergency decision-making in real time, thereby supporting damage mitigation and effective relief activities. Specifically, the goal is to provide high-resolution data to affected areas at an update frequency of every 10 to 30 minutes when a disaster occurs or is predicted to occur. For example, in the event of a flood, information such as predicted inundation extent, temporal changes in water levels, and spatial distribution of precipitation will be provided in real time to support the optimization of evacuation plans and the efficient distribution of relief supplies.

[0014] 5. Development of a satellite system capable of directly delivering high-resolution climate data to areas with limited communication infrastructure: Conventional approaches make it difficult to access high-resolution climate data in areas with limited communication infrastructure. This invention aims to eliminate the global disparity in access to climate information by developing a satellite system that can directly deliver high-resolution climate data to areas with limited communication infrastructure. Specifically, it aims to broadcast data in a format that can be received by standard receivers (e.g., receivers equipped with small parabolic or planar antennas), enabling agricultural workers, local governments, and disaster prevention agencies, in particular, in developing countries and remote areas, to access high-quality climate information. For example, it aims to directly deliver data to areas where access to high-quality climate information has traditionally been limited, such as sub-Saharan Africa and mountainous regions of Asia.

[0015] 6. Development of technologies to efficiently perform high-resolution processing within the limited computing and energy resources of satellites: Computational processing on satellites is subject to severe constraints such as weight, power, and thermal management. This invention aims to develop a technology that enables high-resolution processing within the limited computing and energy resources of a satellite by efficiently implementing and optimizing a consistent model (CM). Specifically, it aims to reduce computational load and energy consumption by two to three orders of magnitude compared to conventional SDE diffusion models by introducing technologies such as quantization and lightweighting of the CM, sparse computation, pipeline processing, and hardware acceleration. For example, it aims to generate high-resolution data of 0.75° × 0.9375° from low-resolution ESM data of 3° × 3.75° in about 0.1 seconds, while keeping power consumption to about 10 to 20W.

[0016] 7. Development of adaptive methods to ensure accuracy and reliability of high-resolution imaging under various weather and climatic conditions: Weather and climate conditions vary greatly depending on the region and season. This invention aims to develop an adaptive method that automatically adjusts the optimal spatial scale according to observation conditions in order to ensure the accuracy and reliability of high-resolution imaging under various weather and climate conditions. Specifically, it implements a function that analyzes the power spectral density (PSD) of observation data and ESM data in real time, identifies the wavenumber (spatial scale) at which the two PSDs intersect, and automatically selects this as the spatial scale to be maintained during high-resolution imaging. It also implements a function that dynamically adjusts the spatial scale according to cloud cover, atmospheric stability, precipitation intensity, topographic complexity, land cover heterogeneity, seasonal variations, etc. The goal is to achieve consistent, high-quality high-resolution imaging under various conditions.

[0017] 8. Implementation of a function to comprehensively assess the uncertainty of high-resolution data and provide probabilistic predictions: It is important to represent predictions as probability distributions, rather than deterministic single predictions. The present invention aims to implement a function that comprehensively evaluates the uncertainty of high-resolution data and provides probabilistic predictions. Specifically, a function is implemented that generates multiple high-resolution realizations from a single ESM output to evaluate sampling uncertainty, and performs comprehensive uncertainty evaluation that takes into account the uncertainty of observation sensors, spatiotemporal correlation structures, and differences in predictions from multiple ESMs. This enables probabilistic predictions such as "the probability of precipitation exceeding 50 mm / day at a specific location is 80%", and aims to support risk-based decision-making.

[0018] 9. Development of self-optimization functions that enable long-term operation and continuous improvement of satellite systems: Since satellite systems are difficult to repair or update after launch, self-optimization functions that enable long-term operation and continuous improvement are critical. The present invention aims to develop self-optimization functions that continuously evaluate the performance of the satellite system and automatically perform parameter adjustment and operation mode optimization. Specifically, a function is implemented that collects information such as prediction verification feedback, user feedback, and system performance feedback, and optimizes the operation of the system based thereon. This aims to provide high-quality services over a long period of 5 to 10 years.

[0019] 10. Provision of customizable data products tailored to the specific needs of different application fields (weather forecasting, disaster response, agricultural management, water resource management, etc.): Different application fields require different types of climate information, temporal and spatial resolutions, update frequencies, and the like. The present invention aims to provide customizable data products tailored to the specific needs of different application fields. Specifically, a function is implemented to provide data products optimized for each application field, such as high-temporal-resolution short-term forecasts for weather forecasting, probabilistic forecasts of extreme phenomena for disaster response, seasonal forecasts and soil moisture information for agricultural management, and watershed-scale water balance forecasts for water resource management. This aims to improve the quality and efficiency of decision-making in each application field. [MEANS FOR SOLVING THE PROBLEMS]

[0020] The present invention provides a satellite system that directly mounts a generative machine learning algorithm based on a consistency model (CM) on a satellite, and performs real-time high-resolution enhancement on low-resolution ESM data. The satellite system of the present invention integrates observation and high-resolution processing on the same platform, thereby eliminating the delay associated with ground processing and providing high-quality data in real time.

[0021] The satellite system of the present invention consists of the following main components:

[0022] 1. Satellite-mounted high-performance computing module: Provides low-power consumption, high-performance computing hardware optimized for the space environment. Specifically, a hybrid architecture combining a dedicated radiation-hardened machine learning processor and an energy-efficient FPGA (Field-Programmable Gate Array) is adopted. This module provides a computing infrastructure for efficiently executing the consistency model (CM) within the limited power resources of the satellite. The dedicated radiation-hardened machine learning processor has resistance to radiation (especially high-energy particle beams) in the space environment, and is optimized for the execution of machine learning algorithms. Specifically, a tensor processing unit (TPU) using radiation-hardened silicon-on-insulator (SOI) technology is adopted, which has the capability to execute basic machine learning operations such as matrix operations, convolution operations, and activation functions at high speed and efficiently. The computing performance is assumed to be about 10 to 20 TFLOPS (tera floating-point operations per second), and the power consumption is assumed to be about 50 to 100 W. Energy-efficient FPGAs are reconfigurable hardware used to accelerate specific computations. Specifically, certain computations of a computer program (CM) (e.g., the encoder and decoder parts of a U-Net) are offloaded to the FPGA to speed up processing. FPGAs are manufactured using a special radiation-resistant process and are resistant to single-event upsets (SEUs). Computational performance is expected to be around 5 to 10 TFLOPS, and power consumption around 30 to 60W. This hybrid architecture can provide the computational power necessary for efficient execution of CMs while addressing the stringent constraints of the space environment (radiation, power, heat, etc.). Furthermore, by utilizing the reconfiguration capabilities of FPGAs, algorithm updates and optimizations can be performed in orbit, allowing for the adoption of the latest technologies over long periods.

[0023] 2. Real-time observation sensor array: The system provides a sensor array that observes the Earth across multiple wavelength bands. Specifically, it integrates high-resolution visible light and near-infrared imagers, microwave radiometers, rainfall radar, infrared sounders, and scattermeters. These sensors provide high-resolution observational data used as training data for the CM (Common Microwave Observatory) and also function as input for real-time data assimilation. The high-resolution visible light / near-infrared imager is a sensor for observing cloud patterns and surface characteristics, employing a CMOS image sensor with a spatial resolution of approximately 100 to 250 m, a spectral resolution of approximately 10 to 20 nm, and a dynamic range of approximately 12 to 14 bits. This sensor will perform observations in multiple bands in the visible light range (approximately 400 to 700 nm) and the near-infrared range (approximately 700 to 1100 nm) to acquire information such as cloud distribution, surface vegetation, and land cover. The sensor's field of view is expected to be approximately 10 to 15 degrees, and its scanning range is expected to be approximately 1000 to 1500 km. The microwave radiometer is a sensor used to observe water vapor and precipitation in the atmosphere. It employs a microwave radiometer with a frequency band of approximately 10 to 90 GHz, a spatial resolution of approximately 5 to 25 km, and a temperature resolution of approximately 0.3 to 0.5 K. This sensor acquires information on atmospheric water vapor, cloud water, and precipitation, and is particularly effective in estimating precipitation in cloud-covered areas. The sensor's field of view is expected to be approximately 45 to 60 degrees, and its scanning range approximately 1500 to 2000 km. The rainfall radar is a sensor for observing three-dimensional precipitation structures. It employs a dual-frequency rainfall radar with frequency bands of approximately 13.6 GHz (Ku band) and 35.5 GHz (Ka band), a spatial resolution of approximately 2 to 5 km, and a sensitivity of approximately 15 to 20 dBZ. This sensor acquires information such as the three-dimensional structure of precipitation, precipitation intensity, and precipitation type, and is particularly effective for observing extreme precipitation events such as torrential rains and typhoons. The sensor's scanning range is expected to be approximately 200 to 300 km. The infrared sounder is a sensor for observing atmospheric temperature and humidity profiles. It employs a high-resolution infrared sounder with a wavelength range of approximately 3.7 to 15.4 μm, a spectral resolution of approximately 0.5 to 1.0 cm⁻¹, and a temperature resolution of approximately 0.1 to 0.3 K. This sensor acquires information such as atmospheric temperature profiles, humidity profiles, and cloud top altitude, making it effective for understanding the thermodynamic state of the atmosphere. The sensor's field of view is expected to be approximately 30 to 45 degrees, and its scanning range approximately 1000 to 1500 km. The scattermeter is a sensor for observing sea surface winds, employing a frequency band of approximately 5.3 GHz (C-band), a spatial resolution of approximately 25 km, and a wind speed accuracy of approximately 1 to 2 m / s. This sensor acquires information on wind direction and speed at sea and is particularly effective in understanding the structure of meteorological systems such as typhoons and low-pressure systems. The sensor's field of view is expected to be approximately 45 to 60 degrees, and its scanning range is expected to be approximately 1500 to 2000 km. By integrating these diverse sensors, comprehensive observations of the atmosphere and Earth's surface can be achieved, providing the data necessary for monitoring and predicting various weather and climate phenomena. In particular, observations across a wide wavelength range from visible light to radio waves allow for the acquisition of diverse information such as clouds, precipitation, water vapor, and surface characteristics. Furthermore, satellite constellations increase the frequency of observations at the same location, enabling the capture of temporal changes.

[0024] 3. Onboard data processing unit: Preprocessing of observational data and ESM data is performed on the satellite. Specifically, this includes calibration and correction of sensor data, cloud masking and quality control, spatial interpolation and low-pass filtering of ESM data, bias correction using quantile delta mapping (QDM), and data normalization and transformation. While these processes were conventionally performed on the ground, this invention minimizes data processing delays by performing them in real time on the satellite. Sensor data calibration and correction are processes that convert raw data from each sensor into physical quantities and correct for noise and distortion caused by the sensor's characteristics. Specifically, this includes calibration of radiometric measurements (dark current correction, nonlinearity correction, etc.), geometric correction (correction based on satellite attitude and position), and atmospheric correction (removal of the effects of absorption and scattering by the atmosphere). These processes are performed using algorithms optimized for each sensor. For example, in the case of visible light and near-infrared imagers, dark current correction, flat field correction, geometric correction, and atmospheric correction are applied. In the case of microwave radiometers, antenna pattern correction, nonlinearity correction, and temperature calibration are applied. Cloud masking and quality control are processes that identify the influence of clouds from observational data and evaluate the quality of the data. Specifically, cloud detection algorithms are applied to visible light and near-infrared imager data to identify areas covered by clouds. In addition, quality control algorithms are applied to the data from each sensor to detect outliers, evaluate noise, and assess data reliability. These processes ensure the quality of the data used in subsequent processing. Spatial interpolation and low-pass filtering of ESM data is a process that interpolates low-resolution ESM data received from the ground into a high-resolution grid and removes small artifacts resulting from the interpolation. Specifically, ESM data is interpolated into a high-resolution grid using methods such as bilinear interpolation or cubic spline interpolation, and then low-pass filters such as Gaussian filters are applied to remove artifacts. These processes prepare ESM data suitable as input for CM (Compression Mechanism). Bias correction using quantile delta mapping (QDM) is a process for removing distribution bias from ESM simulations. Specifically, it compares the distribution of ESM data with that of observational data and performs bias correction for each quantile. This process uses approximately 500 quantiles to perform detailed bias correction, bringing the statistical characteristics of the ESM data closer to those of the observational data. This allows for the preparation of statistically appropriate ESM data as input for CM. Data normalization and transformation are processes that transform data into an appropriate range and optimize it as input for a computational model (CM). Specifically, a logarithmic transformation (x^~ = log(x + E) - log[E], E = 0.0001) is applied, followed by normalization to a range of approximately [-1, 1]. These transformations adjust the data distribution and improve the efficiency of CM training and inference. These preprocessing steps are efficiently performed using technologies such as parallel processing, pipelining, and hardware acceleration, reducing the time from sensor data reception to preprocessing completion from a few seconds to a few minutes. This allows for the timely provision of data necessary for real-time data assimilation and CM execution.

[0025] 4. Satellite-mounted CM implementation module: This provides an optimized implementation for running the CM algorithm proposed by Hess et al. on satellites. Specifically, it introduces optimization techniques such as quantization and optimization of model parameters, sparse computation, pipeline processing, and hardware acceleration. This makes it possible to efficiently execute CM within the limited computing resources of a satellite. Model parameter quantization and optimization are techniques that reduce memory usage and computational complexity by quantizing model parameters. Specifically, they involve techniques such as quantizing model parameters from 32-bit floating-point numbers to 8-bit integers, pruning (removing unnecessary connections), and knowledge distillation (transferring knowledge from a large model to a smaller model). This can reduce model size by approximately 75 to 90% and improve inference speed by approximately 3 to 5 times. For example, if the original CM model has approximately 27M parameters, quantization and optimization can reduce it to approximately 3 to 7M parameters. Sparse computation is a technique that improves energy efficiency by omitting unnecessary calculations. Specifically, it applies techniques such as dynamic pruning (omitting calculations that are not important at runtime), conditional computation (selecting a computation path based on the input), and early exit (terminating the computation when sufficient accuracy is obtained from the output of the hidden layer). This can reduce the amount of computation by approximately 50 to 70% and improve energy efficiency. Pipelining is a technique that optimizes data flow and improves throughput. Specifically, it applies techniques such as pipeline processing, which executes each layer of a model in parallel; batch processing (processing multiple inputs simultaneously); and memory access optimization (efficient use of cache). This can improve throughput by approximately 2 to 4 times. Hardware acceleration is a technique that speeds up specific operations using FPGAs. Specifically, it offloads computationally intensive operations such as convolution, matrix multiplication, and activation functions to the FPGA and executes them at high speed using dedicated hardware. This can improve the execution speed of these operations by approximately 5 to 10 times. These optimization techniques give the CM implementation features high-speed inference (generating high-resolution data in a single step), low memory usage (reducing memory usage during inference to approximately 1 to 2 GB), low power consumption (reducing power consumption during inference to approximately 10 to 20 W), high adaptability (adapting to various input conditions), and robustness (tolerance for input data quality issues and calculation errors). This allows for efficient high-resolution processing within the limited computing and energy resources of the satellite.

[0026] 5. Dynamic scaling adaptive system: The system provides a function to automatically adjust the optimal spatial scale according to observation conditions. Specifically, it implements functions such as real-time PSD analysis, adaptive scale selection, regional optimization, and seasonal variation handling. This makes it possible to obtain optimal high-resolution results under various observation conditions. Real-time PSD analysis is a process that analyzes the power spectral density (PSD) of observational data and ESM data in real time. Specifically, it uses a two-dimensional Fourier transform to convert spatial domain data to the wavenumber domain and calculates the power for each wavenumber. This analysis helps to understand the spatial scale characteristics of observational data and ESM data and to identify the wavenumber (spatial scale) at which the PSDs of both data intersect. This intersection represents the smallest spatial scale at which the ESM data is reliably supported, and it becomes a natural choice for the spatial scale to be retained when increasing resolution. Adaptive scale selection is a process of dynamically selecting the optimal spatial scale according to observation conditions (cloud cover, atmospheric conditions, etc.). Specifically, the spatial scale is selected by considering factors such as cloud cover (in cloud-covered areas, the reliability of visible light and near-infrared observations decreases, so a larger spatial scale is selected), atmospheric stability (in areas with an unstable atmosphere, small-scale phenomena become important, so a smaller spatial scale is selected), precipitation intensity (in areas with heavy precipitation, the spatial distribution of precipitation becomes complex, so a smaller spatial scale is selected), and the quality of the observation sensor (the spatial scale is adjusted according to the quality of the sensor data). Regional optimization is the process of optimizing the scale according to the characteristics of the region (topography, land cover, etc.). Specifically, the scale is optimized by considering regional characteristics such as the complexity of the topography (in areas with complex topography such as mountainous regions, topographic effects are important, so a smaller spatial scale is selected), the heterogeneity of land cover (in areas with heterogeneous land cover such as urban areas or areas where farmland and forests are mixed, local effects are important, so a smaller spatial scale is selected), the distribution of land and sea (in areas where land and sea are mixed, such as coastal areas, land and sea contrast is important, so a smaller spatial scale is selected), and the climate zone (the scale is adjusted according to the climate zone, such as tropical, mid-latitude, and polar regions). Seasonal variation adaptation is the process of adjusting the scale in accordance with seasonal changes in weather patterns. Specifically, the scale is adjusted considering seasonal variations such as monsoons (in areas where precipitation patterns change significantly between monsoon and non-monsoon seasons, the scale is adjusted according to the season), snow and ice cover (in areas covered with snow and ice in winter, the scale is adjusted according to the presence or absence of snow and ice cover), vegetation changes (the scale is adjusted according to seasonal changes in vegetation), and solar radiation conditions (the scale is adjusted according to seasonal changes in solar radiation conditions). These features give the dynamic scale adaptation system characteristics such as adaptability (adapting to various observation conditions, regional characteristics, and seasonal variations), automation (fully automated scale selection process), real-time performance (responding to changes in observation conditions in real time), reflection of regional characteristics (applying different scales to each region), and self-optimization (continuous evaluation and optimization of system performance through a feedback mechanism). This ensures accuracy and reliability of high-resolution imaging under various conditions.

[0027] 6. Multimodal Uncertainty Quantification Module: It provides uncertainty assessment by integrating multiple information sources. Specifically, it implements functions such as ensemble generation, sensor uncertainty integration, spatiotemporal correlation analysis, multi-model integration, and confidence interval mapping. This makes it possible to comprehensively assess the uncertainty of high-resolution data and provide it to decision-makers. Ensemble generation is a process that generates multiple high-resolution realizations from a single ESM output and evaluates the uncertainty of sampling. Specifically, it involves running CM multiple times on the same ESM output using different random number seeds, and constructing an ensemble from the multiple high-resolution realizations generated. By calculating the statistical properties of this ensemble (mean, standard deviation, quantiles, etc.), the intrinsic uncertainty of the high-resolution process can be quantified. For example, it is possible to generate 100 to 1000 ensemble members and estimate the probability distribution of precipitation for each grid cell. Sensor uncertainty integration is a process that considers the uncertainties of observation sensors during evaluation. Specifically, it integrates sensor uncertainties by considering factors such as sensor measurement errors (uncertainties arising from the measurement accuracy and calibration errors of each sensor), sampling errors (uncertainties arising from spatiotemporal sampling of observations), representativeness errors (errors when estimating grid average values ​​from point observations), and algorithmic errors (errors arising from the estimation algorithms for physical quantities). These uncertainty elements are represented by different models for each sensor and integrated using Bayesian statistical methods. Spatiotemporal correlation analysis is a process that analyzes the spatiotemporal correlation structure of uncertainty. Specifically, it analyzes correlation structures such as spatial correlation (correlation of uncertainty between different locations), temporal correlation (correlation of uncertainty between different time steps), and inter-variable correlation (correlation of uncertainty between different meteorological variables). These correlation structures are estimated based on statistical analysis or theoretical models of ensemble members. For example, spatial correlation is expressed using variational methods or Gaussian process models, while temporal correlation is expressed using autoregressive models or hidden Markov models. Multi-model integration is a process for assessing uncertainty by integrating predictions from multiple ESMs. Specifically, it involves receiving predictions from multiple ESMs, individually refining the predictions from each ESM, and integrating the refined results to construct a multi-model ensemble. This multi-model ensemble allows for the assessment of structural uncertainty based on differences between models. For example, by integrating predictions from multiple ESMs in CMIP6 (the 6th-order coupled model intercomparison project), it is possible to quantify uncertainty arising from differences between models. Confidence interval mapping is the process of visualizing spatial confidence intervals. Specifically, it visualizes information such as point estimates (e.g., ensemble mean), confidence intervals (e.g., 90% confidence interval), probability thresholds (e.g., the probability of exceeding a specific threshold), and the spatial distribution of uncertainty. This information is overlaid on a map to facilitate intuitive understanding. For example, in the case of precipitation forecasting, the mean values ​​can be displayed on a color map, and the width of the confidence interval can be displayed using transparency or contour lines. These features give the multimodal uncertainty quantification module characteristics such as inclusivity (a comprehensive assessment considering various sources of uncertainty), probabilistic representation (representing predictions as probability distributions), multimodal integration (integrating information from multiple sources), consideration of spatiotemporal correlation (considering the spatiotemporal correlation structure of uncertainty), and application orientation (considering the impact of uncertainty on downstream applications). This enables a comprehensive assessment of uncertainty in high-resolution data and supports risk-based decision-making.

[0028] 7. Direct distribution communication system: It provides direct data delivery capabilities to end users. Specifically, it implements features such as direct broadcasting, adaptive data compression, priority-based delivery, regional customization, and two-way communication. This makes it possible to directly deliver high-resolution climate data to users worldwide, including in areas with limited communication infrastructure. The direct broadcast function is a feature that broadcasts data in a format that can be received by standard receivers. Specifically, it broadcasts data using multiple frequency bands, such as L-band broadcasting (1.5 to 1.6 GHz), X-band broadcasting (7.8 to 8.4 GHz), and Ka-band broadcasting (25.5 to 27 GHz). These broadcasts comply with international standards (e.g., DVB-S2 and CCSDS standards) and are provided in a format that can be received by commercially available receivers. The data format also complies with international standards (e.g., NetCDF, GRIB, etc.) and is provided in a format that can be used with common weather data processing software. Adaptive data compression is a function that dynamically adjusts the data compression ratio according to the communication bandwidth. Specifically, it uses techniques such as hierarchical data structures (structuring data at multiple resolution levels and selecting the appropriate resolution according to the available bandwidth), adaptive quantization (adjusting the quantization level according to the importance of the data), context-adaptive entropy coding (selecting the optimal coding method according to the statistical characteristics of the data), and predictive coding (reducing data redundancy using temporal and spatial predictive models) to dynamically adjust the data compression ratio within a range of 10:1 to 100:1 according to the communication bandwidth. Priority-based delivery is a function that prioritizes the delivery of data with high urgency. Specifically, it sets data priorities and delivers them preferentially based on criteria such as the priority of extreme phenomena (data related to extreme phenomena such as heavy rain, strong winds, high temperatures, and low temperatures is delivered with the highest priority), regional vulnerability (data for areas vulnerable to specific hazards, such as areas with a high risk of flooding or landslides is delivered with a high priority), temporal urgency (data related to phenomena with a short predicted lead time is delivered with a high priority), and user requests (data based on requests from specific users is delivered with a priority according to the urgency of the request). Regional customization is a function that delivers data products tailored to the needs of each region. Specifically, it implements regional customization such as climate zone optimization (providing data products that focus on the most relevant weather variables and phenomena according to climate zones such as tropical, mid-latitude, and polar regions), industry optimization (providing data products according to the major industries of the region, such as regions where agriculture is the main industry, regions where fisheries are the main industry, and regions where tourism is the main industry), disaster risk optimization (providing data products according to the major disaster risks of the region, such as regions with a high risk of floods, regions with a high risk of droughts, and regions with a high risk of typhoons), and language and culture optimization (providing data representations and metadata that are appropriate for the local language and culture). Two-way communication is a function that delivers data in response to user requests. Specifically, it implements two-way communication functions such as request-based delivery (users request data on specific regions, times, and variables, and data is delivered accordingly), feedback mechanisms (user feedback is collected and used to improve the system), alert subscriptions (users subscribe to alerts based on specific conditions and receive notifications when those conditions are met), and data quality information (users request data quality information and receive information to evaluate the reliability of the data). These features give the direct delivery communication system global access (directly providing high-resolution climate data to users worldwide, including in areas with limited communication infrastructure), adaptive delivery (adaptively adjusting delivery content according to communication bandwidth, user needs, and data urgency), user-centric design (providing data products and interfaces tailored to user needs and capabilities), high reliability (maintaining stable communication under various conditions through redundancy and error correction), and bidirectionality (enabling bidirectional communication based on user requests and feedback). This enables the direct delivery of high-resolution climate data to users worldwide, including in areas with limited communication infrastructure.

[0029] 8. Real-time data assimilation engine: It provides the ability to instantly integrate observational data and ESM predictions. Specifically, it implements functions such as variational data assimilation, ensemble Kalman filtering, hybrid assimilation, multiscale assimilation, and non-Gaussian error handling. This makes it possible to optimally integrate observational data and ESM predictions and improve prediction accuracy. Variational data assimilation is a function that implements data assimilation using the four-dimensional variational method. Specifically, it performs variational data assimilation through processes such as defining a cost function (defining a cost function based on the difference between observations and model predictions, and deviations from the background field), calculating gradients (implementing an adjoint model to calculate the gradient of the cost function), optimization (minimizing the cost function using optimization algorithms such as the conjugate gradient method), and generating an analysis field (generating optimal initial conditions based on the optimization results). This variational data assimilation is particularly effective when continuous observational data is available and can generate a physically consistent analysis field. The ensemble Kalman filter is a function that implements ensemble-based data assimilation. Specifically, it performs ensemble forecasting (running ESM on multiple initial conditions to generate a forecast ensemble), estimation of error covariance (estimating the forecast error covariance from the variances of the ensemble members), calculation of Kalman gain (calculating the Kalman gain from the forecast error covariance and observation error covariance), updating the analysis field (updating the forecast field to more closely resemble the observation using the Kalman gain), and resampling the ensemble (resampling the ensemble to reflect the analysis error covariance). This ensemble Kalman filter is robust to the nonlinearity and non-Gaussianity of the model and can naturally represent the uncertainty of the forecast. Hybrid assimilation is a function that implements hybrid assimilation by combining variational methods and ensemble methods. Specifically, it performs hybrid assimilation through processes such as hybrid representation of background error covariance (combining static climatological covariance with flow-dependent covariance estimated from the ensemble), optimization by variational method (generating the analysis field using variational method with hybrid background error covariance), and ensemble update (updating ensemble members using the results of the variational method). This hybrid assimilation is a method that combines the physical consistency of variational methods with the flow-dependent error representation of ensemble methods, enabling highly accurate assimilation by leveraging the advantages of both. Multiscale assimilation is a function that implements multiscale assimilation, integrating information from different spatial scales. Specifically, it performs multiscale assimilation through processes such as scale decomposition (decomposing observations and model predictions into different spatial scales), scale-dependent assimilation (applying the optimal assimilation method to each scale), and scale integration (integrating assimilation results from different scales to generate the final analysis field). This multiscale assimilation can perform appropriate assimilation for phenomena at different spatial scales and can improve prediction accuracy, especially in areas with complex topography or uneven land cover. Non-Gaussian error handling is a feature that implements data assimilation that supports non-Gaussian error distributions. Specifically, it addresses non-Gaussian errors using methods such as particle filters (a non-parametric assimilation method based on Monte Carlo sampling), rank transformation (a method that transforms non-Gaussian distributions into Gaussian distributions before assimilation, and then performs the inverse transformation), and Gaussian mixture models (a method that approximates non-Gaussian distributions as a mixture of multiple Gaussian distributions). These methods can improve the assimilation accuracy, especially for variables with non-Gaussian distributions, such as precipitation. These features give the real-time data assimilation engine characteristics such as real-time performance (minimizing the time from receiving observational data to completing assimilation), high accuracy (optimally integrating observations and model predictions using state-of-the-art data assimilation techniques), adaptability (a flexible data assimilation system that adapts to various observation types, meteorological conditions, and regional characteristics), computational efficiency (efficiently performing data assimilation within limited computational resources), and uncertainty representation (explicitly representing the uncertainty of the assimilation results). This enables the immediate integration of observational data and ESM predictions, improving prediction accuracy.

[0030] The most innovative aspect of this invention lies in its ability to directly mount the CM algorithm on the satellite and integrate observation and data processing on the same platform, thereby eliminating delays associated with ground processing and achieving real-time performance. Furthermore, efficient CM implementation on the satellite enables high-resolution processing within limited computing and energy resources. In addition, the dynamic scale adaptation system and multimodal uncertainty quantification module ensure the accuracy and reliability of high-resolution processing under various weather and climatic conditions.

[0031] By integrating these components, the satellite system of the present invention can eliminate delays associated with ground processing and provide high-quality, high-resolution climate data in real time. This can improve the quality and efficiency of decision-making in fields such as weather forecasting, disaster response, agricultural management, and water resource management. [Effects of the Invention]

[0032] The satellite system of the present invention provides the following effects:

[0033] 1. Improved time efficiency through real-time processing: By eliminating data transfer delays associated with ground processing and performing high-resolution ESM data in real time, processing time is significantly reduced compared to conventional approaches. Specifically, processing that previously took several hours to several days with conventional ground processing approaches can be reduced to several minutes to several tens of minutes. This will greatly improve the effectiveness of early warning systems, especially for extreme weather events such as heavy rainfall and floods. For example, in the case of precipitation forecasting during a typhoon approach, conventional approaches resulted in a delay of approximately 6 to 12 hours between observation and distribution of forecast information, but this can be shortened to approximately 1 to 2 hours with the present invention. This reduction of 4 to 10 hours is extremely important for disaster prevention responses such as the implementation of evacuation plans and the operation of floodgates. For example, when heavy rainfall is predicted due to a typhoon, it usually takes about 4 to 6 hours from the issuance of evacuation orders to the actual completion of evacuation, and the time reduction according to the present invention greatly improves the feasibility of this evacuation process. Furthermore, in the case of flood prediction, conventional approaches only provided a few hours of buffer time between the issuance of a warning and the actual occurrence of a flood, whereas the approach of the present invention makes it possible to secure a buffer of 6 to 12 hours. This additional time is extremely important for implementing evacuation plans and preparing for emergency responses. For example, in areas with a high risk of flooding, it takes about 4 to 8 hours to open evacuation centers, implement traffic restrictions, and distribute relief supplies, and the increase in buffer time provided by the present invention significantly improves the effectiveness of these responses. Furthermore, in agricultural management, conventional approaches resulted in a delay of approximately one to two weeks between updating seasonal forecasts and providing information to farmers. This delay can be reduced to approximately one to two days with the present invention. This reduction in time allows farmers to respond quickly to changes in weather conditions and make timely decisions such as adjusting planting times, optimizing irrigation schedules, and implementing pest and disease control measures.

[0034] 2. Improving prediction accuracy through real-time data assimilation: By instantly integrating observational data and ESM predictions on satellite, the time gap between observation and prediction is minimized, improving prediction accuracy. Conventional approaches require considerable time to transmit observational data to the ground, perform data assimilation, and distribute updated predictions, making it difficult to reflect the latest observational information in predictions. The present invention's approach enables the immediate reflection of the latest observational information in predictions by performing real-time data assimilation on satellite. This significantly improves the accuracy of predicting meteorological phenomena, especially those that develop rapidly. For example, it can improve the accuracy of predicting the rapid development (rapid intensification) of tropical cyclones by 30 to 50 percent. Rapid intensification of tropical cyclones is a phenomenon in which the central pressure drops by more than 30 hPa within 24 hours, and while it was difficult to predict with conventional approaches, the approach of the present invention can significantly improve the accuracy of predicting this phenomenon by immediately reflecting the latest observational information. As a result, disaster prevention agencies in coastal areas can formulate and implement evacuation plans based on more accurate information. Furthermore, the accuracy of predicting convective precipitation (such as thunderstorms and localized heavy rainfall) is also significantly improved. While conventional approaches had limited accuracy in predicting the location and intensity of convective precipitation, the present invention's approach, by immediately incorporating the latest observational information, can improve the prediction accuracy of convective precipitation by 20 to 40%. This can improve the quality of decision-making in areas such as urban flood control and traffic management. Furthermore, it is expected that the spatial agreement index (Spatial Agreement Index) for 24-hour precipitation forecasts will improve by 15 to 25 percent. This index indicates the spatial agreement between predicted precipitation patterns and actual precipitation patterns, and this improvement will enhance the accuracy of planning in areas such as water resource management and agricultural management.

[0035] 3. Strengthening emergency response: By providing real-time monitoring and forecasting of extreme weather events such as torrential rain and floods, this invention enhances disaster response and emergency decision-making. Conventional approaches have made it difficult to rapidly provide high-resolution climate data when a disaster occurs, and situation assessment tends to be delayed, especially in areas where communication infrastructure is damaged. The approach of this invention, by delivering high-resolution data directly from satellites, makes it possible to provide necessary information to disaster response agencies regardless of the state of communication infrastructure. This will help mitigate damage and support effective relief efforts. For example, improving the accuracy of flood inundation predictions during floods will enable the optimization of evacuation plans and the efficient distribution of relief supplies. Specifically, improving the accuracy of flood inundation predictions by 30 to 50 percent will allow for more precise identification of areas requiring evacuation and more effective allocation of limited relief resources. Furthermore, the risk assessment of landslides will also be strengthened. In conventional approaches, the accuracy of predicting the spatial distribution of precipitation was limited, resulting in high uncertainty in landslide risk assessment. The approach of the present invention makes it possible to improve the accuracy of landslide risk assessment by 20 to 40% by providing high-resolution precipitation forecasts. This makes it possible to identify areas with a high risk of landslides and issue appropriate warnings. Furthermore, the ability to understand the situation after a disaster is also enhanced. In conventional approaches, understanding the situation after a disaster relies mainly on reports from the ground, making it difficult to grasp the overall picture, especially in the case of widespread disasters. The approach of the present invention combines real-time observations from satellites with high-resolution data, making it possible to quickly and comprehensively understand the situation after a disaster. This enables prioritization of relief activities and optimization of resource allocation.

[0036] 4. Improved communication efficiency: By generating high-resolution data on satellite, the amount of communication with the ground is reduced, and communication efficiency is improved. Conventional approaches require transmitting high-resolution observation data to the ground, processing it, and then redistributing it, which necessitates a large communication bandwidth. The present invention's approach processes observation data on satellite and transmits only the necessary information to the ground, significantly reducing the amount of communication required. Specifically, it is expected that the amount of data transmitted will be reduced by 50 to 80% compared to conventional approaches. For example, when transmitting high-resolution global precipitation observation data to the ground, conventional approaches required a communication bandwidth of several hundred Mbps to several Gbps, but the approach of the present invention transmits only processed high-resolution data, so a communication bandwidth of several tens of Mbps is sufficient. This makes it possible to efficiently utilize limited communication bandwidth and provide more information. Furthermore, the reliability of communication is also improved. Conventional approaches require the transmission of large amounts of data, making them susceptible to communication failures and bandwidth limitations. The approach of the present invention transmits a small amount of data, making it less susceptible to communication failures and bandwidth limitations, thus enabling more reliable communication. This ensures that necessary information can be reliably provided, especially in areas with unstable communication environments or during disasters. Furthermore, communication costs are also reduced. Conventional approaches required expensive communication systems to transmit large amounts of data, but the approach of the present invention transmits less data, so a lower-cost communication system is sufficient. This reduces the overall system cost and enables the provision of services over a wider area.

[0037] 5. Expanding global access: By directly distributing high-resolution data to areas with limited communication infrastructure, this invention aims to eliminate disparities in access to global climate information. Conventional approaches have primarily limited access to high-resolution climate data available to meteorological agencies and research institutions in developed countries, leaving developing countries and remote areas without sufficient access. This invention's approach, by directly distributing data from satellites, makes it possible to provide high-resolution climate data to users worldwide, regardless of the state of communication infrastructure. This will enable agricultural workers, local governments, and disaster management agencies, particularly those in developing countries and remote areas vulnerable to the effects of climate change, to access high-quality climate information and take appropriate countermeasures. For example, in sub-Saharan Africa, access to high-quality climate information has traditionally been limited, but the approach of this invention will allow agricultural workers in this region to access high-quality seasonal forecasts, enabling them to optimize planting plans and water resource management. This is expected to improve agricultural productivity and strengthen the ability to adapt to climate change. Furthermore, this technology enables access to high-quality climate information even in areas with complex topography and limited communication infrastructure, such as mountainous regions in Asia. In these areas, local weather patterns due to topographic effects are important, and high-resolution climate information is particularly useful. The approach of this invention will allow residents and local governments in these areas to access high-quality climate information, enabling disaster prevention planning and agricultural management based on local weather patterns. Furthermore, access to high-quality climate information will become possible even in geographically isolated regions such as small island nations. These regions are strongly influenced by oceanic climates and are therefore greatly affected by climate change, making high-quality climate information particularly important. The approach of the present invention will enable residents and government agencies in these regions to access high-quality climate information and develop and implement climate change adaptation measures.

[0038] 6. Improved energy efficiency: By generating data in a single step, high-resolution processing is achieved within the limited power resources of the satellite, improving overall energy efficiency. Conventional diffusion model-based approaches require numerous iterative calculations to generate high-resolution data, consuming significant computational resources and energy. The approach of this invention, by generating CMs in a single step, makes it possible to drastically reduce computational load and energy consumption. Specifically, it is expected that the computational load and energy consumption will be reduced by two to three orders of magnitude compared to conventional SDE diffusion models. For example, when generating high-resolution data of 0.75° × 0.9375° from low-resolution ESM data of 3° × 3.75°, the conventional SDE diffusion model requires approximately 500 network evaluations, with a processing time of approximately 40 seconds and a power consumption of approximately 500 to 1000W. In the CM-based approach of the present invention, high-resolution data is generated in a single step, so the processing time can be reduced to approximately 0.1 seconds and the power consumption to approximately 10 to 20W. This makes it possible to perform high-resolution processing within the limited power resources of a satellite. Furthermore, overall energy consumption can be reduced compared to processing on large-scale ground-based computer systems. In conventional approaches, high-resolution processing is performed in large-scale ground-based data centers, requiring significant energy for power consumption and cooling of the data centers. In the present invention, high-resolution processing is performed on a satellite, reducing the load on ground-based data centers and lowering overall energy consumption. This reduces the environmental impact of the entire system and enables sustainable operation. Furthermore, the energy required for communication is also reduced. Conventional approaches required significant communication energy to send and receive large amounts of data. In the approach of the present invention, the amount of data sent and received is small, thus reducing the energy required for communication. This improves energy efficiency at both the satellite and the ground station.

[0039] 7. Improving accuracy and reliability through adaptive scale selection: By automatically adjusting the optimal spatial scale according to observation conditions, the accuracy and reliability of high-resolution imaging under various weather and climate conditions are improved. Conventional approaches generally involve fixed scale selection, making it difficult to adapt to various weather and climate conditions. In the approach of the present invention, a dynamic scale adaptation system makes it possible to dynamically select the optimal spatial scale according to observation conditions (cloud cover, atmospheric conditions, etc.), regional characteristics (topography, land cover, etc.), and seasonal variations. This improves the accuracy and reliability of high-resolution imaging under various conditions. For example, in mountainous regions with complex terrain, it becomes possible to select a smaller spatial scale and generate high-resolution data that reflects fine topographic effects. Specifically, by adjusting the spatial scale according to the complexity of the terrain, it is expected that the accuracy of precipitation forecasts in mountainous regions can be improved by 20 to 40 percent. This can improve the accuracy of risk assessments for floods and landslides in mountainous regions. Furthermore, in areas where convective precipitation occurs, adjusting the spatial scale according to atmospheric instability and precipitation intensity can improve the accuracy of predicting the spatial distribution of convective precipitation by 15 to 30%. This can improve the accuracy of assessing urban flood risk due to localized heavy rainfall. Furthermore, adjusting the spatial scale according to the season and region improves the accuracy of seasonal forecasts. For example, in regions where precipitation patterns change significantly between monsoon and non-monsoon seasons, adjusting the spatial scale according to the season can improve the accuracy of seasonal forecasts by 10 to 20 percent. This can improve the accuracy of agricultural management and water resource management planning.

[0040] 8. Enhancing decision support through comprehensive uncertainty assessment: This invention supports decision-makers' judgments by providing uncertainty assessments that integrate multiple information sources. Conventional approaches have limited uncertainty assessment capabilities, and in particular, integrating uncertainties from multiple information sources has been difficult. The present invention's approach, through a multimodal uncertainty quantification module, enables comprehensive uncertainty assessments using techniques such as ensemble generation, sensor uncertainty integration, spatiotemporal correlation analysis, and multi-model integration. This allows decision-makers to make appropriate judgments while taking uncertainty into account. For example, in the case of flood prediction, providing probabilistic predictions that take into account the uncertainty of the extent of flooding makes it possible to develop risk-based evacuation plans. Specifically, by providing probabilistic information such as "there is an 80% probability that flooding of 50 cm or more will occur at a particular location," it becomes possible to prioritize evacuations and optimize resource allocation. Furthermore, in agricultural management, providing probabilistic information that takes into account the uncertainty of seasonal forecasts makes it possible to formulate risk management strategies. For example, by providing probabilistic information such as "there is a 60% probability that rainfall for the next three months will be less than 80% of the average," appropriate preparations for drought risk (e.g., selecting drought-tolerant varieties, preparing irrigation facilities, etc.) can be made. Furthermore, uncertainty assessment that considers spatiotemporal correlations enhances the risk assessment of wide-area extreme events. For example, in assessing wide-area flood risk spanning multiple river basins, it is important to consider spatiotemporal correlations of precipitation. The approach of the present invention provides uncertainty assessment that considers spatiotemporal correlations, making it possible to improve the accuracy of wide-area flood risk assessment by 15 to 30%. This enables the formulation of wide-area disaster prevention plans and the optimization of resource allocation.

[0041] 9. Self-optimizing features that enable long-term operation and continuous improvement: By continuously evaluating the performance of the satellite system and automatically adjusting parameters and optimizing operating modes through a self-optimization function, long-term operation and continuous improvement are achieved. In conventional approaches, system optimization relies mainly on human intervention on the ground, and corrections after launch are particularly difficult for satellite systems. The approach of this invention collects information such as predictive verification feedback, user feedback, and system performance feedback, and implements a function that automatically optimizes the operation of the system based on this information. This makes it possible to provide high-quality services over a long period of 5 to 10 years. For example, prediction validation feedback allows for continuous adjustment of CM parameters and dynamic scale adaptation system parameters, maintaining and improving prediction accuracy. Specifically, by comparing predictions with actual observations and analyzing patterns of prediction errors, weaknesses in the system can be identified and improved. Furthermore, user feedback allows for continuous improvement of the content and format of data products. For example, based on user feedback, data products tailored to specific application areas (e.g., agricultural management or water resource management) can be developed and improved. This enables the provision of services that meet user needs and enhances the usefulness of the system. Furthermore, system performance feedback allows for the optimization of computing resource allocation and communication bandwidth usage. For example, by dynamically adjusting processing priorities and precision according to system load and power conditions, maximum performance can be achieved within limited resources. This helps maintain system stability and efficiency.

[0042] 10. Providing customizable data products tailored to different application areas: By providing customizable data products that meet the specific needs of different application fields (such as weather forecasting, disaster response, agricultural management, and water resource management), we aim to improve the quality and efficiency of decision-making in each field. Conventional approaches often provide general-purpose data products that do not adequately address the needs of specific application fields. The approach of this invention makes it possible to provide customizable data products that meet the specific needs of different application fields. For example, for weather forecasting, it can provide short-term forecasts with high temporal resolution. Specifically, by updating precipitation forecasts 1 to 6 hours in advance every 10 to 30 minutes and providing forecasts with a temporal resolution of about 30 minutes, the accuracy of short-term forecasts can be improved. This can improve the quality of decision-making in fields such as air traffic control, road traffic management, and outdoor event management. For disaster response, we can provide probability predictions of extreme weather events. Specifically, by providing predictions of the probability and intensity of extreme weather events such as heavy rain, strong winds, and storm surges, we can support the formulation and implementation of disaster prevention plans. For example, by providing probabilistic information such as "there is a 70% probability that the 3-hour rainfall at a specific location will exceed 100 mm," we can support the formulation of evacuation plans and decisions regarding floodgate operations. For agricultural management, we can provide seasonal forecasts and soil moisture information. Specifically, by providing forecasts of temperature and precipitation one to six months in advance, as well as information on current soil moisture levels, we can support agricultural management such as planting plans, irrigation schedules, and pest and disease control. For example, by combining a seasonal forecast such as "there is a high probability that precipitation for the next three months will be within the range of 80 to 120% of the average year" with current information such as "the current soil moisture level is around 90% of the average year," we can support decision-making by farmers. For water resource management, we can provide water balance forecasts at the basin level. Specifically, by providing forecasts of precipitation, evapotranspiration, and runoff for each basin, we can support water resource management such as dam operations, irrigation water allocation, and hydroelectric power generation planning. For example, by providing forecast information such as "average precipitation in the basin over the next month is expected to be around 110 to 130% of the normal level," we can support the optimization of dam operations and preparations for flood control measures.

[0043] These effects will enable faster and more accurate decision-making in areas such as weather forecasting, disaster response, agricultural management, and water resource management, thereby improving society's ability to adapt to the impacts of climate change. In particular, by significantly improving access to high-quality climate information in areas with limited communication infrastructure and in developing countries, it will help bridge the global gap in access to climate information and contribute to achieving the Sustainable Development Goals (SDGs). [Modes for carrying out the invention]

[0044] Embodiments of the present invention will be described in detail below. ### Overall configuration of the satellite system The satellite system of the present invention is implemented as a constellation consisting of multiple satellites positioned in low Earth orbit (LEO). Specifically, it comprises a constellation of 12 to 24 satellites positioned in polar orbits at an altitude of approximately 500 to 800 km, providing a real-time, high-resolution data stream covering the entire Earth. The satellite constellation has the following orbital parameters: - Orbital altitude: Approximately 500 to 800 km (altitude to ensure optimal observation and communication conditions) - Orbital inclination: Approximately 98 degrees (the inclination angle required to achieve a sun-synchronous orbit) - Number of orbital planes: 4 to 6 planes (number of orbital planes to ensure global coverage) - Number of satellites in each orbital plane: 3 to 4 (number of satellites to ensure observation frequency) - Orbital period: Approximately 95 to 100 minutes (typical period for low Earth orbit) This orbital configuration enables observations approximately every 3 to 6 hours globally, and every 30 minutes to 1 hour in specific regions. Furthermore, by employing a sun-synchronous orbit, observations at each location can always be performed at the same local time, ensuring consistency in time-series data. Each satellite will carry the following main components: 1. Satellite-mounted high-performance computing module 2. Real-time observation sensor array 3. Onboard Data Processing Unit 4. Satellite-mounted CM implementation module 5. Dynamic Scale Adaptation System 6. Multimodal Uncertainty Quantification Module 7. Direct Distribution Communication System 8. Real-time data assimilation engine These components will be integrated with the satellite's bus system (power system, attitude control system, thermal control system, communications system, etc.) to function as a single integrated system. The total weight of the satellite is expected to be approximately 500 to 800 kg, and its power consumption is expected to be approximately 1.5 to 2.5 kW. The satellite measures approximately 2m x 2m x 3m and is launched in a folded state on the launch vehicle. After orbital insertion, the solar panels and observation sensor antennas are deployed, and the satellite becomes operational. The solar panels have an area of ​​approximately 10 to 15 m² and generate approximately 2 to 3 kW of power. This power is used to operate observation sensors, computing modules, communication systems, and other equipment. The satellite's thermal control system employs a comprehensive thermal management system combining heat pipes, heat dissipation fins, multi-layer insulation, and heaters to cope with the harsh temperature conditions of the space environment (fluctuating between -100°C and +100°C). In particular, thermal management to efficiently dissipate heat from the high-performance computing module is crucial, and by providing a dedicated heat dissipation path, the operating temperature of the computing module is maintained within an appropriate range (approximately -20 to +60°C). The satellite's attitude control system employs a three-axis stabilization method to achieve high-precision observation and communication. Specifically, it uses an attitude control system that combines reaction wheels, magnetic torquers, star trackers, solar sensors, and Earth sensors, achieving an attitude determination accuracy of approximately 0.01 degrees and an attitude control accuracy of approximately 0.1 degrees. This enables highly accurate directional control of observation sensors and precise directional control of communication antennas. A satellite constellation forms an end-to-end system by coordinating with a ground station network, data processing center, and user interface. The ground station network consists of multiple ground stations located around the world, responsible for communication with the satellite, sending commands, and receiving telemetry. The data processing center is responsible for storing, analyzing, and distributing the data received from the satellite. The user interface provides end users with tools to access, visualize, and analyze the data. Intersatellite communication links allow for the direct exchange of data and commands between satellites within a constellation. This enables continued operation even outside the visible range of ground stations, improving the overall flexibility and redundancy of the system. Specifically, an intersatellite communication link in the Ka band (approximately 26 GHz) is employed, achieving a data rate of approximately 100 Mbps. The operational lifespan of the satellites is expected to be approximately 5 to 7 years. By sequentially replacing satellites within this period, continuous operation of the entire system will be ensured. Furthermore, on-orbit software updates will allow for the introduction of new algorithms and functions, enabling continuous improvement of the system. ### Satellite-mounted high-performance computing module The satellite-mounted high-performance computing module provides low-power, high-performance computing hardware optimized for the space environment. This module provides a computing foundation for efficiently running consistent models (CMs) within the limited power resources of a satellite. Specifically, it consists of the following hardware components: 1. A dedicated machine learning processor with radiation resistance: A dedicated processor optimized for executing machine learning algorithms, possessing resistance to radiation (especially high-energy particle beams) in the space environment. Specifically, it employs a tensor processing unit (TPU) using radiation-resistant silicon-on-insulator (SOI) technology. This processor has the ability to perform basic machine learning operations such as matrix operations, convolution operations, and activation functions at high speed and efficiency. SOI technology prevents charge accumulation and leakage due to radiation by placing an insulating layer (usually silicon dioxide) beneath the silicon layer, improving resistance to radiation effects such as single-event latch-up (SEL) and single-event function interruption (SEFI). Furthermore, by combining it with technologies such as triple modular redundancy (TMR) and error detection and correction (EDAC) circuits, it also improves resistance to single-event upset (SEU). This TPU is optimized for 8-bit integer operations and can efficiently execute quantized machine learning models. It is expected to have a computational performance of approximately 10 to 20 TFLOPS (tera floating-point operations per second) and a power consumption of approximately 50 to 100W. It also features on-chip memory (approximately 32 to 64MB) to reduce memory access latency and energy consumption. This processor has the following features: - Radiation resistance: Total ionization dose (TID) resistance approximately 100 krad (Si), single event resistance approximately 100 MeV-cm² / mg - Low power consumption: Computation efficiency of approximately 100 to 200 GFLOPS / W - High performance: 8-bit integer arithmetic performance of approximately 10 to 20 TFLOPS - Flexibility: Supports various machine learning operations (matrix multiplication, convolution, activation functions, etc.) - Reconfigurable: Some computing units are reconfigurable and can be optimized for different algorithms. 2. Energy-efficient FPGAs: By using FPGAs, which are reconfigurable hardware, specific calculations are accelerated. Specifically, certain calculations of the CM (for example, the encoder and decoder parts of U-Net) are offloaded to the FPGA to speed up processing. FPGAs are manufactured using a special radiation-resistant manufacturing process and are resistant to single-event upsets (SEUs). Radiation-tolerant FPGAs employ special manufacturing processes and circuit designs compared to standard FPGAs. Specifically, they utilize techniques such as epitaxial layers, trench isolation, and redundant circuits to improve radiation tolerance. They also implement periodic scrubbing (reloading) of the configuration memory to mitigate the effects of bit inversion caused by radiation. This FPGA has the following features: - Radiation resistance: Total ionization dose (TID) resistance approximately 100 krad (Si), single event resistance approximately 100 MeV-cm² / mg - Low power consumption: Computation efficiency of approximately 50 to 100 GFLOPS / W - High performance: 8-bit integer arithmetic performance of approximately 5 to 10 TFLOPS - Reconfigurable: Hardware can be reconfigured in orbit. - Dedicated circuits: Dedicated circuits optimized for specific operations (e.g., convolution, pooling, upsampling, etc.) can be implemented. The expected computing performance is approximately 5 to 10 TFLOPS, and power consumption is approximately 30 to 60W. 3. High-speed, large-capacity memory: High-speed, high-capacity memory for storing data and model parameters necessary for CM execution. Specifically, it employs radiation-resistant DDR4 memory with error correction capabilities. This memory has the following characteristics: - Radiation resistance: Error detection and correction (EDAC) circuitry automatically corrects single-bit errors and detects multi-bit errors. - Large capacity: Approximately 64 to 128GB - High speed: Bandwidth approximately 100 to 200 GB / s - Low power consumption: Approximately 10 to 20W The expected memory capacity is approximately 64 to 128 GB, and the bandwidth is approximately 100 to 200 GB / s. 4. Non-volatile storage: Non-volatile storage for long-term storage of observational data, ESM data, CM model parameters, etc. Specifically, it employs radiation-resistant 3D NAND flash memory. This storage has the following features: - Radiation resistance: Error detection and correction (EDAC) circuitry and redundant storage mitigate the effects of radiation. - Large capacity: Approximately 1 to 2TB - High speed: Read / write speeds of approximately 1 to 2 GB / s - Low power consumption: Approximately 1 to 2W in idle mode, approximately 5 to 10W in active mode. The storage capacity is expected to be approximately 1 to 2 TB, and the read / write speed is expected to be approximately 1 to 2 GB / s. 5. High-speed interconnection: A high-speed interconnection connects the above hardware components, enabling high-speed data transfer. Specifically, it employs a radiation-resistant PCIe Gen4 interface. This interconnection has the following characteristics: - Radiation immunity: Error detection and correction (EDAC) circuitry and retransmission function reduce communication errors caused by radiation. - High speed: Bandwidth approximately 16 to 32 GB / s - Low latency: Latency of approximately 100 to 200 ns - Flexibility: Supports connection of various devices The expected bandwidth is approximately 16 to 32 GB / s. 6. Thermal Management System: A thermal management system for efficiently dissipating heat generated by computing hardware. Specifically, it employs a passive thermal management system that combines heat pipes, heat sink fins, and radiant cooling plates. This thermal management system has the following features: - High efficiency: Thermal resistance of approximately 0.1 to 0.2°C / W - Passive: Has no moving parts and is highly reliable. - Lightweight: Weighs approximately 5 to 10 kg - Adaptability: Can handle various thermal load conditions. This system maintains the operating temperature of the computing hardware within an appropriate range (approximately -20 to +60°C). By integrating these hardware components, it is possible to provide the computing power necessary for efficient execution of CM while addressing the stringent constraints of the space environment (radiation, power, heat, etc.). In particular, a hybrid architecture combining TPUs and FPGAs can achieve both high performance and flexibility.

[0045] Furthermore, this module supports the following operating modes: 1. Standard mode: This mode is used during normal operation. It utilizes both the TPU and FPGA to achieve maximum processing power. Power consumption is approximately 100 to 200W. 2. Power saving mode: This mode is used when power resources are limited (for example, in poor sunlight conditions). It uses only a portion of the TPU and FPGA, sacrificing processing power to reduce power consumption. Power consumption is approximately 50 to 100W. 3. Emergency Mode: This mode is used when there are extreme power constraints (for example, when some solar panels fail). It maintains only the minimum processing power and minimizes power consumption. Power consumption is approximately 20 to 50W. 4. Update Mode: This mode is used when updating software or hardware configurations. It performs software updates and FPGA reconfigurations based on commands from the ground. 5. Diagnostic mode: This mode is used for system diagnosis and troubleshooting. It collects detailed status information for each component and transmits it to the ground. These operating modes allow for adaptation to various situations and enable optimal performance within limited resources. ### Real-time observation sensor array A real-time observation sensor array provides a group of sensors that observe the Earth across multiple wavelength bands. This sensor array provides high-resolution observational data used as training data for CMs and also functions as input for real-time data assimilation. Specifically, it consists of the following sensors: 1. High-resolution visible and near-infrared imager: A high-resolution visible-light and near-infrared imager for observing cloud patterns and surface characteristics. This sensor performs observations in multiple bands in the visible light range (approximately 400 to 700 nm) and the near-infrared range (approximately 700 to 1100 nm) to acquire information on cloud distribution, surface vegetation, land cover, and more. This imager has the following specifications: - Spatial resolution: Approximately 100 to 250 m (at the point directly beneath) - Spectral resolution: approximately 10 to 20 nm - Number of bands: 8 bands (blue, green, red, near-infrared 1, near-infrared 2, etc.) - Dynamic range: Approximately 12 to 14 bits - Scan width: Approximately 1000 to 1500 km - Field of view: Approximately 10 to 15 degrees - Signal-to-noise ratio (SNR): Approximately 200 to 400 (under standard ground reflectance conditions) - Data rate: Approximately 50 to 100 Mbps The sensor's optical system consists of a reflecting telescope with an aperture of approximately 20 to 30 cm and a CMOS detector array with approximately 4000 x 6000 pixels. It also features an onboard calibration function, ensuring the quality and consistency of observation data over long periods by performing periodic calibrations using sunlight or internal light sources. This imager provides information such as cloud distribution and characteristics, surface vegetation index (NDVI), land cover classification, and snow and ice cover, playing a crucial role in monitoring and predicting meteorological and climate systems. In particular, cloud distribution and characteristics are essential for highly accurate climate prediction because they have a significant impact on the radiation balance and precipitation processes. 2. Microwave radiometer: A microwave radiometer used to observe atmospheric water vapor and precipitation. This sensor acquires information on atmospheric water vapor, cloud water, and precipitation, and is particularly effective in estimating precipitation in cloud-covered areas. This microwave radiometer has the following specifications: - Frequency band: Approximately 10 to 90 GHz (multiple channels) - Main channels: approximately 19GHz (sea surface temperature), 22GHz (water vapor), 37GHz (cloud water), 85GHz (precipitation) - Spatial resolution: Approximately 5 to 25 km (frequency dependent) - Temperature resolution: Approximately 0.3 to 0.5K - Scan width: Approximately 1500 to 2000 km - Field of view: Approximately 45 to 60 degrees Polarization: Vertical polarization and horizontal polarization - Data rate: Approximately 1 to 5 Mbps The sensor's antenna employs an offset parabolic antenna with a diameter of approximately 60 to 80 cm, and observations are performed using a conical scan method. Calibration is also performed periodically by observing black bodies in outer space. This microwave radiometer is an all-weather sensor, capable of observation even in cloud-covered areas. This allows for the acquisition of information about the interior of clouds and the Earth's surface covered by clouds, which cannot be observed with visible light or infrared sensors. It is particularly effective for estimating precipitation at sea and observing the distribution of sea ice in polar regions. 3. Rainfall radar: A rainfall radar for observing three-dimensional precipitation structures. This sensor acquires information such as the three-dimensional structure of precipitation, precipitation intensity, and precipitation type, and is particularly effective for observing extreme precipitation events such as torrential rains and typhoons. This rainfall radar has the following specifications: - Frequency band: Dual frequency band of approximately 13.6 GHz (Ku band) and 35.5 GHz (Ka band) - Spatial resolution: approximately 2 to 5 km (horizontal), approximately 0.25 to 0.5 km (vertical) - Scanning width: Approximately 200 to 300 km - Sensitivity: Approximately 15 to 20 dBZ (minimum detectable reflectivity) - Observation altitude: Approximately 20 km above the Earth's surface - Polarization: Single polarization or dual polarization - Data rate: Approximately 5 to 10 Mbps The sensor's antenna employs a parabolic antenna with a diameter of approximately 1.5 to 2.0 m and performs observations using a cross-track scanning method. Furthermore, the quality of the observation data is ensured by combining an internal calibration function with external calibration (comparison with ground radar). This rainfall radar uses dual-frequency observation to estimate the size distribution of precipitation particles and the type of precipitation (rain, snow, hail, etc.). This can contribute to improving the understanding and prediction of precipitation processes. In particular, Ku-band radar is suitable for observing moderate to heavy precipitation, while Ka-band radar is suitable for observing light precipitation and snow. By combining the two, a wide range of precipitation intensities can be covered. 4. Infrared sounder: An infrared sounder for observing atmospheric temperature and humidity profiles. This sensor acquires information such as atmospheric temperature profiles, humidity profiles, and cloud top altitude, making it effective for understanding the thermodynamic state of the atmosphere. This infrared sounder has the following specifications: - Wavelength range: approximately 3.7 to 15.4 μm - Spectral resolution: approximately 0.5 to 1.0 cm⁻¹ - Number of channels: Approximately 1000 to 2000 channels - Spatial resolution: Approximately 10km (direct point) - Temperature resolution: approximately 0.1 to 0.3K - Scan width: Approximately 1000 to 1500 km - Field of view: Approximately 30 to 45 degrees - Data rate: Approximately 10 to 20 Mbps The sensor's optical system consists of a telescope with an aperture of approximately 10 to 15 cm and a Fourier transform spectrometer or diffraction grating spectrometer. Calibration is also performed periodically by observing internal black bodies and outer space. This infrared sounder allows for detailed observation of the vertical structure of the atmosphere through high spectral resolution. In particular, the vertical profiles of temperature and humidity are important for evaluating atmospheric stability and the potential for convection. Furthermore, it can observe the distribution of trace gases in the atmosphere (ozone, carbon dioxide, methane, etc.), contributing to the understanding of atmospheric chemical processes. 5. Scatterometer: A scattermeter used to observe sea surface winds. This sensor acquires information on wind direction and speed at sea, and is particularly effective in understanding the structure of weather systems such as typhoons and low-pressure systems. This scattermeter has the following specifications: - Frequency band: Approximately 5.3GHz (C band) - Spatial resolution: approximately 25km - Scan width: Approximately 1500 to 2000 km - Field of view: Approximately 45 to 60 degrees - Wind speed accuracy: approximately 1 to 2 m / s - Wind direction accuracy: about 20 degrees - Wind speed range: approximately 3 to 30 m / s - Data rate: Approximately 1 to 2 Mbps The sensor's antenna employs a system combining multiple fixed-beam antennas to observe the sea surface at different incidence and azimuth angles. This allows for the estimation of wind direction and speed from the directional dependence of the backscatter coefficient on the sea surface. This scattermeter can observe the global distribution of sea surface winds and is effective in understanding the structure and intensity of meteorological systems such as tropical and extratropical cyclones. It is also important as input data for ocean circulation models and wave prediction models. These sensors have the following characteristics: 1. Multi-wavelength observations: Observations are conducted across a wide wavelength range, from visible light to radio waves, capturing various characteristics of the atmosphere and Earth's surface. This allows for the acquisition of comprehensive information on clouds, precipitation, water vapor, and surface characteristics. Each wavelength band has sensitivity to different atmospheric and surface characteristics, and by combining these, a more complete understanding of the Earth system's state can be obtained. 2. High temporal resolution: By using satellite constellations, the frequency of observations at the same location can be increased, capturing changes over time. Specifically, observations can be made approximately every 3 to 6 hours globally, and approximately every 30 minutes to 1 hour in specific regions. This high-temporal-resolution observation makes it possible to capture rapidly developing weather systems (for example, the rapid development of tropical cyclones and the occurrence of convective precipitation). 3. High spatial resolution: Observations are conducted with spatial resolution appropriate to the characteristics of each sensor to capture local phenomena. In particular, visible light and near-infrared imagers perform observations with a high resolution of approximately 100 to 250 m, acquiring detailed information on cloud patterns and surface characteristics. This high spatial resolution observation allows for the evaluation of local influences such as topographic effects and land cover effects. 4. Three-dimensional observation: Rainfall radar and infrared sounders are used to observe the three-dimensional structure of the atmosphere. This allows us to understand the three-dimensional structure of precipitation systems and the thermodynamic structure of the atmosphere. Three-dimensional observations are important for evaluating the vertical stability of the atmosphere and the possibility of convection, and are particularly effective in predicting extreme phenomena such as heavy rainfall and thunderstorms. 5. All-weather observation: Microwave radiometers and rainfall radar are used to conduct observations even in cloud-covered areas. This enables continuous observation regardless of weather conditions. All-weather observation is particularly important in cloudy regions such as the tropics and polar regions. 6. Complementary observation: By combining the characteristics of different sensors, information that cannot be obtained with a single sensor can be acquired. For example, by combining a visible light / near-infrared imager with a microwave radiometer, both the optical and microphysical properties of clouds can be understood. Also, by combining a rainfall radar with a scattermeter, the relationship between precipitation systems and sea surface winds can be evaluated. 7. Calibration and Verification Function: Each sensor is equipped with the capability for periodic calibration and validation. Specifically, it ensures the quality and consistency of observational data over long periods by combining methods such as internal calibration sources (blackbody, lamp, etc.), external calibration sources (sun, moon, outer space, etc.), and comparison with ground validation sites. These features enable real-time observation sensor arrays to provide comprehensive observations of the atmosphere and the Earth's surface, offering training data and input for real-time data assimilation of CMs. In particular, integrating information from multiple sensors allows for a comprehensive understanding of the Earth system's state that cannot be obtained from a single sensor.

[0046] ### Onboard Data Processing Unit The onboard data processing unit performs preprocessing of observational and ESM data on the satellite. This unit prepares the input data for the CM and is also responsible for preprocessing for real-time data assimilation. Specifically, the following processes are performed: 1. Calibration and correction of sensor data: The raw data from each sensor is converted into physical quantities, and noise and distortion caused by the sensor's characteristics are corrected. Specifically, the following processes are performed: a. Calibration of radiometric equipment: - Dark current correction: Measure the sensor's dark current (output signal when there is no light input) and subtract it from the observed data. - Nonlinearity correction: Corrects the nonlinearity of the sensor's response based on a calibration curve. - Relative response correction: Corrects the sensitivity difference between each detection element within the sensor. - Absolute calibration: Using an internal calibration source (blackbody, lamp, etc.) or an external calibration source (sun, moon, outer space, etc.), the sensor's response is converted into physical quantities (brightness temperature, reflectance, etc.). b. Geometric correction: - Calculate the geographic coordinates of the observation data based on the satellite's position and attitude information. - The intersection point of the sensor's line-of-sight vector and the Earth's ellipsoid is calculated to determine the latitude and longitude of each pixel. - Considering topographic effects, a digital elevation model (DEM) is used to correct for positional shifts caused by terrain. - Corrects optical distortion in the sensor to generate geometrically accurate images. c. Atmospheric correction: - Correct for the effects of atmospheric absorption, scattering, and radiation to estimate the true properties of the Earth's surface. - Consider the effects of molecular scattering (Rayleigh scattering), aerosol scattering (Mie scattering), and molecular absorption (water vapor, ozone, carbon dioxide, etc.). - Using a radiative transfer model, the reflectance and emissivity of the Earth's surface are estimated from the radiation levels at the top of the atmosphere. - Corrects for adjacent area effects (the influence of scattered light from the surrounding ground surface). These processes are performed using algorithms optimized for each sensor. For example, in the case of visible light and near-infrared imagers, dark current correction, flat field correction, geometric correction, and atmospheric correction are applied. In the case of microwave radiometers, antenna pattern correction, nonlinearity correction, and temperature calibration are applied. In the case of rainfall radar, range profile correction, surface clutter removal, and rainfall attenuation correction are applied. The accuracy of calibration and correction is a crucial factor affecting the quality of observational data. In particular, stable calibration is essential to ensure the consistency of observational data over long periods. Therefore, regular calibration and monitoring of changes in calibration parameters over time are important. Furthermore, cross-calibration between different sensors is also performed to ensure compatibility between sensors. 2. Cloud masking and quality control: We identify the effects of clouds from observational data and evaluate the quality of the data. Specifically, we perform the following processes: a. Cloud detection: - Apply cloud detection algorithms based on spectral, spatial, and temporal characteristics to visible light and near-infrared imager data. - Clouds are detected by combining the ratio of reflectance in the visible and near-infrared regions, the brightness temperature in the thermal infrared region, and texture information. - Improve cloud detection accuracy under complex conditions using machine learning algorithms (e.g., random forests, support vector machines, convolutional neural networks, etc.). - Detected clouds are classified by cloud type (stratocumulus, cumulus, cirrus, etc.) and cloud height (low-level clouds, mid-level clouds, high-level clouds). b. Quality control: - The data from each sensor is subjected to anomaly detection, noise evaluation, and data reliability evaluation. - Detect outliers using statistical methods (e.g., Z-score, modified Z-score, local outlier coefficient, etc.). - Data quality is evaluated using metrics such as signal-to-noise ratio (SNR), coefficient of variation, and spatial consistency. - Monitor the operating status of sensors (e.g., temperature, voltage, current, etc.) and set a data quality flag if an anomaly is detected. - Detect known issues such as terrain influences (e.g., terrain echoes in rainfall radar in mountainous areas), solar glints (specular reflection of sunlight), and radio interference, and flag the corresponding data. c. Data reliability assessment: - Assign a quality flag and a confidence score to each data point. - Quality flags are discrete indicators that show the state of the data (e.g., good, cloudy, outliers, missing data, etc.). - The confidence score is an index that represents the reliability of data as a continuous value between 0 and 1, and is used for weighting in subsequent processing. - When integrating data from multiple sensors, adjust the contribution of each sensor based on its confidence score. These processes ensure the quality of data used in subsequent processing. In particular, cloud effects are a significant issue in many remote sensing applications, and proper cloud masking is essential for generating high-quality data products. Furthermore, quality control processes mitigate the impact of outliers and low-quality data, leading to more reliable analysis results. 3. Spatial interpolation and low-pass filtering of ESM data: Low-resolution ESM data received from the ground is interpolated into a high-resolution grid, and small artifacts resulting from the interpolation are removed. Specifically, the following processes are performed: a. Spatial interpolation: - Bilinear interpolation: Uses the values ​​of four neighboring grid points to calculate high-resolution grid values ​​using linear interpolation. The calculation is simple and fast, but the interpolation results may not be smooth. - Cubic spline interpolation: Using the values ​​of 16 neighboring grid points, the high-resolution grid values ​​are calculated using cubic polynomial interpolation. The calculation is more complex, but the interpolation results are smoother. - Spherical harmonic interpolation: This interpolation method is suitable for data on a sphere and is suitable for interpolating global data. In particular, it exhibits less distortion near the polar regions. - Conservative interpolation: This interpolation method preserves conserved quantities such as mass and heat, and is important for maintaining physical consistency. - Terrain-aware interpolation: This interpolation method takes terrain information into account and is suitable for interpolating variables where terrain effects are important (e.g., precipitation, temperature, etc.). b. Low-pass filtering: - Gaussian filter: Smoothes data using a weight function with the shape of a normal distribution. It is computationally simple and fast, and allows for isotropic smoothing. - Bilateral filter: This filter smooths edges while preserving them by using a weighting function that considers not only spatial distance but also the difference in values. This allows for the preservation of physical boundaries (e.g., fronts, coastlines, etc.). - Anisotropic diffusion filter: Adjusts the smoothing intensity according to the direction of the local gradient, smoothing while preserving the physical structure. - Wavelet filter: Separates components from different scales and removes noise only from a specific scale. Suitable for multiscale analysis. - Kalman filter: A filtering method that takes temporal consistency into consideration and is suitable for smoothing time-series data. c. Ensuring physical consistency: - Conservation of mass: Ensure that mass (e.g., atmospheric mass, water vapor content, etc.) is conserved during the interpolation and filtering process. - Energy conservation: Ensure that energy (e.g., thermal energy, kinetic energy, etc.) is conserved during the interpolation and filtering process. - Physical constraints: Impose constraints so that the values ​​fall within a physically meaningful range (for example, relative humidity is between 0 and 100%). - Spatial consistency: Adjustments are made to maintain the physical relationships between different variables (e.g., the relationship between temperature and density). These processes allow for the preparation of appropriate ESM data as input for CM. In particular, it is important to remove artifacts caused by interpolation (e.g., Gibbs effect, aliasing, etc.), thereby improving the quality of high-resolution processing. Furthermore, ensuring physical consistency allows for the generation of physically meaningful high-resolution data. 4. Bias correction using quantile delta mapping (QDM): Remove the distribution bias from the ESM simulation. Specifically, perform the following steps: a. Quantile calculation: - Calculate the cumulative distribution function (CDF) of the ESM data and the observed data. - Calculate approximately 500 quantiles (0.1%, 0.2%, ..., 99.9%, etc.) to capture the detailed shape of the distribution. - Calculate spatially local CDFs and consider the distribution characteristics of each region. - Calculate the CDF for each season and take seasonal variations into account. b. Bias correction: - Calculate the bias correction coefficient for each quantile. - For each value in the ESM data, apply the bias correction coefficient for the corresponding quantile. - For values ​​between quantiles, the bias correction coefficient is calculated using linear interpolation. - For extreme values ​​(below the minimum value or above the maximum value), the bias correction coefficient is calculated using extrapolation. c. Save trends: - For future climate scenarios, apply trend-preserving QDM to preserve climate change signals. - Consider changes in the distribution between the base period and future periods, and preserve relative changes. - Depending on the type of variable, choose either additive (e.g., temperature) or multiplicative (e.g., precipitation) correction. d. Ensuring spatial consistency: - Applying QDM independently to each grid cell may result in spatial discontinuities. - To generate a spatially smooth bias correction field, information from neighboring grid cells is considered. - Multivariate QDM is applied to ensure consistency between physically related variables. These processes can bring the statistical characteristics of ESM data closer to those of observational data. QDM can correct not only the mean and standard deviation but also the overall shape of the distribution, thus improving the representation of extreme values. This is particularly important in forecasting extreme precipitation events. Furthermore, trend-preserving QDM can correct biases while preserving climate change signals.

[0047] 5. Data normalization and transformation: The data is converted to an appropriate range and optimized as input for the CM. Specifically, the following processes are performed: a. Logarithmic transformation: - Apply a logarithmic transformation to variables with non-negative and skewed distributions, such as precipitation. The transformation is performed in the form - x^~ = log(x + E) - log[E], and zero values ​​are handled appropriately. Here, E is a small positive constant (e.g., 0.0001). - Logarithmic transformation reduces the skewness of the distribution, bringing it closer to a normal distribution. This improves the efficiency of training and inference for machine learning models. b. Scaling: - Normalize the converted data to a range of approximately [-1, 1]. - Minimum-maximum scaling: Scaling is performed in the format (x - min) / (max - min) * 2 - 1. - Z-score normalization: Scaling is performed in the format (x - mean) / std, and then adjusted to an appropriate range. - Robust scaling: Scaling is performed using the median and interquartile range to mitigate the impact of outliers. c. Feature Engineering: - Extract physically meaningful features and add them as input to the CM. - For example, it calculates derived variables such as saturation (calculated from temperature and relative humidity), horizontal divergence (calculated from wind speed and direction), and geostrophic wind (calculated from pressure gradient). - These features can improve the accuracy of CM training and inference because they more directly represent the physical process. d. Dimensionality reduction: - Reduce the dimensionality of the data using techniques such as principal component analysis (PCA) and autoencoders. - This improves computational efficiency and reduces the impact of noise. - However, this should be applied with caution as it may reduce physical interpretability. These transformations can adjust the data distribution and improve the efficiency of CM training and inference. Logarithmic transformations, in particular, are effective for variables with non-negative and skewed distributions, such as precipitation, and can improve the representation of extreme values. Furthermore, appropriate scaling can adjust for differences in scale between different variables, improving the stability of CM training. 6. Data compression and storage: The processed data is compressed and stored in non-volatile storage. Specifically, the following processes are performed: a. Data compression: - Lossless compression: A compression method that completely preserves the information of the data. For example, it uses algorithms such as Huffman coding, arithmetic coding, and LZ77 / LZ78. - Lossy compression: A compression method that achieves a high compression ratio at the expense of some information. For example, it uses discrete cosine transform (DCT) or discrete wavelet transform (DWT). - Predictive coding: A compression method that reduces data redundancy using temporal and spatial predictive models. Examples include differential pulse code modulation (DPCM) and adaptive differential pulse code modulation (ADPCM). - CCSDS-recommended compression algorithm: Use the compression algorithm recommended by the Inter-Space Agency Council (CCSDS). For example, use CCSDS 121.0-B-2 (lossless data compression), CCSDS 122.0-B-2 (image data compression), etc. b. Data storage: - Hierarchical storage management: Data is placed in different storage layers (high-speed memory, flash storage, high-capacity storage, etc.) based on access frequency. - Data retention policy: Set retention periods according to the importance and freshness of the data. For example, raw data should be stored for a short period (a few days to a few weeks), processed data for a medium period (a few weeks to a few months), and important event data for a long period (a few months to a few years). - Redundant storage: Critical data is stored redundantly on multiple storage devices to improve resilience against hardware failures. - Error Detection and Correction (EDAC): Applies error detection and correction codes during reading and writing to storage devices to ensure data integrity. c. Data management: - Metadata management: Records data attributes (e.g., observation time, sensor type, processing level, quality information, etc.) to enable efficient data retrieval and management. - Version control: Tracks data reprocessing due to updates to data processing algorithms and manages different versions of data. - Province tracking: Records the processing history from data generation to the current state to ensure data reliability and reproducibility. These processes improve storage utilization efficiency while ensuring data integrity and availability. In particular, in space environments, where the risk of hardware failures such as bit inversion due to radiation is high, measures such as error detection and correction and redundant storage are crucial. Furthermore, appropriate data compression and retention policies are necessary to efficiently utilize limited storage capacity. 7. Data indexing and retrieval: Create an index to efficiently search the stored data. Specifically, perform the following steps: a. Spatiotemporal index: - Time Index: Create an index based on observation time or prediction time to efficiently search for data within a specific time range. - Spatial index: Creates an index based on geographic coordinates (latitude and longitude) to efficiently search for data in a specific region. Spatial index structures such as R-trees, quadtrees, and grid indexes are used. - Spatiotemporal combined index: Creates an index that considers both time and space, making spatiotemporal searches (for example, data for a specific region within a specific period) more efficient. b. Attribute index: - Create indexes based on attributes such as sensor type, variable type, processing level, and quality flag, enabling efficient retrieval of data that meets specific criteria. - High-speed attribute searching is achieved using data structures such as B-trees and hash tables. c. Content-based index: - Create an index based on the content of the data (e.g., specific patterns or characteristics) to enable searching for similar data. - We achieve efficient content-based search using techniques such as feature vector extraction, dimensionality reduction, and clustering. - For example, you can search for data that includes specific weather patterns (typhoons, weather fronts, etc.). d. Diversification index: - Create distributed indexes that span multiple satellites and ground systems, enabling data retrieval across the entire system. - We enable efficient searching in distributed environments using technologies such as federated indexes and distributed hash tables (DHTs). These indexes allow for quick retrieval of necessary data and efficient acquisition of data required for real-time data assimilation and CM execution. In particular, spatiotemporal indexes are crucial for searching Earth observation data, enabling efficient spatiotemporal queries. Attribute indexes, on the other hand, allow for efficient extraction of data that meets specific criteria (e.g., high-quality precipitation data). These processes have the following characteristics: 1. Real-time processing: The system performs data reception and processing in real time, minimizing processing delays. Specifically, it reduces the time from sensor data reception to pre-processing completion from a few seconds to a few minutes. This allows for the timely provision of data necessary for real-time data assimilation and CM execution. To achieve real-time processing, the following technologies will be employed: - Pipeline processing: Data processing is divided into multiple stages, and each stage is executed in parallel. This can improve throughput. - Stream processing: Data is divided into small chunks, and processing begins as each chunk becomes available. This reduces latency. - Event-driven architecture: Processing is driven based on events such as data arrival or process completion. This enables efficient resource utilization and processing flexibility. - In-memory processing: Minimizes disk I / O and processes data in memory. This can improve processing speed. 2. Adaptive processing: Processing parameters are adaptively adjusted according to the characteristics and quality of the data. Specifically, atmospheric correction parameters are adjusted according to observation conditions such as cloud cover and atmospheric conditions, and the filtering intensity is adjusted according to the quality of the data. This allows for obtaining high-quality processing results under various conditions. To achieve adaptive processing, the following technologies will be employed: - Data-driven parameter tuning: Automatically adjusts processing parameters based on the statistical characteristics of the data (e.g., mean, variance, skewness, etc.). - Machine learning-based parameter optimization: Learns optimal parameters from past data and processing results, and applies them to new data. - Feedback control: Continuously adjust processing parameters based on quality indicators of the processing results. - Multi-model approach: Prepare multiple processing models optimized for different conditions and select the appropriate model according to the characteristics of the data. 3. Efficient implementation: The goal is to efficiently execute processing within limited computing resources. Specifically, this involves utilizing technologies such as parallel processing, pipeline processing, and hardware acceleration to improve processing efficiency. Furthermore, computational load is reduced through optimization of processing algorithms and the introduction of approximation calculations. This allows for the execution of diverse preprocessing tasks within the satellite's limited computing resources. To achieve efficient implementation, the following technologies will be adopted: - Parallel processing: Parallelize data processing by utilizing the parallel processing capabilities of multi-core processors and FPGAs. - Hardware acceleration: Offloading specific operations (e.g., FFT, matrix operations, image processing, etc.) to hardware accelerators to speed up processing. - Memory optimization: Improves the efficiency of memory access through efficient cache utilization, optimization of memory access patterns, and optimization of data layout. - Algorithm optimization: Reduce computational complexity by selecting algorithms with low computational complexity, introducing approximation calculations, and reusing calculations. - Dynamic resource allocation: Dynamically allocates computing resources based on processing priority and available resources.

[0048] 4. Robustness: This system provides robust processing against data loss and noise. Specifically, it addresses data quality issues using methods such as anomaly detection and interpolation, noise reduction filters, and robust statistical techniques. This allows processing to continue as much as possible even in the event of temporary sensor failures or communication problems. To achieve robustness, the following technologies are employed: - Robust statistical methods: Robust statistical methods such as median, interquartile range, and M-estimator are used to mitigate the impact of outliers. - Anomaly detection and handling: Statistical and machine learning methods are used to detect, exclude, or correct for anomalies. - Data interpolation: Missing data is estimated through spatiotemporal interpolation to ensure data continuity. - Multi-sensor data fusion: By integrating data from multiple sensors, the resilience of a single sensor to failure is improved. - Fallback strategy: If the primary algorithm fails, switch to a simpler but more robust alternative algorithm. In this way, the onboard data processing unit performs preprocessing of observation data and ESM data in real time on the satellite, providing preprocessed data for CM input and real-time data assimilation. This eliminates delays associated with ground processing and achieves real-time performance. Furthermore, adaptive processing and robustness enable high-quality processing results under various conditions. In addition, efficient implementation allows for the execution of diverse preprocessing within the satellite's limited computing resources. ### Satellite-mounted CM implementation module The satellite-borne CM implementation module provides an optimized implementation for running the CM algorithm proposed by Hess et al. on satellite. This module implements a core algorithm that generates high-resolution data from low-resolution ESM data in a single step. Specifically, the following technologies will be implemented: 1. Quantization and lightweighting: We reduce memory usage and computational complexity by quantizing model parameters. Specifically, we apply the following techniques: a. Parameter quantization: - Quantization of model parameters from 32-bit floating-point numbers to 8-bit integers - Quantization scheme: Select an appropriate scheme such as symmetric quantization, asymmetric quantization, or channel-by-channel quantization. - Optimization of quantization parameters (scale, zero point, etc.) - Retraining (Quantization-Aware Training) or post-training (Post-Training Quantization) to minimize quantization errors. - Mixed-precision quantization: Critical layers are quantized with high precision (e.g., 16 bits), while others are quantized with lower precision (e.g., 8 bits). b. Pruning: - Identify and remove parameters of low importance (e.g., those with small weights or low activation). - Structured pruning: Improves hardware efficiency through channel-level, filter-level, and layer-level pruning. - Iterative pruning: Repeated retraining after pruning achieves high sparsity while maintaining accuracy. - Importance Criteria: Importance assessment based on L1 norm, L2 norm, Fisher divergence, Hessian matrix, etc. - Dynamic sparsity: Temporarily skips low-activation neurons during runtime. c. Knowledge distillation: - Transferring knowledge from a larger model (teacher model) to a smaller model (student model) - Mimicking the output probability distribution: Training the student model to mimic the output probability distribution (soft target) of the teacher model. - Feature representation imitation: Train the student model to imitate the feature representations of the intermediate layer of the teacher model. - Distillation of relational knowledge: Transferring the relationships between samples learned by the teacher model to the student model. - Self-distillation: Knowledge transfer from a large model to a smaller model of the same architecture. d. Architecture optimization: - Adoption of efficient architectures: Applying the design principles of efficient architectures such as MobileNet, EfficientNet, and ShuffleNet. - Depth-directed convolution: Standard convolutions are decomposed into depth-directed convolutions and point-wise convolutions to reduce computational complexity. - Bottleneck structure: Temporarily reduce the number of channels to reduce computational load. - Skip connection: Improves the flow of information and enables high performance even in shallower networks. - Shared parameters: Share parameters between different layers to reduce model size. These techniques can reduce model size by approximately 75 to 90% and improve inference speed by approximately 3 to 5 times. For example, if the original CM model has approximately 27 million parameters, quantization and optimization can reduce it to approximately 3 to 7 million parameters. This significantly reduces memory usage, allowing the model to run efficiently within the limited memory resources of the satellite. Furthermore, the reduction in computational load improves inference speed and energy efficiency. 2. Sparse computation: We improve energy efficiency by eliminating unnecessary calculations. Specifically, we apply the following techniques: a. Dynamic pruning: - Identify and omit calculations that are not important at runtime. - Skip calculations based on activation threshold - Prioritizing calculations based on the weighting of attentional mechanisms - Optimization of sparse matrix operations (e.g., calculations in compressed sparse row / column format) - Zero skipping: Omits multiplication of zero values b. Conditional calculations: - Select the calculation path according to the characteristics of the input. - Conditional branching network: Selects different subnetworks based on the input. - Gating mechanism: Dynamically controls the execution of each layer and module. - Adaptive computation time: The computation ends when the required precision is reached. - Input-dependent computation: Allocates computational resources according to the complexity of the input. c. Early Exit: - The calculation will terminate when sufficient accuracy is obtained from the output of the hidden layer. - Confidence estimation: Evaluate the confidence level of the output of the hidden layer, and if it is sufficiently high, skip the remaining calculations. - Multistage inference: Process with a simple model and apply a complex model only when necessary. - Cascade structure: Building a chain of models that gradually increase in complexity. - Adaptive precision: Adjusts the computational depth according to the required precision. d. Optimization of attention mechanisms: - Sparse attention mechanism: Focusing attention only on important locations. - Local attention mechanism: Attention calculations are performed only within a local range, not for all position pairs. - Linear attention mechanism: An approximation method that linearly reduces quadratic computation complexity. - Hierarchical attention mechanism: Computes attention at multiple resolution levels. - Memory-efficient attention mechanism: Implicit processing without explicitly calculating the attention map. These technologies can reduce computational complexity by approximately 50 to 70%, improving energy efficiency. In particular, the CM inference process can involve many redundant calculations, and sparse computation techniques can eliminate these redundant calculations. This allows for efficient model execution within the satellite's limited power resources. Furthermore, the reduction in computational complexity improves inference speed, enhancing real-time capabilities. 3. Pipeline processing: Optimize data flow and improve throughput. Specifically, apply the following techniques: a. Model pipeline: - Pipelining process that executes each layer of the model in parallel. - Layered pipeline: Each layer of the network is executed on a separate processor. - Block pipeline: Pipelining is performed using blocks consisting of multiple layers as the unit. - Micro-batch pipeline: Divides a batch into smaller micro-batches for pipeline processing. - Bubble Avoidance: Optimization to minimize pipeline bubbles (empty slots) - Asynchronous Execution: Efficiency improvement through asynchronous execution of each pipeline stage b. Batch Processing: - Process multiple inputs simultaneously - Dynamic Batch Size: Adjust the batch size according to available resources and input characteristics - Batch Prioritization: Determine the processing order of batches based on urgency and importance - Adaptive Batch Formation: Group inputs with similar characteristics to improve processing efficiency - Intra-Batch Parallel Processing: Partially process each sample in a batch in parallel c. Memory Access Optimization: - Efficient Utilization of Cache - Data Layout Optimization: Data arrangement adapted to memory access patterns - Tiling: Divide data into tiles of a size that fits in the cache for processing - Prefetching: Load next required data into the cache in advance - Memory Coalescing: Promote contiguous memory accesses - Bank Conflict Avoidance: Load balancing across memory banks d. Overlap of Computation and Data Transfer: - Execute computation and data transfer in parallel - Double Buffering: Load data into one buffer while performing computation on the other buffer - Asynchronous Data Transfer: Asynchronous data transfer using DMA or the like - Computation Kernel Splitting: Split into small kernels that can overlap with data transfer - Scheduling Optimization: Optimally schedule computation and data transfer tasks These technologies can improve throughput by approximately two to four times. They are particularly effective when processing multiple satellite observation and ESM data simultaneously. Pipelining allows different parts of the model to be executed in parallel, improving the efficiency of hardware resource utilization. Furthermore, memory access optimization reduces memory bandwidth bottlenecks and improves computational efficiency.

[0049] 4. Hardware acceleration: Accelerate specific calculations using FPGAs. Specifically, apply the following techniques: a. Acceleration of convolution operations: - Implement a dedicated convolutional engine on the FPGA. - Spatial parallelization: Execute multiple convolution kernels in parallel. - Channel parallelization: Processes multiple input / output channels in parallel. - Window buffering: Efficient buffering of input feature maps - Window sliding: A sliding window implementation that reduces computational redundancy. - Selective application of direct convolution and FFT-based convolution b. Acceleration of matrix multiplication: - Implement a dedicated array processor for matrix multiplication on an FPGA. - Overall parallelization: Calculate multiple matrix elements in parallel. - Block matrix multiplication: Divides a large matrix into smaller blocks for processing. - Implementation of matrix multiplication optimization algorithms (Strassen, Winograd, etc.) - Optimization of sparse matrix multiplication: Efficient implementation that skips zero values - Low-precision matrix multiplication: Matrix multiplication using low precision such as 8 bits or 4 bits. c. Acceleration of activation functions: - Dedicated hardware implementation of activation functions (ReLU, sigmoid, tanh, etc.) - Lookup Table (LUT) based implementation: Replaces calculations with faster table lookups. - Piecewise linear approximation: Approximating complex activation functions with piecewise linear functions. - Parallel evaluation: Calculates multiple activation values ​​in parallel. - Pipeline implementation: Pipelining the calculation of activation functions - Approximation algorithms: Simplify calculations while maintaining accuracy. d. Memory access acceleration: - Efficient use of on-chip memory - Smart Caching: Optimal caching strategies based on access patterns - Memory controller optimization: Utilizing burst transfers and prefetching - Data flow optimization: Efficient supply of data required for computation - Bank-parallel access: Parallel access to multiple memory banks - Optimizing the memory hierarchy: Efficient combination of memory of different sizes and speeds. These technologies can improve the execution speed of these operations by approximately 5 to 10 times. In particular, accelerating convolution and matrix multiplication, which constitute the majority of CM calculations, using FPGAs can significantly improve overall inference speed. Leveraging the reconfigurability of FPGAs, hardware accelerators optimized for specific workloads can be implemented. This results in higher energy and computational efficiency compared to general-purpose processors. 5. Memory Optimization: Minimize memory usage and optimize memory access. Specifically, apply the following techniques: a. Memory reuse: - Temporary storage and reuse of calculation results - Feature map reuse: Reusing the same feature map in multiple operations. - Weight reuse: Applying the same weight to multiple inputs. - Computational graph optimization: removing redundant computations and reusing intermediate results - Memory pooling: efficient reallocation of used memory - In-place operation: reusing input buffers as output buffers where possible b. Memory layout optimization: - Ingenious data arrangement - Selection between channel-last format (NHWC) and channel-first format (NCHW) - Memory alignment: data arrangement adapted to the hardware memory access unit - Padding optimization: padding for improved memory access efficiency - Interleaving: efficient arrangement of different data structures - Cache line consideration: data layout optimized for cache line size c. Efficient utilization of memory hierarchy: - Appropriate differentiated use of cache, main memory, and storage - Cache blocking: splitting computations to fit the cache size - Utilization of cache prefetching: prefetching for predictable access patterns - Utilization of scratchpad memory: efficient use of small-scale programmable high-speed memory - Memory access pattern optimization: maximizing stride-1 accesses - Software-controlled caching: efficiency improvement through explicit cache management d. Compressed memory representation: - Sparse data structure: efficient data structure that does not explicitly store zero values - Quantization representation: reducing memory usage via low bit-width representation - Mixed-precision representation: representing important data with high precision and other data with low precision - Lossless compression such as Huffman coding: representing frequent patterns with short codes - Run-length encoding: efficiently representing consecutive identical values - Approximate data structures: Representations that significantly reduce memory usage at the expense of some accuracy. These techniques can reduce memory usage by approximately 30 to 50% and improve the efficiency of memory access. These memory optimization techniques are particularly effective during the CM inference process, where a large number of intermediate feature maps are generated. Reducing memory usage allows for the execution of larger input sizes and more complex models within limited memory resources. Furthermore, improved memory access efficiency mitigates memory bandwidth bottlenecks and enhances inference speed. 6. Parallel processing: We will implement parallel processing utilizing multiple processor cores and FPGA blocks. Specifically, we will apply the following technologies: a. Data parallelism: - Process multiple input data in parallel - Batch parallel processing: Processing different batch samples on different processors. - Spatial parallel processing: Processing different spatial regions of input using different processors. - Channel parallel processing: Different input / output channels are processed by different processors. - Sub-batch processing: Dividing a batch into multiple sub-batches for parallel processing. - Dynamic load balancing: Dynamically distributes data according to processing load. b. Model parallelism: - Execute different parts of the model in parallel. - Layer parallelism: Running different layers on different processors - Filter parallelism: Different filters within the same layer are executed on different processors. - Attention head parallelism: Different attention heads run on different processors. - Expert parallelism: Running different subnetworks (experts) on different processors. - Hybrid parallelism: A combination of data parallelism and model parallelism. c. Pipeline parallelism: - Run different layers of the model in parallel. - Micro-batch pipeline: Pipeline processing in small batch units - Bubble reduction technology: Minimizing pipeline bubbles (empty slots) - Adaptive pipeline: Dynamically adjusts pipeline stages according to computational load. - Asynchronous pipeline: Asynchronous execution reduces synchronous overhead. - Optimizing pipeline scheduling: Efficient task allocation d. Task parallelism: - Execute independent tasks in parallel - Parallel execution of preprocessing and inference: Data preprocessing and inference are performed in parallel. - Parallel execution of post-processing and inference: Inference and post-processing of results are executed in parallel. - Parallel execution of multiple models: Execute different models in parallel. - Asynchronous execution: Minimizes dependencies between tasks and executes them asynchronously. - Work stealing: An idle processor takes over tasks from other processors. These technologies can improve processing speed by approximately two to four times. In particular, in satellite-borne high-performance computing modules equipped with multiple processor cores or FPGA blocks, these parallel processing technologies can maximize the efficiency of hardware resource utilization. Combining data parallelism and model parallelism enables efficient parallel processing for various workloads. Furthermore, pipeline parallelism allows for the parallel execution of different parts of a model, improving throughput. 7. Adaptive precision control: The calculation precision is adaptively adjusted according to the required accuracy. Specifically, the following techniques are applied: a. Dynamic precision scaling: - Adjust accuracy according to the complexity of the input. - Confidence estimation: Evaluate the uncertainty and complexity of the input and estimate the required accuracy. - Dynamic allocation of computing resources: Allocate computing resources according to the required precision. - Adaptive quantization: Adjusting quantization parameters according to the characteristics of the input. - Accuracy Feedback: Adjusts accuracy based on output quality. - Risk-based precision control: Adjusts precision according to the impact of errors. b. Mixed precision calculation: - Important calculations are performed with high precision, while others are performed with low precision. - Layer-specific precision settings: Set the precision according to the sensitivity of each layer. - Precision settings for each operation: Set appropriate precision for each operation, such as addition, multiplication, and activation. - Improved accuracy of cumulative calculations: Performs calculations prone to cumulative errors (e.g., cumulative addition of matrix multiplication) with high accuracy. - Gradient scaling: Gradient scaling to ensure numerical stability - Optimal placement of precision conversion: The optimal placement that minimizes the overhead of precision conversion. c. Approximate calculation: - Selectively apply fast but approximate calculations. - Function approximation: Replacing complex functions (e.g., exponential functions, trigonometric functions) with polynomial approximations, etc. - Matrix approximation: Approximation using techniques such as low-rank approximation and random projection. - Fast Fourier Transform (FFT) approximation: FFT approximation reduces computational complexity at the expense of accuracy. - Approximate activation function: Piecewise linear approximation of complex activation functions, etc. - Selective computation: Perform only essential computations, approximating or omitting others. d. Error compensation technology: - Techniques to compensate for errors in low-precision calculations - Residual calculation: Calculates the difference between the low-precision calculation result and the true result with high precision and corrects it. - Iterative refinement: A low-precision solution is used as the initial solution, and the process is iteratively refined with higher precision. - Error prediction: Predicts and corrects future errors based on past error patterns. - Statistical correction: Statistically estimate and correct for bias in low-precision calculations. - Ensemble method: Estimates high-precision results by combining multiple low-precision calculation results.

[0050] These techniques allow for the optimization of the balance between accuracy and computational efficiency. In particular, by dynamically adjusting computational accuracy according to different input conditions and required precision, limited computational resources can be used efficiently. By performing critical calculations with high precision and others with lower precision, computational efficiency can be improved while maintaining overall accuracy. Furthermore, by combining approximation and error compensation techniques, the trade-off between computational accuracy and computational efficiency can be optimized. These optimization techniques give the CM implementation the following characteristics: 1. Fast inference: This method generates high-resolution data in a single step. Specifically, it can generate high-resolution data of 0.75°×0.9375° from low-resolution 3°×3.75° ESM data in approximately 0.1 seconds. This is about 400 times faster than conventional SDE diffusion models. This high-speed inference enables real-time high-resolution processing, offering a significant advantage in time-constrained applications (e.g., heavy rain forecasting and disaster response). 2. Low memory usage: It operates efficiently within limited memory resources. Specifically, it can keep memory usage during inference to approximately 1 to 2 GB. This allows CM to be executed within the satellite's limited memory resources. Low memory usage is particularly important when running CM in parallel with other tasks (e.g., data preprocessing and data assimilation). 3. Low power consumption: It operates efficiently within limited power resources. Specifically, power consumption during inference can be kept to around 10 to 20W. This allows CM to be performed within the satellite's limited power resources. Low power consumption is important for long-term satellite operation, and enables continuous operation, especially during periods of poor sunlight conditions (for example, when the Earth is in shadow). 4. High adaptability: It adapts to various input conditions. Specifically, it can adaptively operate to ESM data of different resolutions, different weather and climate conditions, and different regional characteristics, generating high-quality, high-resolution data. This adaptability allows it to be applied to various ESMs and regions without retraining, improving the system's versatility. 5. Robustness: It is robust against input data quality issues and calculation errors. Specifically, it is robust against input data loss and noise, and resistant to calculation errors (e.g., temporary bit flips due to radiation). This enables long-term stable operation in the space environment. Robustness is particularly important for operation in radiation environments, improving the reliability and availability of the system. The specific implementation of CM employs a 2D U-Net architecture, featuring four down / upsampling layers and an attention mechanism. This architecture consists of the following components: 1. Encoder section: The input data is downsampled in stages to extract features. Specifically, it has four downsampling layers, each consisting of convolution, group normalization, and sigmoid linear unit activation. The number of channels increases in stages: 128, 128, 256, and 256. Each downsampling layer performs the following: - Convolution: 3x3 kernel size, stride 2 convolution reduces spatial resolution by half and increases the number of channels. - Group Normalization: Group normalization of 32 groups achieves stable normalization that is independent of batch size. - Sigmoid linear unit (SiLU) activation: Activation functions of the form f(x) = x * sigmoid(x) mitigate the vanishing gradient problem and improve expressiveness. In the encoder section, the output of each layer is saved for skip connections, and then combined with the corresponding layer in the decoder section. This allows for the recovery of spatial details lost due to downsampling. 2. Bottleneck section: We apply an attention mechanism to the lowest-resolution feature map to extract global features. Specifically, we use a self-attention mechanism to capture the relationships between different locations on the feature map. The attention mechanism performs the following actions: - Generating queries (Q), keys (K), and values ​​(V): Generate Q, K, and V from feature maps using linear transformations. - Calculation of attention score: The attention score is calculated in the form Q * K^T / sqrt(d_k) (where dk is the number of dimensions of the key). - Applying Softmax: Apply the softmax function to the attention score and convert it to a probability distribution. - Weighted aggregation: Weighting V using a probability distribution. - Output projection: Generates the output by linearly transforming the aggregated results. The bottleneck attention mechanism allows us to capture long-range dependencies between different locations on the feature map. This is important for representing large-scale weather patterns (e.g., fronts, cyclones, etc.). 3. Decoder section: The output of the bottleneck portion is upsampled in stages to generate a high-resolution output. Specifically, it has four upsampling layers, each consisting of transposed convolution, group normalization, and sigmoid linear unit activation. It also has skip connections from the corresponding layers in the encoder portion to preserve detailed features. Each upsampling layer performs the following: - Transposed Convolution: Transposed convolution with a 3x3 kernel size and stride of 2 doubles spatial resolution and reduces the number of channels. - Skip connection: Combines feature maps from the corresponding layers in the encoder portion. - Group normalization: Group normalization of 32 groups - Sigmoid linear unit (SiLU) activation The decoder integrates detailed features from the encoder through skip connections while performing stepwise upsampling. This allows for the generation of high-resolution output while preserving important features of the input. 4. Output layer: The final high-resolution output is generated. Specifically, a 1x1 convolution is used to convert the feature map into an output with 1 channel. The output layer performs the following processes: - 1x1 convolution: Converts a feature map to an output with 1 channel. - Scaling and inverse transformation: Restore normalized data to its original scale and apply inverse logarithmic transformation as needed. The output layer generates the final high-resolution precipitation data. This data has the same physical units (e.g., mm / day) as the original ESM data and is defined on a high-resolution grid. 5. Time embedding: The model is provided with information on the noise level (time step). Specifically, time information is transformed using sine and cosine position embeddings and added to the feature maps of each layer. Time embedding performs the following process: - Sine-Cosine Transform: Converts a time step t into its sine and cosine frequency components. - Linear projection: Linear projection of transformed time information into the appropriate dimension. - Adding to feature maps: Add projected time information to the feature maps of each layer. Time embedding allows the model to perform different operations depending on the noise level (time step). This is important in training the CM self-consistency function and is necessary to produce consistent outputs at different noise levels. This architecture is optimized to meet the constraints of satellite-borne high-performance computing modules. Specifically, it applies techniques such as quantization, pruning, pipeline processing, and hardware acceleration to enable efficient execution within limited computing resources. For example, the convolutional layers of the encoder are offloaded to an FPGA and executed at high speed using dedicated hardware. The attention mechanism in the bottleneck section is optimized using techniques such as sparse attention mechanisms and linear attention mechanisms to improve computational efficiency. Similarly, the transposed convolutional layers of the decoder are also offloaded to an FPGA and executed at high speed. Furthermore, skip-connected data transfer mitigates memory bandwidth bottlenecks through efficient memory management and data flow optimization. These optimizations allow CM to run efficiently within the satellite's limited computing and energy resources, enabling real-time high-resolution imaging. ### Dynamic Scale Adaptation System The dynamic scale adaptation system provides the ability to automatically adjust the optimal spatial scale according to observation conditions. This system plays a crucial role in ensuring the accuracy and reliability of high-resolution imaging under various weather and climatic conditions. Specifically, the following features will be implemented: 1. Real-time PSD analysis: This method analyzes the power spectral density (PSD) of observational data and ESM data in real time. Specifically, it uses a two-dimensional Fourier transform to convert spatial domain data to the wavenumber domain and calculates the power for each wavenumber. This analysis helps to understand the spatial scale characteristics of observational data and ESM data and identifies the wavenumber (spatial scale) at which the PSDs of both data intersect. This intersection represents the smallest spatial scale at which the ESM data is reliable and is a natural choice for the spatial scale to be preserved when increasing resolution. Real-time PSD analysis is performed in the following steps: a. Data preprocessing: - Interpolating observational data and ESM data into the same spatial grid - Trend removal: Removes large-scale spatial trends and focuses on local fluctuations. - Apply window: Apply a window function (e.g., Hanning window, Hamming window) to reduce edge effects. - Normalization: Normalizes data to enable comparison between different datasets. b. Two-dimensional Fourier transform: - Calculate the two-dimensional Fourier transform using the Fast Fourier Transform (FFT) algorithm. - Spherical Harmonic Transform: For global data, a more accurate wavenumber decomposition can be achieved using the spherical harmonic transform. - Parallel FFT: Applying parallel FFT algorithms to large-scale data improves computational efficiency. c. PSD calculation: - Calculate power for each wavenumber: Calculate the PSD in the form |F(k)|^2 (F is the Fourier transform, k is the wavenumber) - Directional averaging: When anisotropy is not considered, the power of waves of the same wavenumber in different directions is averaged. - Logarithmic scaling: Displays the PSD on a logarithmic scale to visualize a wide range of power values. d. Identifying PSD intersections: - Numerically identify the intersection of the PSD curves of the observational data and the ESM data. - Robust intersection detection: Robust intersection detection algorithms to mitigate the effects of noise - Handling of multiple intersections: Selection criteria when there are multiple intersections (e.g., select the intersection with the fewest wavenumbers) e. Temporal smoothing: - Smoothing to obtain time-stable PSD analysis results - Exponential Moving Average: Calculates the exponential moving average of the new analysis results and past results. - Anomaly detection: Detects abnormal PSD analysis results and excludes them from smoothing. This analysis is performed on global data approximately every 10 to 30 seconds, ensuring that the latest spatial scale characteristics are always understood. Regional PSD analysis is also performed, enabling the selection of spatial scales appropriate to regional characteristics.

[0051] 2. Adaptive scale selection: The optimal spatial scale is dynamically selected according to observation conditions (cloud cover, atmospheric conditions, etc.). Specifically, the spatial scale is selected considering the following factors: a. Cloud cover and observation quality: - In cloud-covered areas, the reliability of visible light and near-infrared observations decreases, so a larger spatial scale (lower wavenumber) is selected. - Quantitative evaluation of cloud cover: The cloud coverage rate for each grid cell is calculated using a cloud masking algorithm. - Observation quality index: Evaluation of observation quality based on quality flags and signal-to-noise ratio (SNR) of each sensor. - Consideration of cloud type: Evaluation of observation quality according to cloud type (stratocumulus, cumulus, cirrus, etc.) - Integration of multi-wavelength observations: Comprehensive evaluation of observational quality by integrating observations from different wavelength bands. b. Atmospheric stability: - In regions with unstable atmospheres (e.g., regions with active convection), small-scale phenomena become important, so a smaller spatial scale (higher wavenumber) is selected. - Calculation of stability index: Calculate atmospheric stability index (e.g., CAPE, CIN, LI) from temperature and humidity profiles. - Detection of convective activity: Convective activity is detected from cloud top temperature, cloud top altitude, precipitation intensity, etc. - Front detection: Fronts are detected based on temperature gradients, changes in wind direction, etc. - Time evolution of atmospheric stability: Consider the temporal trend of changes in atmospheric stability. c. Precipitation intensity: - In areas with heavy precipitation, the spatial distribution of precipitation becomes complex, so a smaller spatial scale is selected. - Quantification of precipitation intensity: Measurement of precipitation intensity using microwave radiometers and rainfall radar. - Classification of precipitation types: Classification of stratiform precipitation and convective precipitation - Life stages of precipitation systems: Identifying life stages such as development, maturity, and decline. - Spatial heterogeneity of precipitation: Evaluation of the spatial coefficient of variation and local gradient of precipitation. d. Quality of observation sensors: - Adjust the spatial scale according to the quality of the sensor data (noise level, missing data, etc.) - Sensor-specific quality indicators: Evaluation of each sensor's operating status, calibration status, noise level, etc. - Data Missing Patterns: Evaluation of the spatial distribution and missing rate of missing data. - Detection of outliers: Detection and evaluation of outliers using statistical methods - Inter-sensor consistency: Evaluating the consistency of observations from different sensors. These factors are comprehensively evaluated to select the optimal spatial scale for each region. This selection is updated approximately every 1 to 5 minutes to constantly adapt to the latest observation conditions. The selected spatial scale is converted into the noise level (time step) in the CM's conditional sampling, which controls the resolution enhancement process. 3. Regional optimization: The scale is optimized according to regional characteristics (topography, land cover, etc.). Specifically, the scale is adjusted considering the following regional characteristics: a. Topographic complexity: - In areas with complex topography, such as mountainous regions, topographic effects become important, so a smaller spatial scale is selected. - Topographic complexity index: Quantifying topographic complexity from digital elevation models (DEMs). - Terrain gradient: Calculates the gradient (angle of slope) and direction of the terrain. - Topographic roughness: Calculate an index that represents the roughness of the terrain (e.g., standard deviation of elevation, relief). - Classification of topographic features: Classification of topographic features such as mountains, hills, plains, and basins. b. Non-uniformity of land cover: - In areas with uneven land cover, such as urban areas or areas where farmland and forests are mixed, local effects become important, so a smaller spatial scale is selected. - Land cover diversity index: Quantifies the diversity of land cover types (e.g., Shannon index, Simpson index) - Land cover boundary density: Calculate the length of the land cover boundary per unit area. - Urbanization level: Evaluate the proportion and distribution of urban areas. - Spatial variation of vegetation indices: Evaluating the spatial variation of vegetation indices such as NDVI. c. Land and sea distribution: - In coastal areas and other regions where land and sea coexist, land-sea contrast is important, so a smaller spatial scale is selected. - Coastline density: Calculate the length of coastline per unit area. - Ocean / Land Mask: Calculates the ratio of ocean to land in each grid cell. - Sea-land temperature contrast: Evaluating the difference between sea surface temperature and land surface temperature. - Sea and land breeze circulation: Evaluate the intensity and range of sea and land breeze circulation. d. Climate zones: - Adjust the scale according to climate zones such as tropical, mid-latitude, and polar regions. - Climate classification: Classification of climate zones based on the Köppen climate classification, etc. - Intensity of seasonality: Evaluates the magnitude of seasonal variations in temperature and precipitation. - Sensitivity to climate change: Consider regional differences in sensitivity to climate change. - Large-scale climate patterns: Assessing the impact of large-scale climate patterns such as El Niño-Southern Oscillation (ENSO) and North Atlantic Oscillation (NAO). These regional characteristics are pre-databased and combined with satellite positioning information to select the optimal scale. This optimization allows for the application of different scales to each region, resulting in high-resolution images that reflect local characteristics. It also takes into account temporal changes in regional characteristics (e.g., seasonal vegetation changes, urbanization), enabling adaptation to long-term changes. 4. Adapting to seasonal variations: The scale is adjusted according to seasonal changes in weather patterns. Specifically, the scale is adjusted considering the following seasonal variations: a. Monsoon: - In areas where precipitation patterns change significantly between monsoon and non-monsoon seasons, adjust the scale according to the season. - Monsoon index: Quantifying monsoon intensity based on seasonal changes in wind direction, precipitation, etc. - Monsoon start and end: Identify the start and end dates of the monsoon season. - Intra-monsoon variation: Identifying active and resting periods within the monsoon season. - Spatial distribution of monsoon precipitation: Evaluating the spatial distribution patterns of monsoon precipitation. b. Snow and ice cover: - In areas covered with snow and ice during winter, adjust the scale according to the presence or absence of snow and ice cover. - Snow and ice coverage rate: Calculate the snow and ice coverage rate for each grid cell. - Snow and ice conditions: Classification of snow and ice conditions such as fresh snow, transformed snow, and melted snow. - Snow and ice thickness: Estimate the thickness of snow and ice. - Reflectance of snow and ice: Measuring the reflectance (albedo) of snow and ice. c. Vegetation changes: - Adjust the scale according to seasonal changes in vegetation (e.g., leaf unfolding and leaf fall in deciduous broadleaf forests). - Seasonal changes in vegetation indices: Tracking seasonal changes in vegetation indices such as NDVI. - Phenological indicators: Identify phenological indicators such as germination date, flowering date, autumn foliage date, and leaf fall date. - Seasonal changes by vegetation type: Evaluate seasonal change patterns for each vegetation type, such as evergreen trees, deciduous trees, grasslands, and farmlands. - Vegetation evapotranspiration: Evaluating seasonal changes in vegetation evapotranspiration. d. Solar radiation conditions: - Adjust the scale according to seasonal changes in solar radiation conditions (e.g., polar night and midnight sun in polar regions). - Seasonal variations in solar radiation: Calculate seasonal variations in solar zenith angle and sunshine duration. - Seasonal variations in cloud cover: Evaluating seasonal variations in cloud cover. - Seasonal variations in atmospheric transmittance: Evaluating seasonal variations in atmospheric transmittance, such as aerosol optical thickness. - Seasonal variations in ground surface reflectance: Evaluating seasonal variations in ground surface reflectance. These seasonal variations are estimated from date and location data, and the optimal scale is selected according to the season. This approach enables consistent, high-quality, high-resolution imaging even in areas with significant seasonal variations. It also takes into account year-to-year variations (e.g., the difference between El Niño and La Niña years), allowing it to adapt to natural climate variability. 5. Feedback mechanism: We will evaluate the results of the resolution enhancement and use the feedback to guide scale selection. Specifically, we will provide feedback using the following evaluation metrics: a. Spatial consistency: - Revert the high-resolution results back to low resolution and evaluate the spatial agreement with the original ESM data. - Correlation coefficient: Calculate correlation indicators such as the Pearson correlation coefficient and the Spearman rank correlation coefficient. - Mean Squared Error (MSE): Calculate the MSE between the original ESM data and the high-resolution data that has been reverted to a lower resolution. - Structural Similarity Index (SSIM): Calculate the SSIM to evaluate the structural similarity of the images. - Spectral Match: Evaluates the similarity of power spectral densities. b. Statistical consistency: - Evaluate the statistical properties (histograms, spatial correlation structures, etc.) of the high-resolution results and confirm their consistency with the statistical properties of the observed data. - Similarity of distributions: Perform comparative distribution tests such as the Kolmogorov-Smirnov test and the Cramer-von Mises test. - Moment Comparison: Compare statistical moments such as mean, variance, skewness, and kurtosis. - Comparison of extreme value distributions: Comparing the parameters of the generalized extreme value distribution (GEV). - Comparison of spatial correlation structures: Comparison of spatial correlation indices such as variograms and Moran indices. c. Physical integrity: - Evaluate the physical consistency of the high-resolution results (e.g., conservation of mass, conservation of energy, etc.). - Conservation of mass: Evaluation of conserved quantities such as global total precipitation. - Energy balance: Evaluation of energy balance such as radiation balance and heat balance. - Consistency of physical relationships: Evaluate the consistency of physical relationships such as temperature and humidity, wind and pressure gradient. - Temporal continuity: Evaluation of physically valid time-dependent rate of change. Based on these evaluation results, the scale selection parameters are adjusted to improve the quality of high-resolution image processing. This feedback mechanism is executed approximately every 1 to 3 hours, enabling system self-optimization. Specifically, the system learns the optimal scale selection parameters based on the evaluation metrics and reflects them in future scale selections. This allows for the maintenance of high-quality high-resolution image processing over a long period. These features give the dynamically scaling adaptive system the following characteristics: 1. Adaptability: It adapts to various observation conditions, regional characteristics, and seasonal variations, always selecting the optimal spatial scale. This enables high-quality, high-resolution imaging under a variety of conditions. In particular, it can adaptively adjust the scale to rapidly changing weather conditions (e.g., the passage of weather fronts, the development of convective systems), resulting in consistently high-quality results over time. 2. Automation: The scale selection process is fully automated, selecting the optimal scale without human intervention. This ensures operational efficiency and consistency. The automated system operates 24 / 7, 365 days a year, always selecting the optimal scale. Furthermore, it can continuously learn from new observational data and evaluation results to improve the accuracy of scale selection. 3. Real-time capabilities: The system responds to changes in observation conditions in real time, always selecting the spatial scale best suited to the latest conditions. This enables appropriate high resolution even under rapidly changing weather conditions. Real-time processing is particularly important in predicting extreme weather events such as heavy rainfall and typhoons, and can improve the effectiveness of early warning systems.

[0052] 4. Reflection of regional characteristics: By applying different scales to each region, high-resolution imaging that reflects local characteristics is achieved. This allows for the appropriate representation of local influences such as topographic effects and land cover effects. Reflecting regional characteristics is particularly important for improving prediction accuracy in areas with complex topography (e.g., mountainous regions, coastal regions, etc.). 5. Self-optimization: The feedback mechanism continuously evaluates and optimizes system performance, enabling the maintenance of high-quality, high-resolution images over extended periods. The self-optimization function improves system robustness and adaptability, ensuring stable performance under various conditions. Thus, the dynamic scale adaptation system is a crucial component for ensuring the accuracy and reliability of high-resolution imaging under various conditions, and forms the foundation supporting the adaptability and robustness of the satellite system of the present invention. In particular, its ability to respond to diverse weather and climate conditions on a global scale and provide high-resolution data optimized for each region is a major advantage compared to conventional approaches. ### Multimodal Uncertainty Quantification Module The multimodal uncertainty quantification module provides an uncertainty assessment that integrates multiple sources of information. This module supports risk-based decision-making by comprehensively evaluating and providing uncertainty in high-resolution data to decision-makers. Specifically, the following features will be implemented: 1. Ensemble generation: We generate multiple high-resolution realizations from a single ESM output and evaluate the sampling uncertainty. Specifically, we generate the ensemble using the following process: a. Probabilistic sampling: - Run CM multiple times on the same ESM output using different random number seeds. - Ensuring diversity in random number seeds: Optimizing the method for selecting random number seeds (e.g., selecting a pseudo-random number generator, minimizing correlation between seeds) - Optimizing sampling efficiency: A sampling strategy that balances the required number of samples with computational costs. - Parallel sampling: Generates multiple samples in parallel to improve computational efficiency. - Adaptive sampling: The need for additional samples is determined based on the analysis of preliminary samples. b. Ensemble composition: - Construct an ensemble from multiple high-resolution realizations that have been generated. - Optimizing ensemble size: Selecting an ensemble size that balances prediction accuracy and computational cost. - Weighting of ensemble members: Weighting of ensemble members based on quality and diversity. - Sub-ensemble selection: Selecting sub-ensembles that maximize diversity - Multi-model ensemble: Integrates samples from different models and parameter settings. c. Calculation of statistical properties: - Calculate the statistical properties of the ensemble (mean, standard deviation, quantiles, etc.) - Robust statistics: Calculation of robust statistics (median, interquartile range, etc.) that mitigate the effects of outliers. - Spatial aggregation: Calculation of statistics at different spatial scales (e.g., grid cell units, watershed units, administrative district units) - Temporal aggregation: Calculation of statistics on different time scales (e.g., hours, days, months) - Conditional statistics: Calculation of conditional statistics for specific conditions (e.g., when precipitation exceeds a threshold). This ensemble generation leverages the probabilistic properties of CM to quantify the inherent uncertainty in the high-resolution process. For example, it can generate 100 to 1000 ensemble members and estimate the probability distribution of precipitation for each grid cell. This enables probabilistic predictions such as "there is an 80% probability of precipitation of 10 mm / day or more at a particular location." To improve the computational efficiency of ensemble generation, techniques such as batch processing and parallel processing are utilized. For example, batch processing, which generates multiple samples simultaneously in a single CM evaluation, and parallel processing, which generates multiple samples in parallel using multiple processor cores or FPGA blocks, are implemented. This allows for the efficient generation of large-scale ensembles within limited computational resources. 2. Sensor uncertainty integration: The evaluation will take into account the uncertainty of the observation sensor. Specifically, the sensor uncertainty will be integrated by considering the following factors: a. Sensor measurement error: - Uncertainty arising from the measurement accuracy and calibration errors of each sensor - Error model based on sensor specifications: Construction of an error model based on sensor specifications (e.g., sensitivity, dynamic range, quantization error) - Calibration curve uncertainty: Propagation of uncertainty in the sensor's calibration curve. - Temperature dependence: Evaluation of errors caused by the temperature dependence of the sensor. - Changes over time: Evaluation of errors caused by the deterioration of the sensor over time. b. Sampling error: - Uncertainty arising from spatiotemporal sampling of observations - Spatial sampling error: Sampling error due to finite spatial resolution - Time sampling error: Sampling error caused by finite time resolution - Observation geometry: Sampling error caused by observation angle and field of view. - Orbital sampling: Non-uniformity of sampling patterns due to satellite orbits. c. Representativeness error: - Error when estimating grid average values ​​from point observations - Scale gap: Error caused by the difference in scale between point observations and grid average values. - Spatial heterogeneity: Representativeness error caused by spatial heterogeneity of the observed object. - Terrain effects: Representativeness errors caused by terrain complexity - Land cover effect: Representativeness error due to the non-uniformity of land cover d. Algorithmic error: - Errors caused by the algorithm for estimating physical quantities - Model approximation error: Error caused by approximating physical processes. - Parameter uncertainty: Uncertainty of algorithm parameters - Uncertainty of input data: Propagation of uncertainty in the algorithm's input data. - Algorithm convergence: Errors related to the convergence of iterative algorithms These uncertainty factors are represented by different models for each sensor and integrated using Bayesian statistical methods. For example, the uncertainty in precipitation estimation from a microwave radiometer is modeled as a function dependent on precipitation intensity and taken into consideration when generating the ensemble. To integrate sensor uncertainty, the following approach is used: - Bayesian integration: Integration of information from different sensors based on Bayes' theorem. - Optimal interpolation method: Optimal interpolation considering observation error and background error. - Kalman filtering: Optimal integration of time series data - Multivariate uncertainty propagation: Multivariate uncertainty assessment using error propagation. - Monte Carlo simulation: Simulation-based evaluation of complex error structures This enables high-resolution data enhancement that explicitly takes into account the uncertainty of observational data, thereby providing a more reliable uncertainty assessment. 3. Spatiotemporal Correlation Analysis: We will analyze the spatiotemporal correlation structure of uncertainty. Specifically, we will analyze the following correlation structures: a. Spatial correlation: - Correlation of uncertainty between different locations - Estimation of spatial correlation functions: Estimation of spatial correlation functions such as variograms and covariograms. - Non-stationary spatial correlation: Estimation of location-dependent non-stationary spatial correlation structures - Anisotropic spatial correlation: Estimation of direction-dependent anisotropic spatial correlation structure - Distance decay of spatial correlation: Evaluation of the correlation decay pattern according to distance. b. Time correlation: - Correlation of uncertainty between different time steps - Estimation of autocorrelation function: Estimation of autocorrelation function and partial autocorrelation function - Long-term memory: Evaluation of long-term correlation (long-term memory). - Non-stationary time correlation: Estimation of time-dependent, non-stationary time correlation structures. - Periodic correlation: Evaluation of periodic correlation patterns such as diurnal and seasonal cycles. c. Inter-variable correlation: - Correlation of uncertainties between different weather variables - Inter-variable correlation matrix: Estimation of the correlation matrix between different variables - Conditional correlation: Evaluation of correlation between variables under specific conditions. - Nonlinear correlation: Evaluation of nonlinear correlations between variables - Evaluation of causality: Evaluation of causality using methods such as the Granger causality test. These correlation structures are estimated based on statistical analysis and theoretical models of the ensemble members. For example, spatial correlations are represented using variational methods or Gaussian process models, while temporal correlations are represented using autoregressive models or hidden Markov models. The following methods are used to analyze the correlation structure: - Empirical correlation analysis: Empirical correlations calculated directly from ensemble members. - Parametric models: Parametric correlation models such as exponential functions and Gaussian functions. - Nonparametric models: Nonparametric correlation models such as kernel density estimation. - Bayesian Hierarchical Models: Bayesian Hierarchical Models of Correlative Structures - Copra theory: Copra functions that represent complex dependency structures Considering these correlational structures allows for a more realistic uncertainty assessment. In particular, spatially correlated uncertainties are important in large-scale phenomena (e.g., flood risk assessment) and can provide a more realistic assessment than uncertainty assessments based on simple independence assumptions. 4. Multi-model integration: We perform uncertainty assessment by integrating predictions from multiple ESMs. Specifically, we perform multi-model integration using the following process: a. Predictive reception of multiple ESMs: - Receive predictions from multiple ESMs (e.g., POEM, GFDL-ESM4, SpeedyWeather.jl, etc.) - Evaluation of differences between models: Evaluation of differences in resolution, representation of physical processes, initial conditions, etc. - Assessment of systematic bias in the models: Identification and assessment of systematic bias in each model. - Evaluation of model interdependencies: Evaluation of interdependencies between models (e.g., common components and assumptions). - Evaluation of the model's past performance: Evaluation of the model based on past prediction accuracy. b. Individual high-resolution enhancement: - Improve the resolution of each ESM's prediction individually. - Model-specific optimization: Optimal scale selection according to the characteristics of each model. - Model-specific bias correction: Individual correction for systematic bias in each model. - Model-specific uncertainty assessment: Individual assessment of the uncertainty of each model's predictions. - Ensuring consistency between models: Adjustments to ensure consistency between different models. c. Multi-model ensemble configuration: - Integrate the high-resolution results and construct a multi-model ensemble. - Simple average: Equal weighted average of all models - Weighted average: Weighting based on the model's past performance and reliability. - Bayesian Model Mean (BMA): Model integration based on Bayesian statistics - Super Ensemble: Optimal linear combination based on training data - Nonlinear Integration: Nonlinear integration using machine learning methods

[0053] d. Assessment of structural uncertainty: - Evaluate structural uncertainties based on differences between models. - Inter-model variance: An uncertainty assessment based on the variability of predictions across models. - Model agreement: An assessment of the degree of agreement between models regarding a specific outcome. - Assessment of extreme forecasts: Identifying and evaluating the most extreme forecasts. - Scenario-based evaluation: An evaluation that treats different models as different scenarios. - Confidence weighting: Weighting of uncertainty based on the confidence level of the model. This multi-model integration enables robust predictions and uncertainty assessments that do not rely on a single ESM. For example, by integrating predictions from multiple ESMs in CMIP6 (the 6th-order coupled model intercomparison project), it is possible to quantify the uncertainty caused by differences between models. The advantages of multi-model integration are as follows: - Model bias offsetting: Biases from different models are offset, improving overall prediction accuracy. - Improved prediction robustness: Reduced impact from single model defects or extreme predictions. - Comprehensive assessment of uncertainty: A comprehensive assessment of uncertainty, including structural uncertainties in the model. - Realistic estimation of confidence intervals: Avoid overly confident (too narrow) confidence intervals. - Improved prediction skills: In many cases, multi-model prediction demonstrates higher skill than single-model prediction. 5. Confidence interval mapping: Visualize spatial confidence intervals. Specifically, visualize the following confidence interval information: a. Point estimate: - Point estimates such as ensemble mean and median - Maximum Likelihood Estimate: Point estimation based on the maximum likelihood estimation of the probability distribution. - Conditional expected value: Expected value under specific conditions - Optimal prediction: The prediction that minimizes a specific loss function. - Robust estimate: A robust estimate that reduces the impact of outliers. b. Confidence interval: - Confidence intervals such as 90% confidence interval and 95% confidence interval. - Percentile-based confidence interval: A confidence interval based on the percentiles of an empirical distribution. - Parametric confidence interval: A confidence interval based on assumptions about the probability distribution. - Bootstrap confidence interval: A confidence interval based on the bootstrap method. - Bayesian confidence intervals: Confidence intervals based on Bayesian estimation. c. Probability threshold: - Probability of exceeding a specific threshold - Probability of exceeding the danger threshold: The probability of exceeding the danger level (e.g., probability of precipitation of 50 mm / day or more) - Exceedance probability of composite thresholds: Exceedance probability for combinations of multiple conditions - Conditional excess probability: Excess probability under specific conditions - Time-integrated excess probability: Excess probability within a specific period. d. Spatial distribution of uncertainty: - Spatial distribution of the magnitude of uncertainty - Standard deviation map: Spatial distribution of the standard deviation of the forecast - Coefficient of variation map: Spatial distribution of the ratio of standard deviations to the mean. - Entropy map: A spatial distribution that represents the uncertainty of predictions using entropy. - Confidence map: Spatial distribution of prediction confidence (the inverse of uncertainty). - Uncertainty Hotspots: Identifying areas with particularly high levels of uncertainty. This information is overlaid on a map to facilitate intuitive understanding. For example, in the case of precipitation forecasts, the average value can be displayed on a color map, and the width of the confidence interval can be shown using transparency or contour lines. It is also possible to display the probability of exceeding a specific threshold (e.g., 50 mm / day) using different colors. The following approaches are used to represent confidence interval mappings: - Layered approach: Different information is displayed as different layers, allowing the user to select the information they need. - Combined visualization: Integrating multiple pieces of information into a single visual representation (e.g., using hue to represent the average value and saturation to represent confidence). - Interactive visualization: Interactive visualization that changes the displayed content in response to user actions. - Animation: Expressing changes over time as video. - 3D visualization: Facilitating intuitive understanding of complex information through three-dimensional representation. This allows for a spatial understanding of the uncertainty in high-resolution data and enables the evaluation of the reliability of predictions for each region. This is particularly important in evaluating phenomena with spatially heterogeneous uncertainty (for example, precipitation strongly influenced by topography). 6. Probabilistic verification: We will verify the accuracy of probabilistic predictions. Specifically, we will evaluate probabilistic predictions using the following validation metrics: a. Consecutive Rank Probability Score (CRPS): - An index for comprehensively evaluating the accuracy of probability predictions. - CRPS calculation: The square integral of the difference between the cumulative distribution function of predictions and the empirical distribution function of observations. - Normalized CRPS: Normalization to enable comparisons between different variables and regions. - Decomposed CRPS: Decomposed into reliability, resolution, and uncertainty components. - Conditional CRPS: CRPS assessment under specific conditions - Multivariate CRPS: An extension of CRPS for simultaneous prediction of multiple variables b. Briarscore: - An index for evaluating the accuracy of probability prediction for binary events (e.g., threshold exceedance). - Briar score calculation: The mean square of the difference between the predicted probability and the actual occurrence (0 or 1). - Briar score decomposition: Decomposes into components of reliability, resolution, and uncertainty. - Briar score reliability diagram: A diagram showing the relationship between predicted probability and actual occurrence frequency. - Multi-category Briar score: Briar score extension for multiple categories - Threshold Dependence: Evaluation of changes in the Briar score for different thresholds. c. Confidence diagram: - A diagram evaluating the relationship between predicted probability and actual occurrence frequency. - Confidence curve: A plot of actual occurrences against predicted probabilities. - Perfect confidence line: A diagonal line representing ideal confidence levels. - Confidence bias: Evaluation of systematic overestimation or underestimation - Resolution: Variation in frequency of occurrence between different predicted probability categories. - Sharpness: The degree of concentration in the predicted probability distribution. d. Rank histogram: - A figure for evaluating the statistical consistency of ensemble predictions. - Rank calculation: The rank of the observed value among the ensemble members. - Rank histogram creation: Histogram of the frequency distribution of ranks - Uniformity test: Statistical test of uniformity of rank histograms - Bias assessment: Bias assessment from the shape of the rank histogram - Variance assessment: Assessment of ensemble variance from the shape of the rank histogram. These verification results demonstrate the reliability of uncertainty assessment and also serve as feedback for continuous improvement of the system. For example, based on regional and seasonal evaluation results of CRPS and Briar score, weaknesses in uncertainty assessment under specific conditions can be identified and improved. The following approach is used to conduct probabilistic verification: - Cross-validation: Data is split into a training set and a validation set for evaluation. - Temporal verification: Evaluation by comparing past predictions and observations. - Spatial validation: Evaluating prediction accuracy in different regions. - Conditional validation: Evaluating prediction accuracy under specific weather conditions. - Multiscale validation: Evaluating prediction accuracy across different spatial and temporal scales. This allows for a comprehensive evaluation of the accuracy and reliability of probabilistic predictions, and enables the identification of the system's strengths and weaknesses. This is crucial for both the continuous improvement of the system and the provision of appropriate information to users. 7. Uncertainty propagation analysis: This study analyzes how uncertainty propagates to downstream applications (e.g., flood prediction, crop yield prediction). Specifically, it analyzes uncertainty propagation through the following process: a. Quantifying the uncertainty of high-resolution weather data: - Estimation of probability distribution: Estimate the probability distribution of weather variables for each grid cell. - Consideration of spatial correlation: Consideration of correlations between different grid cells. - Consideration of time correlation: Consideration of correlation between different time steps. - Consideration of inter-variable correlations: Consideration of correlations between different weather variables. - Conditional uncertainty: Evaluating how uncertainty changes under specific conditions. b. Supplying uncertain inputs to downstream impact models: - Hydrological models: Predicting runoff based on uncertain inputs such as precipitation and temperature. - Crop models: Predicting crop yields based on uncertainty in weather conditions - Disaster risk model: Assessment of disaster risk based on the uncertainty of extreme weather events - Economic model: Assessment of the impact of weather condition uncertainty on economic activity - Ecosystem models: Assessment of the impact of climate uncertainty on ecosystems. c. Track the propagation of uncertainty: - Monte Carlo simulation: A simulation using a large number of probabilistic samples. - Sensitivity analysis: Evaluating the sensitivity of the output to variations in the input. - Error propagation method: Analytical evaluation of error propagation based on linear approximation - Response surface method: Construction of a response surface that approximates the relationship between input and output. - Metamodeling: Building simplified models as alternatives to computationally expensive models. d. Quantify the uncertainty in the final impact assessment: - Estimation of probability distributions: Estimation of probability distributions of impact variables (e.g., flood level, crop yield) - Risk assessment: Calculation of risk indicators such as the probability of exceeding a threshold. - Decision support: Evaluating the optimal decision-making options while considering uncertainty. - Cost-benefit analysis: Evaluating the costs and benefits of countermeasures under uncertainty. - Robust decision-making: Identifying decisions that perform well even under various uncertain scenarios. This analysis helps us understand how uncertainty in weather forecasts translates into actual impacts and supports risk-based decision-making. For example, in the case of flood forecasting, it can analyze how uncertainty in precipitation forecasts propagates to uncertainty in flood level forecasts and provide a probabilistic assessment of flood risk. Examples of applications of uncertainty propagation analysis include: - Early flood warning: Probabilistic assessment of flood risk based on uncertainty in precipitation forecasts. - Agricultural management: Determining optimal planting times and irrigation schedules while considering the uncertainty of weather forecasts. - Renewable energy management: Probabilistic prediction of power generation based on prediction uncertainties of wind speed and solar radiation. - Public health: Assessing heatwave risk based on uncertainty in temperature forecasts - Traffic Management: Forecasting traffic demand and safety while considering the uncertainty of weather conditions. This allows end users to make appropriate decisions while taking uncertainty into account, mitigating damage from extreme events and enabling the efficient use of resources. These features give the multimodal uncertainty quantification module the following characteristics: 1. Inclusiveness: This module provides a comprehensive assessment that considers various sources of uncertainty (sampling uncertainty, sensor uncertainty, model uncertainty, etc.). This allows for a complete picture of the uncertainty. Traditional approaches often consider only specific sources of uncertainty, potentially leading to an underestimation of uncertainty. This module integrates multiple sources of uncertainty to provide a more realistic uncertainty assessment.

[0054] 2. Probabilistic representation: Instead of representing a single deterministic prediction, the prediction is expressed as a probability distribution. This explicitly communicates the uncertainty of the prediction and supports risk-based decision-making. For example, instead of a deterministic prediction like "Tomorrow's high temperature will be 30°C," providing a probabilistic prediction like "There is an 80% chance that tomorrow's high temperature will be above 30°C" allows users to make risk-aware decisions. 3. Multimodal integration: Integrating information from multiple sources (different sensors, different models, etc.) provides a more robust uncertainty assessment. This reduces the risks associated with relying on a single source. For example, integrating predictions from multiple ESMs allows for an assessment that takes into account the structural uncertainties of the model. Furthermore, integrating observations from different sensors can mitigate the effects of sensor-specific biases and errors. 4. Consideration of spatiotemporal correlations: This module provides an assessment that considers the spatiotemporal correlation structure of uncertainty. This enables a more realistic representation of uncertainty and is useful, for example, in risk assessment of widespread extreme events. Conventional approaches often assumed that uncertainty at different locations and times was independent, but in reality, spatial and temporal correlations exist. This module explicitly models these correlation structures, making it possible to assess, for example, the "probability of floods occurring simultaneously in multiple river basins." 5. Application-oriented: This module provides assessments that consider the impact of uncertainty in downstream applications (flood forecasting, crop yield forecasting, etc.). This enables the provision of uncertainty information directly relevant to final decision-making. While conventional approaches often limited uncertainty assessment to weather forecasts, this module tracks uncertainty propagation and assesses uncertainty in actual impacts. This allows for the provision of information directly relevant to decision-making, such as "the probability that the flood level at a specific location will exceed the danger level" or "the probability that crop yields will fall below 80% of the average year." 6. Dynamic adaptability: The uncertainty assessment method is dynamically adapted according to observation conditions and the target of prediction. This allows for the provision of optimal uncertainty assessments under various circumstances. For example, detailed probability distributions can be estimated in areas with abundant data, while more conservative assessments can be performed in areas with limited data, enabling adaptation to the situation. Furthermore, optimization is also performed according to the target of prediction, such as paying particular attention to the distribution shape of the right tail when predicting extreme phenomena. 7. Verification and Improvement: It has the ability to continuously improve uncertainty assessment methods based on the validation results of probabilistic predictions. This allows for improved prediction accuracy and reliability over the long term. Specifically, it evaluates probabilistic validation indices such as CRPS and Briar score by region, season, and phenomenon, identifying and improving weaknesses under specific conditions. In addition, it periodically updates the parameters of uncertainty models using new observational data and validation results. Thus, the multimodal uncertainty quantification module is a crucial component for comprehensively assessing uncertainty in high-resolution data and supporting risk-based decision-making. In particular, by integrating uncertainty from multiple sources and providing an assessment that considers spatiotemporal correlations, it enables more realistic risk assessments. Furthermore, by tracking uncertainty propagation, it can assess uncertainty in the final impact and provide information directly relevant to decision-making. ### Direct Distribution Communication System The direct distribution communication system provides the capability to deliver data directly to end users. This system enables the direct delivery of high-resolution climate data to users worldwide, including in areas with limited communication infrastructure. Specifically, the following features will be implemented: 1. Direct broadcast function: The data will be broadcast in a format that can be received by standard receivers. Specifically, the following broadcasting methods will be used: a. L-band broadcasting (1.5 to 1.6 GHz): - Uses frequency bands that can be received with small parabolic or planar antennas. - Data rate: Approximately 50 to 100kbps - Coverage: Wide area (field of view approximately 20 to 25 degrees, ground coverage diameter approximately 3000 to 4000 km) - Modulation method: Quadruple Phase Shift Keying (QPSK), Forward Error Correction (FEC) - Data format: Provides basic weather information (precipitation, temperature, wind speed, etc.) at low resolution. - Receiver requirements: A relatively inexpensive dedicated receiver or a general-purpose SDR (Software Defined Radio) receiver. - Power consumption: Satellite side approximately 10 to 20W, receiver side approximately 1 to 2W b. X-band broadcasting (7.8 to 8.4 GHz): - Use high-frequency bands to transmit larger amounts of data. - Data rate: Approximately 1 to 10 Mbps - Coverage: Mid-range (field of view approximately 10 to 15 degrees, ground coverage diameter approximately 1500 to 2500 km) - Modulation scheme: 8-phase shift modulation (8PSK), low-density parity check (LDPC) code - Data format: Detailed weather information (high-resolution precipitation, 3D atmospheric structure, etc.) - Receiver requirements: A dedicated receiver of moderate cost, and a parabolic antenna with a diameter of approximately 60 to 90 cm. - Power consumption: Satellite side approximately 30 to 50W, receiver side approximately 5 to 10W c. Ka-band broadcasting (25.5 to 27 GHz): - Uses ultra-high frequency bands to achieve extremely high data rates - Data rate: Approximately 10 to 100 Mbps - Coverage: Narrow area (field of view approximately 5 to 8 degrees, ground coverage diameter approximately 800 to 1200 km) - Modulation method: Quadrature amplitude modulation (16QAM / 32QAM), turbo code - Data format: Highest resolution weather information and combinations of multiple weather variables - Receiver requirements: High-cost dedicated receiver, high-precision parabolic antenna with a diameter of approximately 1 to 1.2m. - Power consumption: Satellite side approximately 50 to 80W, receiver side approximately 10 to 15W These broadcasts comply with international standards (e.g., DVB-S2 and CCSDS standards) and are provided in a format that can be received by commercially available receivers. The data format also complies with international standards (e.g., NetCDF, GRIB, etc.) and is provided in a format usable by common weather data processing software. The main advantage of direct broadcasting is that it allows data to be distributed without relying on terrestrial communication infrastructure. This enables access to high-resolution climate data even in areas with limited communication infrastructure or in areas where communication infrastructure has been damaged by disasters. Furthermore, because it is a broadcast-type communication, it can distribute data to a large number of users simultaneously, offering high scalability. The implementation of direct broadcast functionality presents the following technical challenges: - Rainfall attenuation: In particular, signal attenuation due to rainfall is significant in the Ka band, so adaptive modulation / coding and securing rainfall margins are necessary. - Frequency regulation: Frequency allocation and regulation in accordance with International Telecommunication Union (ITU) regulations. - Power constraints: Efficient allocation of transmit power within the limited power resources of a satellite. - Antenna design: Design of a multiband antenna that covers multiple frequency bands. - Popularization of receivers: Development and widespread adoption of low-cost, user-friendly receivers. To address these challenges, we will adopt the following technologies: - Adaptive Modulation and Coding (ACM): Dynamically adjusts the modulation scheme and coding rate according to reception conditions. - Beamforming: Forms multiple beams to transmit the optimal signal for each region. - Hierarchical modulation: Providing different quality levels of service with a single carrier wave. - Time-division multiplexing: Allocating time slots to accommodate different regions. - Open source receivers: Development of open-source receiving software that runs on low-cost, general-purpose hardware. 2. Adaptive data compression: The data compression ratio is dynamically adjusted according to the communication bandwidth. Specifically, the following adaptive data compression techniques are employed: a. Hierarchical data structure: - Structure the data at multiple resolution levels and select the appropriate resolution based on available bandwidth. - Basic layer (low resolution): Basic information delivered to all users. - Extension Layer 1 (Medium Resolution): Additional information delivered to users with sufficient bandwidth. - Extended Layer 2 (High Resolution): Detailed information delivered to users with high bandwidth. - Dependencies between layers: Higher layers depend on lower layers and can be decrypted step by step. - Resolution scalability: Gradually improve spatial resolution. - Time scalability: Gradually improving time resolution. - Quality scalability: Gradually improve data quality (e.g., quantization accuracy). b. Adaptive quantization: - Adjust the quantization level according to the importance of the data. - Importance Assessment: Assessment of importance based on factors such as the significance of the phenomenon (e.g., extreme events), the uncertainty of the prediction, and user interest. - Spatial adaptive quantization: Quantization is performed with high precision in spatially important regions (e.g., complex terrain, densely populated areas) and low precision elsewhere. - Variable adaptive quantization: Important variables (e.g., precipitation) are quantized with high precision, while others are quantized with low precision. - Time-adaptive quantization: Quantizes with high precision during critical time periods (e.g., initial prediction time) and with lower precision during other periods. - Perceptual quantization: Quantization that takes into account human perceptual characteristics (e.g., preserving visually important features). c. Context-adaptive entropy coding: - Select the optimal encoding method according to the statistical characteristics of the data. - Context Modeling: Building context models to capture the local statistical characteristics of data. - Arithmetic coding: Highly efficient arithmetic coding based on context models - Huffman coding: Frequency-based variable-length code assignment - Run-length coding: Efficient coding of consecutive identical values - Dictionary-based compression: Dictionary-based compression of repeating patterns (e.g., LZ77, LZ78) - Mixed models: Adaptive combinations of multiple compression models d. Predictive coding: - Reduce data redundancy using temporal and spatial predictive models. - Time forecast: Forecast based on the value of the previous time step. - Spatial prediction: Prediction based on the values ​​of neighboring grid cells - Spatiotemporal prediction: Predictions that consider both time and space. - Physically based prediction: Prediction based on a physical model (e.g., advection model) - Machine learning-based prediction: A predictive model learned from historical data. - Residual coding: Encodes only the difference (residual) between the predicted value and the actual value. By combining these technologies, the data compression ratio can be dynamically adjusted within a range of 10:1 to 100:1 depending on the communication bandwidth. This makes it possible to efficiently transmit necessary information even with limited communication bandwidth. The main advantage of adaptive data compression is that it enables optimal data delivery under various communication conditions. For example, in areas with limited bandwidth, only basic information can be delivered, while in areas with abundant bandwidth, detailed information can also be delivered. Furthermore, even when communication conditions change over time (for example, signal attenuation due to rainfall), the compression ratio can be adjusted according to the conditions.

[0055] The following technical challenges exist in implementing adaptive data compression: - Computational load of compression and decompression: Minimizing the computational load of decompression processing on the receiving end, in particular. - Error tolerance: Ensuring that compressed data is resistant to communication errors. - Adaptive control: Adaptive control of compression parameters in response to changes in communication conditions. - Compatibility: Ensures compatibility with various receivers. - Real-time performance: Real-time execution of compression processing. To address these challenges, we will adopt the following technologies: - Hardware acceleration: Speeding up compression and decompression processes with dedicated hardware. - Error protection: Additional error protection for critical data. - Feedback control: Adjustment of compression parameters based on feedback from the receiving end. - Standards compliance: Adoption of data formats and protocols that comply with international standards. - Pipelining: Ensuring real-time performance through pipelining of compression processing. 3. Priority-based delivery: Prioritize the delivery of data with high urgency. Specifically, set the following priority criteria: a. Priority of extreme phenomena: - Data related to extreme weather events such as heavy rain, strong winds, high temperatures, and low temperatures will be delivered with the highest priority. - Threshold-based detection: Detection of extreme events based on predefined thresholds. - Statistical-based detection: Detection of extreme events based on statistical distributions (e.g., exceeding the 90th percentile). - Impact-based detection: Detection of extreme events based on the expected magnitude of their impact. - Multi-hazard detection: Detection of combinations of multiple hazards (e.g., combination of high temperature and dryness). - Detection of rapidly developing phenomena: Detection of phenomena that develop rapidly (e.g., rapidly developing low-pressure systems) b. Regional vulnerabilities: - Data from areas vulnerable to specific hazards, such as areas at high risk of flooding or landslides, will be delivered with high priority. - Vulnerability Map: A map that assesses the vulnerability to hazards in each region. - Exposure assessment: Assessment of exposure factors such as population and critical infrastructure. - Past disaster history: Vulnerability assessment based on past disaster occurrence history - Socioeconomic vulnerability: Vulnerability assessment based on socioeconomic factors such as income level and access to healthcare. - Adaptive capability: Vulnerability assessment based on adaptive capabilities such as response and recovery capabilities. c. Time urgency: - Data related to phenomena with short predicted lead times (e.g., heavy rainfall occurring in a few hours) will be delivered with high priority. - Lead time evaluation: Evaluation of the remaining time until the phenomenon occurs. - Response time requirements: Evaluation of the minimum time required for an effective response. - Rate of change in prediction: Detection of rapid changes in prediction (e.g., significant updates to predictions) - Threshold approach speed: Evaluation of the speed of approach to the danger threshold. - Cascade effect: Evaluation of a chain reaction effect in which one phenomenon triggers other phenomena. d. User requests: - Data based on requests from specific users is delivered with priority according to the urgency of the request. - Request classification: Request classification such as urgent, high priority, normal, and low priority. - User classification: User classifications such as disaster prevention organizations, critical infrastructure managers, and general users. - Request frequency limit: A frequency limit to prevent system overload caused by excessive requests. - Request aggregation: Efficiency improvements through the aggregation of similar requests. - Conditional request: A request that delivers data only if specific conditions are met. Based on these priorities, data packets are tagged with priority tags and processed preferentially at the communication protocol level. This ensures that the most important information is delivered reliably, even in situations with limited communication bandwidth. The main advantage of priority-based delivery is that it allows for the most effective use of limited communication resources. In particular, during disasters or other situations where communication resources are constrained, it is crucial to prioritize the delivery of information that is most important for protecting lives and property. Furthermore, even during normal times, prioritizing the delivery of information tailored to user needs can improve user satisfaction. Implementing priority-based delivery presents the following technical challenges: - Automated priority evaluation: Automated priority evaluation considering various factors. - Dynamic priority adjustment: Dynamic adjustment of priorities in response to changing circumstances. - Ensuring fairness: Preventing excessive delays in low-priority data. - Priority propagation: Consistent priority handling across the entire communication protocol stack. - Congestion control: Prioritized and effective congestion control To address these challenges, we will adopt the following technologies: - Machine learning-based priority evaluation: Prioritization models based on historical data and expert knowledge. - Adaptive Priority Scheduling: Adaptive scheduling based on system load and priority. - Hierarchical Equitable Queuing: Queuing that ensures fairness between different priority levels - End-to-end priority processing: Consistent priority processing across the entire communication stack. - Predictive congestion avoidance: Preventive congestion control based on predictions of communication load. 4. Regional customization: We will deliver data products tailored to local needs. Specifically, we will implement the following regional customizations: a. Optimization by climate zone: - We provide data products that focus on the most relevant weather variables and phenomena according to climate zones such as tropical, mid-latitude, and polar regions. - Tropical regions: Focus on tropical cyclones, monsoons, convective precipitation, etc. - Mid-latitude regions: Focus on extratropical cyclones, fronts, blocking highs, etc. - Polar Regions: Focusing on sea ice, blizzards, polar vortices, etc. - Arid regions: Focus on droughts, dust storms, heat waves, etc. - Mountainous regions: Focus on orographic precipitation, avalanches, fog, etc. b. Industry-specific optimization: - We provide data products tailored to the main industry of each region, such as regions where agriculture is the main industry, regions where fisheries are the main industry, and regions where tourism is the main industry. - For agriculture: Agricultural variables such as precipitation, temperature, solar radiation, soil moisture, and evapotranspiration. - For fisheries: Ocean-related variables such as sea surface temperature, ocean currents, wave height, wind direction and speed. - For tourism: Tourism-related variables such as weather forecast, UV index, perceived temperature, and visibility. - For energy: Energy-related variables such as wind speed, solar radiation, temperature, and humidity. - For traffic: Traffic-related variables such as visibility, wind speed, road surface temperature, and precipitation type. c. Optimization by disaster risk: - We provide data products tailored to major disaster risks in a region, such as areas with a high risk of flooding, drought, or typhoon. - Flood-risk areas: Flood-related variables such as precipitation, river flow, soil moisture, and snow cover. - Drought-risk areas: Drought-related variables such as precipitation, evapotranspiration, soil moisture, and vegetation conditions. - Typhoon risk areas: Typhoon-related variables such as wind speed, atmospheric pressure, wave height, and storm surge. - Areas at risk of sediment-related disasters: Sediment-related variables such as precipitation, soil moisture, and slope gradient. - Forest fire risk areas: Fire-related variables such as temperature, humidity, wind speed, and vegetation dryness. d. Language and culture optimization: - Provide data representations and metadata appropriate to the local language and culture. - Multilingual support: Metadata and descriptions in the main language of the region. - Cultural context: Information representation appropriate to the local cultural context - Local unit system: A unit system commonly used in a region (e.g., metric system, imperial system). - Regional time zone: Time displayed according to the regional standard time zone. - Local calendar: Displays dates according to the local calendar. These customizations are automatically applied based on regional user profiles, providing the most relevant information to users in each region. The main advantage of regional customization is the ability to efficiently provide users with the most relevant information. This improves the usefulness and availability of information, effectively supporting users' decision-making. Furthermore, customization tailored to regional characteristics allows for efficient use of limited communication bandwidth. Implementing regional customization presents the following technical challenges: - Regional profiling: Accurately understanding the characteristics and needs of each region. - Dynamic boundaries: Handling cases where regional boundaries are not clearly defined. - Scalability: Managing individual customizations for multiple regions - User feedback: Adapting to changing local needs - Content Management: Efficient management of diverse customizable content To address these challenges, we will adopt the following technologies: - Data-driven profiling: Building regional profiles based on usage patterns and feedback - Fuzzy boundary processing: Fuzzy boundary processing that enables gradual transitions between regions. - Template-based customization: Customization based on reusable templates - Continuous user research: Continuous research to understand changes in local needs. - Content Management System: An efficient management system for diverse and customizable content. 5. Two-way communication: We will deliver data in response to user requests. Specifically, we will implement the following bidirectional communication functions: a. Request-based delivery: - Users request data related to specific regions, times, and variables, and the data is delivered accordingly. - Geographical range specification: Specification of a geographical range by latitude / longitude range or administrative area. - Time range specification: Specify a time range by start time and end time. - Variable selection: Select the necessary weather variables. - Resolution specification: Specify the required spatial and temporal resolution. - Format specification: Specify the required data format. - Priority specification: Specify the priority of the request. b. Feedback mechanism: - Collect user feedback (data usefulness, accuracy, etc.) and use it to improve the system. - Quantitative evaluation: Evaluation of the usefulness and accuracy of data using numerical scales. - Qualitative comments: Detailed feedback through free-form comments. - Problem Report: Report data problems or bugs. - Improvement Suggestions: Suggestions for improving the system - Sharing use cases: Sharing actual use cases of the data. - Automated usage statistics: Automatic collection of data usage patterns. c. Alert Subscription: - Users can subscribe to alerts based on specific conditions (e.g., forecasts of 50mm or more of precipitation per day in a particular area) and receive notifications when those conditions are met. - Condition Definition: Defining alert conditions such as thresholds, geographical ranges, and time ranges. - Notification method: Choose your notification method, such as email, SMS, or app notification. - Notification frequency: Settings for notification frequency and duplicate filtering. - Priority setting: Setting priorities for different alerts - Expiration Date: Set the expiration date for alert subscriptions. - Combination of conditions: A logical combination of multiple conditions (AND, OR, NOT, etc.)

[0056] d. Data quality information: - Users request data quality information (uncertainty, verification results, etc.) and receive information to assess the reliability of the data. - Uncertainty indicators: Indicators that show the uncertainty of predictions. - Verification statistics: Verification statistics based on comparisons between past predictions and observations. - Quality Flags: Flags and notes regarding data quality. - Source metadata: Information about the data source and processing history. - Comparison information: Comparison information with different models and data sources. - Known limitations: Known limitations and considerations regarding the data. These two-way communications are implemented via low-bandwidth uplink connections (e.g., cellular networks or the internet), enabling customized data delivery tailored to user needs. The main advantage of two-way communication is that it enables data delivery tailored to the specific needs of users. This allows users to efficiently obtain only the information they need, preventing information overload. Furthermore, collecting user feedback enables continuous improvement of the system. The implementation of bidirectional communication presents the following technical challenges: - Uplink constraints: Efficient communication on low-bandwidth, high-latency uplink connections. - Request processing capability: The ability to efficiently process a large number of requests. - Security: Protection from unauthorized access and excessive requests - Offline support: Support for situations where the upload connection is unavailable. - User Interface: An interface that is easy for a diverse range of users to use. To address these challenges, we will adopt the following technologies: - Lightweight protocol: A lightweight communication protocol optimized for low-bandwidth, high-latency environments. - Distributed request processing: Distributed request processing across multiple satellites and ground stations. - Authentication and authorization: Ensuring security through appropriate authentication and authorization. - Caching and pre-delivery: Caching and pre-delivery of common requests - Multi-platform support: Interface compatible with various devices and platforms. 6. Redundancy and Reliability: Ensure communication redundancy and reliability. Specifically, implement the following redundancy and reliability features: a. Use of multiple frequency bands: - Uses multiple frequency bands such as the L-band, X-band, and Ka-band to maintain communication even if some frequency bands are affected by interference or malfunctions. - Frequency diversity: Use of multiple frequency bands with different propagation characteristics. - Adaptive frequency selection: Selection of the optimal frequency band according to propagation conditions. - Inter-frequency handover: Seamless switching between frequency bands. - Frequency reuse: Reusing the same frequency in spatially separated areas. - Interference avoidance: Transmit scheduling that minimizes interference between frequency bands. b. Temporal redundancy: - Important data will be sent repeatedly in multiple time slots to ensure reliable reception even in the event of temporary communication failures. - Importance-based iteration: Adjusting the number of transmissions based on the importance of the data. - Temporal distribution: Sending messages distributed across different time periods. - Interleaving: Improved resistance to burst errors through interleaving of data blocks. - Incremental update: Incremental update sends only the difference from the previous transmission. - Acknowledgment-based retransmission: Automatic retransmission if no acknowledgment is received. c. Spatial redundancy: - By transmitting the same data from multiple satellites, data can be received even when communication with a specific satellite is difficult. - Satellite diversity: Transmission from multiple satellites in different orbits and positions. - Geographic redundancy: transmission from different geographical locations - Beam redundancy: Redundant coverage using different beam patterns - Handover: Seamless handover between satellites - Mesh network: A mesh network between satellites and ground stations. d. Correction of forward errors: - Using powerful error correction codes, data can be accurately recovered even in the event of communication errors. - Low-density parity check (LDPC) codes: High performance and efficient error correction capability - Turbo codes: High error correction capability through iterative decoding. - Reed-Solomon coding: Codes that are robust against burst errors. - Concatenated code: The concatenation of multiple codes with different characteristics. - Adaptive coding: Adaptive adjustment of the coding rate according to channel conditions. - Non-uniform error protection: Non-uniform error protection based on the importance of the data. These redundancy and reliability features ensure stable communication under various conditions and reliable delivery of important weather information. The main advantage of redundancy and reliability is that it enables stable data delivery even in situations where communication conditions deteriorate (e.g., signal attenuation due to rain, limited satellite visibility, etc.). This is particularly important in situations where communication conditions deteriorate when communication is at its most critical, such as during disasters. In addition, multiple redundancy functions can eliminate single points of failure and improve the overall availability of the system. Implementing redundancy and reliability presents the following technical challenges: - Resource efficiency: The balance between redundancy and resource efficiency - Complexity Management: Integrated management of multiple redundant functions - Ensuring consistency: Ensuring consistency between redundant data. - Fault detection: Rapid detection of communication failures. - Recovery mechanism: Efficient recovery from failures To address these challenges, we will adopt the following technologies: - Adaptive redundancy control: Adaptive control of redundancy levels based on communication conditions and data importance. - Integrated Redundancy Management: A framework for managing multiple redundancy functions in an integrated manner. - Version control: Version control ensures consistency of redundant data. - Proactive monitoring: Early fault detection through continuous monitoring of communication quality. - Self-healing mechanism: Automatic recovery mechanism from detected faults 7. Security and Privacy: We will protect data security and user privacy. Specifically, we will implement the following security and privacy features: a. Data encryption: - Highly sensitive data is protected with appropriate encryption algorithms. - Symmetric encryption: Fast symmetric encryption algorithms such as AES-256 - Public-key cryptography: Public-key cryptographic algorithms such as RSA and ECDH - Key management: The secure process of key generation, distribution, updating, and destruction. - End-to-end encryption: Encryption across the entire communication path. - Selective encryption: Selective encryption based on the level of data confidentiality. - Lightweight cryptography: Lightweight cryptographic algorithms for resource-constrained environments b. Access control: - Implement data access control based on user permissions. - Role-based access control: Managing access rights based on user roles - Attribute-based access control: Flexible access control based on user attributes and data attributes. - Multi-factor authentication: Enhanced access control through multiple authentication factors - Single sign-on: Unified authentication for multiple services - Access Logs: Detailed logging of data access - Anomaly detection: Detection of access patterns that are different from normal. c. Anonymization: - Anonymize user request information to protect privacy. - Data minimization: Collect only the minimum necessary data. - Kana encoding: Replace direct identifiers with kana characters. - k-anonymity: Ensures that users of k or more cannot be distinguished from one another. - Differential Privacy: Privacy protection through the addition of statistical noise - Aggregation: Aggregating and providing individual data. - Location blurring: Provides location information with a blurred appearance. d. Secure Boot: - The communication system's firmware implements secure boot with signature verification to prevent the execution of malicious code. - Hardware trust root: The starting point of hardware-based trust. - Bootloader Verification: Verification of the bootloader using a digital signature. - Firmware Verification: Firmware verification using digital signatures. - Runtime integrity verification: Continuous verification of code integrity at runtime. - Secure Update: A safe firmware update process - Rollback protection: Prevents unauthorized firmware rollbacks. These security and privacy features ensure data confidentiality and integrity, maintaining user trust. The main security and privacy benefits are the ability to protect sensitive weather data and user information. This allows users to use the system with peace of mind and meets the needs of highly confidential users, especially government agencies and corporations. It also facilitates compliance with regulatory requirements (e.g., privacy regulations such as GDPR). Implementing security and privacy presents the following technical challenges: - Computational overhead: Computational overhead for security features such as encryption. - Key management: The complexity of secure key management - Usability: A balance between security and ease of use - Responding to new threats: Continuously responding to evolving security threats - Regulatory compliance: Compliance with regulatory requirements in various countries and regions. To address these challenges, we will adopt the following technologies: - Hardware acceleration: Speeding up cryptographic processing with dedicated hardware. - Secure element: Dedicated hardware for the secure storage and handling of key materials. - User-centered design: A design approach that balances security and ease of use. - Threat Intelligence: Continuous collection and analysis of the latest security threat information. - Privacy Impact Assessment: Assessment and countermeasures for privacy impacts from the design stage. These features give the direct distribution communication system the following characteristics: 1. Global access: This system directly provides high-resolution climate data to users worldwide, including those in areas with limited communication infrastructure. This will bridge the gap in access to global climate information and improve climate change adaptation capabilities, particularly in developing countries and remote regions. Traditional approaches have primarily limited access to high-resolution climate data available to meteorological agencies and research institutions in developed countries, leaving developing countries and remote areas without sufficient access. This system, by distributing data directly from satellites, makes it possible to provide high-resolution climate data to users worldwide, regardless of the state of communication infrastructure. 2. Adaptive delivery: The system adaptively adjusts the content of data delivery according to communication bandwidth, user needs, and data urgency. This maximizes the use of limited communication resources and ensures the reliable delivery of the most important information. Conventional approaches typically use fixed data formats and delivery schedules, making it difficult to adapt to various conditions. This system achieves optimal data delivery adapted to various conditions through technologies such as adaptive data compression, priority-based delivery, and regional customization.

[0057] 3. User-centered design: We provide data products and interfaces tailored to users' needs and capabilities. This maximizes the usefulness and availability of data, providing information that is helpful for actual decision-making. Traditional approaches often involve data products aimed at users with technical expertise, making them difficult for general users to use. This system provides data products that meet the diverse needs of users through features such as regional customization, two-way communication, and user feedback. 4. High reliability: Redundancy and error correction capabilities ensure stable communication under various conditions. This allows for the reliable delivery of important weather information, especially during disasters or other situations where communication environments deteriorate. Conventional approaches carried the risk of information delivery being interrupted due to poor communication environments. This system significantly improves communication reliability through technologies such as the use of multiple frequency bands, temporal redundancy, spatial redundancy, and forward error correction. 5. Interactivity: This system enables two-way communication based on user requests and feedback. This allows for customized data delivery tailored to user needs and continuous system improvement. Traditional approaches primarily rely on one-way broadcast-style communication, making it difficult to address specific user needs. This system achieves two-way communication with users through features such as request-based delivery, feedback mechanisms, alert subscriptions, and data quality information. 6. Security and Privacy: This system protects data security and user privacy. This allows for the protection of sensitive weather data and user information, maintaining user trust. Traditional approaches sometimes lacked sufficient security and privacy considerations. This system achieves a high level of security and privacy protection through technologies such as data encryption, access control, anonymization, and secure boot. Thus, direct distribution communication systems are a crucial component for directly providing high-resolution climate data to users worldwide, significantly improving access to climate information, especially in areas with limited communication infrastructure. Furthermore, features such as adaptive distribution, user-centric design, high reliability, two-way communication, security, and privacy enable effective data distribution that meets the diverse needs of users. ### Real-time data assimilation engine The real-time data assimilation engine provides the ability to instantly integrate observational data and ESM forecasts. This engine is a core component for minimizing the time gap between observation and forecast and improving forecast accuracy. Specifically, the following features will be implemented: 1. Variational data assimilation: We will implement data assimilation using the 4D variational method. Specifically, variational data assimilation will be performed using the following process: a. Definition of the cost function: - Define a cost function based on the difference between observations and model predictions, and the deviation from the background field. - Observation term: The sum of squared differences between observed values ​​and model predicted values ​​(weighted by the observation error covariance matrix) - Background term: Sum of squared deviations from the background field (initial estimate) (weighted by the background error covariance matrix) - Constraint terms: Terms relating to physical constraints (e.g., conservation of mass, conservation of energy). - Regularization term: A regularization term used to ensure the smoothness and physical validity of the solution. - Time integration: In the 4-dimensional variational method, time integration takes into account all observations within the time window. - Weighting: Weighting based on the reliability and importance of the observations. b. Gradient calculation: - Implement an adjoint model to compute the gradient of the cost function. - Adjoint model: An adjoint model based on the linearization and transposition of the original model. - Efficient implementation: Optimization to improve the computational efficiency of the adjoint model. - Checkpoint Strategy: A checkpoint strategy that considers the balance between memory usage and computational complexity. - Automatic differentiation: Automatic differentiation technique for the automatic generation of adjoint models - Approximate adjoints: Approximate adjoint models for reducing computational cost - Parallel computing: Speeding up gradient calculations through parallelization. c. Optimization: - Minimize the cost function using optimization algorithms such as the conjugate gradient method. - Conjugate gradient method: An iterative optimization algorithm effective for nonlinear optimization problems. - Quasi-Newton method: A fast convergence algorithm based on the approximation of Hessian matrices - Limited-memory BFGS method: A variant of the quasi-Newton method suitable for large-scale problems - Trust region method: A trust region approach that improves the stability of optimization. - Preprocessing: Preprocessing techniques to accelerate convergence - Early termination: Early termination criteria that consider the balance between computational efficiency and accuracy. d. Generation of the analysis field: - Generate optimal initial conditions (analysis field) based on the optimization results. - State vector update: Updating the state vector based on the optimization results. - Balance adjustment: Adjustments to ensure physical balance (e.g., geostrophic balance) - Smoothing: Smoothing to remove small amounts of noise. - Physical consistency verification: Verification of the physical consistency of the analysis field. - Extreme value processing: Processing of non-physical extreme values - Uncertainty estimation: Estimation of uncertainty in the analysis field. This variational data assimilation is particularly effective when continuous observational data is available, and it can generate a physically consistent analytical field. For example, assimilating atmospheric temperature and humidity profiles obtained from satellite observations can improve the thermodynamic structure of the atmosphere and enhance the accuracy of precipitation forecasts. The main advantages of variational data assimilation are its ability to generate a physically consistent analysis field and to integrate various types of observational data. It is particularly well-suited for assimilation of indirect observations (e.g., estimating temperature profiles from brightness temperature). Furthermore, it can directly handle the nonlinearity of observational operators. The following technical challenges exist in implementing variational data assimilation: - Computational cost: Computational cost of developing and running the adjoint model - Limitations of linearization: Limitations of linearization in systems with strong nonlinearity - Estimation of error covariance: The difficulty of accurately estimating the background error covariance matrix. - Local Optimal Solutions: Problems involving local optima in non-convex optimization problems - Memory requirements: Memory requirements for large-scale problems To address these challenges, we will adopt the following technologies: - Incremental approach: Incremental variational assimilation with outer and inner loops - Hybrid covariance: A combination of climatological covariance and ensemble-based covariance. - Multiscale optimization: Stepwise optimization at different scales - Parallel computing: Improving computational efficiency through large-scale parallel computing. - Model reduction: Model reduction techniques to improve computational efficiency 2. Ensemble Kalman filter: We will implement ensemble-based data assimilation. Specifically, we will perform an ensemble Kalman filter using the following process: a. Ensemble forecast: - Run ESM on multiple initial conditions to generate a forecast ensemble. - Initial perturbation generation: Method for generating initial ensemble members (singular vectors, bleeding vectors, etc.) - Ensemble size: Choosing an ensemble size that balances computational cost and statistical reliability. - Multi-model ensemble: An ensemble using different models and physical parameterizations. - Time integration: Time integration using forecasting models - Stochastic physical processes: Stochastic parameterization to express uncertainty in physical processes - Ensemble management: Maintaining ensemble diversity and representativeness. b. Estimation of error covariance: - Estimate the forecast error covariance from the variance of ensemble members. - Sample covariance: Sample covariance calculated directly from the ensemble members. - Localization: Localization to suppress spurious correlations at long distances. - Expansion: Expansion to correct for underestimation of ensemble variance. - Sparse approximation: Efficient covariance representation for large-scale problems - Adaptive adjustment: Adaptive adjustment of localization and expansion parameters based on data. - Low-rank approximation: Efficient low-rank approximation of large covariance matrices c. Calculation of Kalman gain: - Calculate the Kalman gain from the forecast error covariance and the observation error covariance. - Matrix operations: Efficient implementation of matrix operations - Observation operator: Transformation from model space to observation space - Observation error covariance: A representation of the statistical characteristics of observation errors. - Parallel computing: Parallelization of Kalman gain calculation - Approximation: Approximation techniques for large-scale problems - Numerical stability: Numerically stable Kalman gain calculation d. Updating the analysis field: - Update the forecast field using Kalman gain to bring it closer to the observation. - Ensemble Update: Updates for each ensemble member. - Stochastic perturbation: A stochastic representation of observation noise - Balance adjustment: Adjustments to ensure physical balance. - Non-negative constraint: Applying constraints to non-negative variables such as precipitation. - Sequential processing: An implementation that processes observations sequentially. - Parallel update: Parallel update of ensemble members e. Re-sampling of the ensemble: - Resample the ensemble to reflect the analysis error covariance. - Deterministic ensemble transformation: Ensemble transformation using a deterministic method - Probabilistic resampling: Ensemble resampling using probabilistic methods - Prevention of filter divergence: Techniques to prevent filter divergence - Maintaining diversity: Techniques for maintaining diversity in an ensemble - Handling abnormal members: Handling ensemble members that are physically impossible to handle. - Ensemble resizing: Dynamic adjustment of ensemble size for computational efficiency This ensemble Kalman filter is robust to the nonlinearity and non-Gaussianity of the model and can naturally represent forecast uncertainties. For example, by assimilating precipitation radar observations, the location and intensity of precipitation systems can be corrected, improving the accuracy of short-term precipitation forecasts. The main advantages of ensemble Kalman filters are their robustness to model nonlinearity and their ability to naturally represent forecast uncertainty. In particular, ensemble forecasting directly provides probabilistic predictions and allows for the quantification of forecast uncertainty. Furthermore, it does not require the development of adjoint models and is relatively easy to implement.

[0058] The implementation of an ensemble Kalman filter presents the following technical challenges: - Sampling error: Sampling error due to limited ensemble size - Far-field spurious correlation: Spurious correlation at a distance due to a limited ensemble size. - Filter divergence: Filter divergence due to model errors and nonlinearity. - Computational cost: Computational cost for multiple ensemble members - Non-Gaussianity: Handling variables with strong non-Gaussianity (e.g., precipitation) To address these challenges, we will adopt the following technologies: - Localization and Expansion: Localization and expansion to reduce sampling error and far-field spurious correlation. - Hybrid approach: Combining climatological covariance and ensemble-based covariance - Probabilistic model error: Probabilistic parameterization to represent model error - Efficient ensemble generation: Ensemble generation techniques to improve computational efficiency - Non-Gaussian transform: Processing of non-Gaussian variables using Gaussian transforms. 3. Hybrid Assimilation: We will implement a hybrid assimilation method that combines variational and ensemble methods. Specifically, we will perform hybrid assimilation using the following process: a. Hybrid representation of background error covariance: - Combine static climatological covariance with flow-dependent covariance estimated from the ensemble. - Linear combination: Weighted linear combination of climatological covariance and ensemble covariance. - Adaptive weights: Adaptive determination of optimal weights based on data. - Multiscale joining: Applying different weights at different scales - Variable-dependent weighting: Applying different weights to different variables - Localization: Localization of ensemble covariance - Expansion: Expansion of ensemble covariance b. Optimization using variational calculus: - Generate the analysis field using a variational method with hybrid background error covariance. - Cost function: Definition of the cost function using hybrid covariance - Preprocessing: Preprocessing for efficient optimization - Incremental approach: Incremental variational assimilation with outer and inner loops - Multiscale optimization: Stepwise optimization at different scales - Parallel computing: Improving computational efficiency through large-scale parallel computing. - Early termination: Early termination criteria that consider the balance between computational efficiency and accuracy. c. Ensemble update: - Update ensemble members using the results of variational calculus. - Deterministic ensemble transformations: Ensemble transformations based on the analytic field of variational calculus - Stochastic perturbation: Stochastic perturbation based on analytical error covariance - Hybrid Gain: A combination of gains from variational calculus and ensemble calculus. - Maintaining diversity: Techniques for maintaining diversity in an ensemble - Balance adjustment: Adjustments to ensure physical balance. - Non-negative constraint: Applying constraints to non-negative variables such as precipitation. This hybrid assimilation method combines the physical consistency of variational methods with the flow-dependent error representation of ensemble methods, enabling highly accurate assimilation by leveraging the advantages of both. For example, in predicting phenomena with strong nonlinearity, such as typhoons, it can generate a physically consistent analysis field while considering the flow-dependent error structure. The main advantage of hybrid assimilation is that it combines the advantages of variational and ensemble methods. Specifically, it combines the physical consistency and assimilation ability of variational methods with the flow-dependent error representation and uncertainty quantification ability of ensemble methods. This enables more accurate assimilation than any single method. The following technical challenges exist in implementing hybrid assimilation: - Computational cost: Computational cost of both variational method and ensemble method - Complexity: Implementation complexity due to the integration of two methods. - Parameter tuning: The difficulty of tuning parameters such as optimal weights. - Ensuring consistency: Ensuring consistency between the two methods. - Scalability: Scalability in large-scale problems To address these challenges, we will adopt the following technologies: - Efficient implementation: Optimization to improve computational efficiency - Modular Design: Modular Integration of Variational and Ensemble Methods - Adaptive parameter tuning: Adaptive tuning of parameters based on data. - Unified framework: A framework that handles two methods in a unified manner. - Parallel computing: Improving scalability through large-scale parallel computing. 4. Multiscale assimilation: We will implement multiscale assimilation, which integrates information from different spatial scales. Specifically, we will perform multiscale assimilation using the following process: a. Scale decomposition: - Decomposing observations and model predictions into different spatial scales - Wavelet decomposition: Multiscale decomposition using wavelet transform - Spectral decomposition: Spectral decomposition by Fourier transform - Empirical mode decomposition: Scale decomposition by empirical mode decomposition - Definition of scale: A definition of a physically meaningful scale. - Adaptive decomposition: Adaptive scaling decomposition according to data characteristics - Consideration of nonlinear scale interactions: Consideration of nonlinear interactions between different scales. b. Scale-dependent assimilation: - Apply the most suitable assimilation method for each scale. - Large-scale assimilation: Assimilation methods for large-scale patterns (e.g., variational methods) - Medium-scale assimilation: Assimilation methods for medium-scale patterns (e.g., hybrid assimilation) - Small-scale assimilation: Assimilation techniques for small-scale patterns (e.g., ensemble method) - Scale-dependent error model: A model suited to the error characteristics of each scale. - Scale-dependent localization: Adjustment of the localization radius according to the scale. - Scale-dependent observation selection: Selection of the optimal observation for each scale. c. Scaling integration: - Integrate assimilation results from different scales to generate the final analysis field. - Additive integration: Addition of analysis increments from different scales. - Weighted integration: Weighted integration based on reliability at each scale. - Hierarchical integration: Hierarchical integration from large to small - Ensuring consistency: Ensuring physical consistency between scales. - Edge processing: Handling discontinuities at scale boundaries - Consideration of nonlinear effects: Consideration of nonlinear effects in the integration process. This multiscale assimilation can perform appropriate assimilation for phenomena at different spatial scales, and can improve prediction accuracy, especially in areas with complex topography or uneven land cover. For example, it enables assimilation that simultaneously considers large-scale weather patterns (e.g., extratropical cyclones) and local convective systems. The main advantage of multiscale assimilation is that it allows the application of optimal assimilation methods to phenomena at different spatial scales. Atmospheric phenomena exist at various spatial scales, and each scale has different physical processes and error characteristics. Multiscale assimilation enables optimal assimilation tailored to these characteristics, improving overall prediction accuracy. The implementation of multiscale assimilation presents the following technical challenges: - Optimization of scale decomposition: Defining a physically meaningful scale decomposition - Inter-scale interaction: Processing interactions between different scales - Computational cost: Computational cost of assimilation across multiple scales - Ensuring consistency: Ensuring consistency between assimilation results at different scales. - Complexity Management: A Multiscale Approach to Complexity Management To address these challenges, we will adopt the following technologies: - Physically-based scale definition: A scale definition based on the physical properties of atmospheric phenomena. - Binding assimilation: Binding assimilation considering interscale interactions - Parallel computing: Improving computational efficiency through parallel processing at different scales. - Physical constraints: Constraints to ensure physical consistency - Modular design: Modular design for managing complexity 5. Non-Gaussian error support: We will implement data assimilation that supports non-Gaussian error distributions. Specifically, we will handle non-Gaussian errors using the following method: a. Particle filter: - Non-parametric assimilation methods based on Monte Carlo sampling - Particle generation: Generation of the initial particle set - Weight calculation: Particle weight calculation based on observation likelihood - Resampling: Weight-based particle resampling - Preventing particle degeneration: Technologies to prevent particle degeneration - Adaptive particle count: Adaptive adjustment of the particle count according to the required accuracy. - Localization: Improving the efficiency of particle filters through localization. b. Rank conversion: - A method that transforms a non-Gaussian distribution into a Gaussian distribution, performs assimilation, and then performs the inverse transformation. - Normalized score transformation: Gaussian transformation using empirical distribution functions - Box-Cox transform: Parametric nonlinear transformation - Logarithmic transformation: Logarithmic transformation for non-negative variables - Optimization of transformation parameters: Estimation of optimal transformation parameters - Inverse transformation consistency: Ensuring consistency during the inverse transformation. - Multivariate transformation: Simultaneous transformation of multiple variables c. Gaussian mixture model: - A method for approximating non-Gaussian distributions as a mixture of multiple Gaussian distributions. - Clustering: Identifying mixed components based on data clustering. - EM (Expected Value Maximization) algorithm: Parameter estimation for Gaussian mixture models - Optimization of the number of components: Determining the optimal number of mixed components. - Assimilation of each component: Individual assimilation for each Gaussian component. - Integration of results: Integration of assimilation results for each component. - Adaptive mixed models: Adaptive adjustment of mixed models according to data. These methods can improve the assimilation accuracy, particularly for variables with non-Gaussian distributions, such as precipitation. For example, assimilation that takes into account the log-normal distribution characteristics of precipitation can improve the accuracy of extreme precipitation predictions. The main advantage of non-Gaussian error handling is its ability to appropriately account for the non-Gaussian nature of actual error distributions. Many meteorological variables, particularly precipitation and wind speed, exhibit non-Gaussian distributions, and conventional assimilation methods based on Gaussian assumptions do not yield optimal results. Methods that handle non-Gaussian errors can improve the assimilation accuracy of these variables. The following technical challenges exist in implementing non-Gaussian error handling: - Computational cost: Computational cost of non-Gaussian methods (especially particle filters) - The Curse of Dimensions: The Difficulty of Sampling in Higher-Dimensional Spaces - Optimizing transformations: Identifying the optimal transformation function. - Multivariate non-Gaussianity: Handling simultaneous non-Gaussianity of multiple variables - Model nonlinearity: Interaction between model nonlinearity and non-Gaussianity To address these challenges, we will adopt the following technologies: - Efficient sampling: Efficient sampling techniques such as importance sampling - Localization: A localization approach to mitigate the curse of dimensions - Data-driven transformation: Estimation of the optimal transformation function based on data characteristics. - Copra Theory: Copra theory for modeling multivariate dependent structures - Hybrid approach: A hybrid approach that combines the advantages of different methods.

[0059] 6. Parameter estimation: We will implement data assimilation that simultaneously estimates model parameters. Specifically, parameter estimation will be performed using the following process: a. State-parameter coupling vector: - Define a vector that combines the model state variables and parameters. - Extended state vector: An extended vector containing state variables and parameters. - Scaling: Appropriate scaling of state variables and parameters at different scales. - Correlation structure: Modeling the correlation structure between state variables and parameters. - Evolutionary model: A time evolution model of parameters (usually a duration model) - Constraints: Incorporation of physical constraints on parameters - Uncertainty representation: Representation of parameter uncertainty b. Extended assimilation: - Apply data assimilation to the bonded vector, updating the state and parameters simultaneously. - Extended Kalman filter: Kalman filter for extended state vectors - Ensemble method: Ensemble Kalman filter for extended state vectors - Variational method: Variational assimilation of extended state vectors - Hybrid methods: Hybrid assimilation for extended state vectors - Localization: Localization to suppress spurious correlations between state variables and parameters. - Expansion: Expansion to improve the stability of parameter estimation c. Parameter constraints: - Limit the range of parameters based on physical constraints and prior information. - Boundary constraints: Setting upper and lower limits for parameters - Smoothness constraints: Constraints on the spatial and temporal smoothness of the parameters. - Physical relation constraints: Constraints based on the physical relationships between parameters - Regularization: Regularization to prevent over-conformity - Bayesian constraint: A constraint based on the prior distribution of the parameters. - Adaptive constraints: Adaptive adjustment of constraints based on data. This parameter estimation can reduce systematic bias in the model and improve prediction accuracy. For example, adjusting the parameters of the precipitation process based on observations can reduce systematic bias in precipitation forecasts. The main advantage of parameter estimation is that it can reduce systematic bias in the model and improve prediction accuracy. Many weather models have uncertainties associated with the parameterization of physical processes, which is a major cause of prediction bias. Parameter estimation allows these parameters to be optimized based on observational data, thereby improving the performance of the model. The following technical challenges exist in implementing parameter estimation: - Identifiability: Identifiability of parameters from limited observations - Nonlinearity: Nonlinear relationships between parameters and state variables - Multicollinearity: Multicollinearity due to correlations between parameters - Time variation: Handling of time-varying parameters - Overfitting: The risk of overfitting to current observations. To address these challenges, we will adopt the following technologies: - Sensitivity analysis: Sensitivity analysis to evaluate the identifiability of parameters - Regularization: Regularization techniques to prevent over-conformity - Stepwise estimation: Estimating stepwise from important parameters. - Time-varying parameter model: Modeling of parameters that vary over time. - Cross-validation: Evaluation of the generalization ability of estimated parameters. 7. Optimization of the observation operator: Optimize the observation operator that performs the transformation from model space to observation space. Specifically, perform the following optimizations: a. Physical-based operators: - Operators that simulate observed values ​​based on physical laws - Radiative transfer model: A radiative transfer model that simulates brightness temperature from satellite observations. - Radar reflection model: A model that simulates the radar reflection factor. - Visualization model: A model that simulates visible and infrared images. - Optimization of physical parameters: Optimization of the physical parameters of operators - Improved computational efficiency: Optimization to improve the computational efficiency of operators. - Improved precision: Improvements to enhance the precision of operators. b. Machine learning-based operators: - A machine learning model that has learned the relationship between observed data and model output. - Neural Networks: Observation Operators Based on Deep Learning - Random Forest: Observation Operator Based on Decision Tree Ensemble - Support Vector Machines: Observation Operators Based on Kernel Methods - Training data generation: Generating the data necessary for training the operators. - Feature selection: Selecting the optimal input features - Prevention of overconformance: Prevention of overconformance through regularization, etc. c. Hybrid operators: - An operator that combines physical and machine learning-based approaches - Physical model correction: Correcting the output of the physical model using machine learning. - Physically Constrained Learning: Machine learning incorporating physical constraints - Multi-model integration: Integration of multiple operator models - Adaptive selection: Selecting the optimal operator based on the conditions. - Considering uncertainty: Integration that takes into account the uncertainty of operators. - Improved Interpretability: Improved interpretability through a hybrid approach These optimized observation operators can reduce systematic differences between model outputs and observed values, thereby improving the accuracy of assimilation. For example, when directly assimilating the brightness temperature of a microwave radiometer, the effectiveness of assimilation can be maximized by improving the accuracy of the radiative transfer model. The main advantage of optimizing observation operators is that they improve the accuracy of the transformation between the model space and the observation space, thereby maximizing the effect of assimilation. In particular, in the assimilation of indirect observations such as satellite observations and radar observations, the accuracy of observation operators is a crucial factor determining the success or failure of the assimilation. Optimized observation operators allow for the effective assimilation of these observational data, improving prediction accuracy. The following technical challenges exist in implementing the optimization of the observation operator: - Computational cost: Especially the computational cost of physical-based operators - Nonlinearity: Strong nonlinearity of the observation operator - Training data: Lack of training data for machine learning-based operators - Generalization ability: The ability of an operator to generalize under various conditions. - Uncertainty assessment: Assessment of the uncertainty of the operator itself. To address these challenges, we will adopt the following technologies: - Approximation calculation: Approximation calculation techniques to improve computational efficiency - Linearization: Proper linearization of operators in the assimilation process - Simulation data: Generation of training data using a physical model - Transfer learning: Improving generalization ability with limited data. - Ensemble methods: Ensemble methods that represent the uncertainty of operators 8. Adaptive observation selection: The observations used for assimilation are adaptively selected. Specifically, observations are selected based on the following criteria: a. Quality of observation: - Quality assessment based on noise level, bias, and typicality error. - Quality control: Automated quality control of observation data - Anomaly detection: Detection of anomalies using statistical methods - Bias assessment: Assessment and correction of observational bias - Assessment of representativeness error: Assessment of the representativeness error of the observations. - Sensor status evaluation: Evaluation of the operating status of the sensor. - Environmental condition assessment: Evaluation of the observation environment (e.g., cloud cover, presence or absence of precipitation). b. Amount of information: - Evaluation of additional information compared to existing observations - Information quantity indicators: Information-theoretic indicators such as mutual information and entropy reduction. - Redundancy assessment: Evaluation of redundancy with existing observations. - Independence assessment: Assessment of independence between observations - Complementarity assessment: Evaluation of information that complements existing observations. - Novelty assessment: Assessment of the ability to provide new information. - Information density: An evaluation of the amount of information per unit cost. c. Sensitivity: - Assessment of the impact on predictions - Sensitivity analysis: Analysis of the influence of observations on predictions. - Singular vector: Sensitivity to the main mode of prediction error growth - Ensemble sensitivity: Sensitivity of ensemble predictions to variance - Objective function sensitivity: Sensitivity to a specific objective function (e.g., precipitation prediction error). - Temporal sensitivity: Sensitivity to different predicted lead times - Spatial sensitivity: Sensitivity to predictions for different regions d. Computational cost: - Evaluation of the computational resources required for assimilation - Observation operator cost: The computational cost of the observation operator. - Data volume: Size of observational data and processing cost - Parallel efficiency: Efficiency in parallel processing - Memory requirements: Evaluation of memory usage - Scheduling: Impact of computing resource scheduling - Cost-benefit ratio: The ratio of the predicted improvement effect to the computational cost. This adaptive observation selection allows for the maximum assimilation effect within limited computational resources. For example, it enables strategies such as assimilating many observations around rapidly developing weather systems and fewer observations in calm regions. The main advantage of adaptive observation selection is that it allows for the most efficient use of limited computational resources. Assimilation of all available observations is computationally expensive and does not necessarily lead to improved prediction accuracy. Adaptive observation selection allows for the selection of the most valuable observations and efficient use of computational resources. It can also address observation quality issues (e.g., bias, outliers, etc.). The following technical challenges exist in implementing adaptive observation selection: - Real-time evaluation: The difficulty of evaluating the value of observations in real time. - Interdependence: Consideration of interdependence between observations - Dynamic change: Temporal and spatial changes in the value of an observation. - Computational overhead: The computational cost of the observation selection itself. - Optimization problems: Solutions to large-scale combinatorial optimization problems To address these challenges, we will adopt the following technologies: - Preliminary assessment: Preliminary assessment of observed value based on similar past cases. - Graphical models: Modeling of dependencies between observations - Online learning: Online learning that responds to the dynamic changes in observed values - Efficient algorithms: Computationally efficient observation selection algorithms - Greedy algorithm: A greedy algorithm that considers the balance between computational efficiency and approximation accuracy. These features give the real-time data assimilation engine the following characteristics: 1. Real-time capabilities: The system minimizes the time from receiving observational data to completing assimilation, ensuring that the latest observational information is always reflected in the forecast. Specifically, it reduces the time from receiving observational data to completing assimilation to within a few minutes, minimizing the time gap between forecast and observation. Conventional approaches involved delays of several hours to half a day between transmitting observational data to the ground, performing data assimilation, and distributing updated forecasts. This system significantly reduces this delay by performing real-time data assimilation on the satellite. This improves the accuracy of forecasts, particularly for rapidly developing weather phenomena (e.g., rapid development of tropical cyclones and sudden convective precipitation).

[0060] 2. High precision: By using state-of-the-art data assimilation techniques, observations and model predictions are optimally integrated. This enables more accurate estimation of atmospheric conditions compared to model predictions or observations alone. For example, the spatial agreement of 24-hour precipitation forecasts can be improved by 15 to 25 percent. Conventional approaches often used relatively simple data assimilation techniques due to limitations in computing resources. This system leverages the capabilities of satellite-borne high-performance computing modules to implement advanced data assimilation techniques such as variational methods, ensemble Kalman filters, and hybrid assimilation. This allows for the effective integration of various types of observational data, maximizing prediction accuracy. 3. Adaptability: This system provides a flexible data assimilation system that adapts to various observation types, meteorological conditions, and regional characteristics. This enables consistent, high-quality assimilation results under diverse circumstances. Conventional approaches typically use fixed data assimilation settings, making it difficult to adapt to various conditions. This system achieves optimal data assimilation adapted to various conditions through technologies such as multiscale assimilation, non-Gaussian error handling, and adaptive observation selection. For example, it can apply assimilation settings optimized for convective precipitation in tropical regions, settings optimized for extratropical cyclones in mid-latitude regions, and settings optimized for sea ice and snow cover in polar regions. 4. Computational efficiency: This system efficiently performs data assimilation within limited computing resources. Specifically, it maximizes computational efficiency by using techniques such as parallel processing, approximation calculations, and adaptive observation selection. This enables real-time data assimilation within the limited computing resources of a satellite. Conventional approaches generally rely on data assimilation methods that assume large-scale ground computer systems, making them difficult to implement in environments with severe computing resource constraints. This system, through implementation and optimization that prioritizes computational efficiency, can perform advanced data assimilation even within the limited computing resources of a satellite. For example, it utilizes techniques such as localization, low-rank approximation, and parallel computing to achieve high-precision assimilation while reducing computational load. 5. Representation of uncertainty: This system explicitly expresses the uncertainty of the assimilation results. In particular, ensemble-based methods naturally express the uncertainty of the analysis field as the variance of the ensemble. This enables reliability assessment of predictions and analysis of uncertainty propagation. Conventional approaches often generate a deterministic single analysis field, limiting the evaluation of uncertainty. This system generates an ensemble of analysis fields using methods such as ensemble Kalman filtering and hybrid assimilation, explicitly expressing uncertainty. This enables probabilistic predictions such as "the probability of precipitation exceeding 50 mm / day at a specific location is 70%", supporting risk-based decision-making. 6. Integration of diverse observations: It possesses the capability to integrate various types of observational data. Specifically, it can integrate diverse observational data such as satellite observations (visible and infrared imagers, microwave radiometers, rainfall radar, etc.), ground observations (weather stations, radar, etc.), aircraft observations, and ocean observations. Conventional approaches generally use data assimilation systems specialized for specific types of observational data, making it difficult to integrate diverse observations. This system effectively integrates diverse observational data through observation operators that correspond to various observation types and data assimilation methods adapted to different observational characteristics. As a result, it can generate a comprehensive analysis field that leverages the strengths of each observation and compensates for its weaknesses. 7. Physical integrity: This system ensures the physical consistency of data assimilation results. Specifically, it generates a physically consistent analysis field by considering constraints based on physical laws such as conservation of mass, energy, and momentum. Conventional approaches sometimes resulted in an imbalance in physical balance due to the assimilation of observational data. This system achieves assimilation that takes physical constraints into account through methods such as variational methods and hybrid assimilation. Furthermore, physical balance adjustment (e.g., geostrophic balance adjustment) can be applied as a post-processing step for the analysis field. This improves the stability and accuracy of predictions after data assimilation. Thus, the real-time data assimilation engine is a crucial component for instantly integrating observational data and ESM forecasts to improve forecast accuracy. In particular, performing real-time data assimilation on satellite eliminates delays associated with ground processing, allowing the latest observational information to be immediately reflected in forecasts. This can significantly improve the accuracy of forecasts, especially for rapidly developing weather phenomena, and enhance the effectiveness of early warning systems. ### Operational Flow of Satellite Systems The satellite system of the present invention operates according to the following procedure: 1. Observe the Earth with a real-time observation sensor array: Multiple sensors mounted on the satellite (high-resolution visible and near-infrared imagers, microwave radiometers, precipitation radar, infrared sounders, scattermeters, etc.) continuously observe the Earth. These sensors capture various characteristics of the atmosphere and the Earth's surface, and collect information, in particular, related to precipitation. Observation data is acquired at an appropriate sampling rate for each sensor. For example, high-resolution visible light and near-infrared imagers acquire data approximately every 1 to 5 minutes, microwave radiometers every 10 to 30 minutes, rainfall radar every 5 to 15 minutes, infrared sounders every 15 to 45 minutes, and scattermeters every 30 to 60 minutes. These observation frequencies are optimized based on the characteristics of the sensor, the rate of change of the observed object over time, and the satellite's orbital characteristics. Observation data is temporarily stored in the satellite's onboard memory and sent to the onboard data processing unit. The amount of data varies depending on the sensor; for example, the high-resolution visible light / near-infrared imager generates a data rate of approximately 50 to 100 Mbps, the microwave radiometer approximately 1 to 5 Mbps, the rainfall radar approximately 5 to 10 Mbps, the infrared sounder approximately 10 to 20 Mbps, and the scattermeter approximately 1 to 2 Mbps. Satellite constellations can increase the frequency of observations at the same location. For example, a constellation of 12 satellites can enable observations globally approximately every 3 to 6 hours, and in specific areas approximately every 30 minutes to 1 hour. This allows for the capture of the time-dependent changes in rapidly developing weather systems. 2. Receive low-resolution ESM prediction data from the ground station: Satellites receive low-resolution ESM forecast data transmitted from ground-based meteorological agencies and research institutions. This data is typically provided with a coarse resolution of around 3° × 3.75° and represents global weather and climate patterns. ESM data is updated regularly (e.g., every 6 hours) to reflect the latest forecast information. The data is usually provided in standard meteorological data formats such as GRIB2 or NetCDF and includes information such as forecast variables (temperature, pressure, wind direction and speed, specific humidity, precipitation, etc.), forecast time (elapsed time from the analysis time), and grid information (latitude, longitude, vertical level). ESM data is received using S-band (2 to 4 GHz) or X-band (7 to 8 GHz) communication links, with data rates of approximately 1 to 10 Mbps. The received ESM data is stored in the satellite's onboard memory and sent to the onboard data processing unit. Predictive data may be received from multiple ESMs (e.g., POEM, GFDL-ESM4, SpeedyWeather.jl), and this data is used to generate multi-model ensemble predictions. Each ESM has different parameterizations and resolutions for physical processes, and each has its own unique strengths and weaknesses. Integrating predictions from multiple ESMs enables more robust predictions. 3. Preprocess observation data and ESM data using the onboard data processing unit: The onboard data processing unit performs preprocessing on observational data and ESM data. This preprocessing is crucial for preparing the input data for subsequent processing (such as high-resolution enhancement and data assimilation). Preprocessing of observational data includes the following steps: - Calibration and correction of sensor data: Calibration of radiometric measurements such as dark current correction, nonlinearity correction, relative response correction, and absolute calibration, as well as geometric correction based on satellite position and attitude information, and atmospheric correction to correct for the effects of absorption, scattering, and radiation by the atmosphere. - Cloud masking and quality control: Application of cloud detection algorithms to visible light and near-infrared imager data, quality control including anomaly detection, noise evaluation, and data reliability evaluation for each sensor's data, and assignment of quality flags and reliability scores based on data reliability evaluation. - Data compression and storage: Compression of processed data using a lossless compression algorithm, storage to non-volatile storage, and setting of retention periods according to the importance of the data. - Data indexing: Creating indexes based on attributes such as time, space, and sensor type, and optimizing data structures for efficient data retrieval. Preprocessing of ESM data includes the following steps: - Spatial interpolation and low-pass filtering: Interpolation of ESM data into high-resolution grids using techniques such as bilinear interpolation and cubic spline interpolation, and removal of artifacts by applying low-pass filters such as Gaussian filters. - Bias correction using quantile delta mapping (QDM): Bias correction for each quantile based on a comparison of the distribution of ESM data and observational data, and detailed bias correction using approximately 500 quantiles. - Data normalization and transformation: Applying a logarithmic transformation (x^~ = log(x + E) - log[E], E = 0.0001), normalizing to a range of approximately [-1, 1], and adjusting the data distribution to improve the efficiency of CM learning and inference. These preprocessing steps are performed efficiently using techniques such as parallel processing, pipelining, and hardware acceleration. The computational cost of preprocessing depends on the amount of data and the complexity of the algorithm, but typically uses 10 to 30% of the computing power of the satellite-borne high-performance computing module. The time required for preprocessing also depends on the amount of data, but is usually between a few seconds and a few minutes. 4. Determine the optimal spatial scale using a dynamic scaling adaptation system: The dynamic scale adaptation system analyzes the power spectral density (PSD) of observational and ESM data in real time to determine the optimal spatial scale. This spatial scale represents the smallest spatial scale for which the ESM data is reliable and is a natural choice for the spatial scale to retain when increasing resolution. Specifically, the spatial scale is determined through the following process: - Real-time PSD analysis: Conversion of observational data and ESM data to the wavenumber domain using two-dimensional Fourier transform, calculation of power for each wavenumber, and identification of intersections by comparing the PSDs of both. - Adaptive scale selection: Dynamic selection of spatial scale based on observation conditions such as cloud cover (selecting a larger spatial scale because the reliability of visible light and near-infrared observations decreases in cloud-covered areas), atmospheric stability (selecting a smaller spatial scale because small-scale phenomena become important in areas with an unstable atmosphere), precipitation intensity (selecting a smaller spatial scale because the spatial distribution of precipitation becomes complex in areas with heavy precipitation), and observation sensor quality (adjusting the spatial scale according to the quality of sensor data).

[0061] - Regional optimization: Optimization of scale based on regional characteristics such as topographic complexity (in areas with complex topography such as mountainous regions, topographic effects become important, so a smaller spatial scale is selected), land cover heterogeneity (in areas with uneven land cover such as urban areas or areas where farmland and forests are mixed, local effects become important, so a smaller spatial scale is selected), sea-land distribution (in areas where sea and land are mixed, such as coastal areas, sea-land contrast becomes important, so a smaller spatial scale is selected), and climate zone (the scale is adjusted according to climate zones such as tropical, mid-latitude, and polar regions). - Seasonal variation adjustments: Adjustments to scales based on seasonal variations such as monsoon (adjust the scale according to the season in areas where precipitation patterns change significantly between monsoon and non-monsoon seasons), snow and ice cover (adjust the scale according to the presence or absence of snow and ice cover in areas covered with snow and ice in winter), vegetation changes (adjust the scale according to seasonal changes in vegetation), and solar radiation conditions (adjust the scale according to seasonal changes in solar radiation conditions). The determined spatial scale is converted into the noise level (time step) in the conditional sampling of CM. Specifically, the relationship between the spatial scale (wavenumber k) and the noise level (time step t) is given by the equation σ^2(t) = N^2PSD(k), where σ^2(t) is the noise variance, N is the grid size, and PSD(k) is the power spectral density at wavenumber k. The spatial scale is updated approximately every 1 to 5 minutes, constantly adapting to the latest observation conditions. The determined spatial scale is sent to the CM implementation module to control the high-resolution process. 5. Generate high-resolution data with satellite-mounted CM implementation modules: The satellite-borne CM implementation module generates high-resolution data in a single step based on pre-processed ESM data and a determined spatial scale. This process leverages the stochastic sampling capabilities of the CM, enabling the efficient generation of physically consistent high-resolution data from low-resolution ESM data. Specifically, high-resolution data is generated using the following process: - Noise addition: Gaussian noise is added to the ESM data based on the noise level (time step t*) corresponding to the selected spatial scale (x^~ESM ≈ N(xESM; σ^2(t*)1)). - CM application: Apply CM to the noisy ESM data to generate high-resolution data in a single step (x^ = f(x^~ESM, t*; θ)) - Inverse Transform: Apply the inverse transform of normalization and logarithmic transformation to return to physical units (e.g., mm / day). CM employs a 2D U-Net architecture with four down / upsampling layers and an attention mechanism. This architecture consists of an encoder (stepwise downsampling of input data to extract features), a bottleneck layer (applying the attention mechanism to the lowest resolution feature map to extract global features), a decoder (stepwise upsampling of the output from the bottleneck layer to generate a high-resolution output), an output layer (generating the final high-resolution output), and a time embedding layer (providing noise level information to the model). The CM implementation is optimized to run efficiently within limited computing resources by applying optimization techniques such as quantization and lightweighting (reducing model size through quantization, pruning, and knowledge distillation of model parameters), sparse computation (omitting unnecessary computations through dynamic pruning, conditional computation, and early exit), pipeline processing (optimizing data flow through pipeline processing that executes each layer of the model in parallel, batch processing, and memory access optimization), and hardware acceleration (accelerating computationally intensive operations such as convolution, matrix multiplication, and activation functions using FPGAs). High-resolution data generation is extremely fast; high-resolution data of 0.75°×0.9375° can be generated from low-resolution 3°×3.75° ESM data in approximately 0.1 seconds. This is about 400 times faster than conventional SDE diffusion models. This high-speed generation enables real-time high-resolution processing, which is a significant advantage in time-constrained applications (e.g., heavy rain forecasting and disaster response). The generated high-resolution data is sent to the multimodal uncertainty quantification module and the real-time data assimilation engine. 6. Evaluate uncertainty using the multimodal uncertainty quantification module: The multimodal uncertainty quantification module generates multiple high-resolution realized values ​​from a single ESM output and assesses uncertainty. This module supports risk-based decision-making by providing a comprehensive assessment of uncertainty in high-resolution data to decision-makers. Specifically, uncertainty is assessed using the following process: - Ensemble generation: Run CM multiple times on the same ESM output using different random number seeds to generate 100 to 1000 ensemble members. - Calculation of statistical properties: Calculate the statistical properties of the ensemble (mean, standard deviation, quantiles, etc.) and estimate the probability distribution of precipitation for each grid cell. - Sensor uncertainty integration: Evaluation that takes into account sensor uncertainties such as measurement error, sampling error, representativeness error, and algorithm error. - Spatiotemporal correlation analysis: Analysis of correlation structures such as spatial correlation (correlation of uncertainty between different locations), temporal correlation (correlation of uncertainty between different time steps), and intervariate correlation (correlation of uncertainty between different meteorological variables). - Multi-model integration: Construction of a multi-model ensemble that integrates predictions from multiple ESMs, and assessment of structural uncertainty based on differences between models. - Confidence interval mapping: Visualization of information such as point estimates (e.g., ensemble mean), confidence intervals (e.g., 90% confidence interval), probability thresholds (e.g., probability of exceeding a specific threshold), and the spatial distribution of uncertainty. - Uncertainty propagation analysis: An analysis of how uncertainty propagates to downstream applications (e.g., flood prediction, crop yield prediction). The computational complexity of uncertainty assessment depends on the ensemble size and the level of detail of the assessment, but typically uses 20 to 40% of the computing power of the satellite-borne high-performance computing module. The time required for the assessment also depends on the ensemble size, but is typically between a few seconds and a few minutes. The results of the uncertainty assessment are sent to the real-time data assimilation engine and the direct distribution communication system. This enables the consideration of uncertainty in the data assimilation process and the provision of uncertainty information to end users. 7. Integrate observational data and high-resolution ESM predictions with a real-time data assimilation engine: The real-time data assimilation engine instantly integrates observational data with high-resolution ESM predictions. This engine is a core component for minimizing the time gap between observation and prediction and improving prediction accuracy. Specifically, data assimilation is performed using the following process: - Preparation of observational data: Quality assessment of pre-processed observational data, estimation of observational errors, application of observational operators, etc. - Background field preparation: Preparation of the background field for high-resolution ESM prediction, estimation of background error, etc. - Selection of assimilation method: Choose the optimal assimilation method from among variations (data assimilation using 4D variational methods), ensemble Kalman filters (ensemble-based data assimilation), hybrid assimilation (assimilation combining variational and ensemble methods), multiscale assimilation (assimilation that integrates information from different spatial scales), non-Gaussian error handling (assimilation that handles non-Gaussian error distributions), etc., depending on the situation. - Data assimilation: Data assimilation using the selected method and generation of the analysis field. - Evaluation of the analysis field: Quality evaluation of the generated analysis field, verification of physical consistency, etc. The computational cost of data assimilation depends on the assimilation method, the amount of observational data, and the complexity of the model, but typically uses about 30 to 60% of the computing power of the satellite-borne high-performance computing module. The time required for assimilation also depends on the assimilation method and the amount of observational data, but is usually several minutes to several tens of minutes. The results of data assimilation (analysis field) are sent directly to the distribution communication system. Furthermore, the analysis field is also used as the background field for the next data assimilation. This enables a continuous data assimilation cycle, allowing for predictions that always reflect the latest observational information. 8. Deliver data to end users via a direct distribution communication system: The direct delivery communication system delivers the final data product, including high-resolution data, uncertainty assessment results, and data assimilation results, directly to end users. This system makes it possible to directly provide high-resolution climate data to users worldwide, including in areas with limited communication infrastructure. Specifically, data is delivered through the following process: - Data product preparation: Preparation of the final data product including high-resolution data, uncertainty assessment results, data assimilation results, metadata addition, data format conversion, etc. - Direct broadcasting: Data broadcasting in a format receivable by standard receivers using multiple frequency bands, such as L-band broadcasting (1.5 to 1.6 GHz, data rate approximately 50 to 100 kbps, basic weather information), X-band broadcasting (7.8 to 8.4 GHz, data rate approximately 1 to 10 Mbps, detailed weather information), and Ka-band broadcasting (25.5 to 27 GHz, data rate approximately 10 to 100 Mbps, highest resolution weather information). - Adaptive data compression: Dynamic adjustment of data compression ratio according to communication bandwidth using techniques such as hierarchical data structures (structuring data at multiple resolution levels and selecting the appropriate resolution according to available bandwidth), adaptive quantization (adjusting the quantization level according to the importance of the data), context-adaptive entropy coding (selecting the optimal coding method according to the statistical characteristics of the data), and predictive coding (reducing data redundancy using temporal and spatial predictive models). - Priority-based delivery: Prioritization and preferential delivery based on criteria such as extreme event priority (data related to extreme events such as heavy rain, strong winds, high temperatures, and low temperatures are delivered with the highest priority), regional vulnerability (data for areas vulnerable to specific hazards, such as areas with high flood risk or landslide risk, is delivered with high priority), time urgency (data related to events with short predicted lead times is delivered with high priority), and user requests (data based on requests from specific users is delivered with priority according to the urgency of the request). - Regional customization: Data products delivered according to regional needs, including climate zone optimization (providing data products focused on the most relevant weather variables and phenomena according to climate zones such as tropical, mid-latitude, and polar regions), industry optimization (providing data products tailored to the major industries of a region, such as regions where agriculture is the main industry, regions where fisheries are the main industry, and regions where tourism is the main industry), disaster risk optimization (providing data products tailored to the major disaster risks of a region, such as regions with high flood risk, regions with high drought risk, and regions with high typhoon risk), and language and culture optimization (providing data representations and metadata appropriate to the local language and culture). - Two-way communication: Providing two-way communication functions such as request-based delivery (users request data on specific regions, times, and variables, and data is delivered accordingly), feedback mechanism (collecting user feedback and using it to improve the system), alert subscription (users subscribe to alerts based on specific conditions and receive notifications when those conditions are met), and data quality information (users request data quality information and receive information to evaluate data reliability).

[0062] The amount of data transmitted during data distribution depends on the type of data product, compression ratio, and target distribution area, but is typically delivered at data rates of approximately 50 to 100 kbps in the L band, 1 to 10 Mbps in the X band, and 10 to 100 Mbps in the Ka band. The time required for delivery also depends on the amount of data and the communication bandwidth, but is usually a few seconds to a few minutes. Data distribution is deli...

Claims

1. A satellite system for real-time high-resolution enhancement of low-resolution Earth System Model (ESM) simulation data, It comprises a satellite-mounted high-performance computing module, a real-time observation sensor array, an onboard data processing unit, a satellite-mounted consistent model (CM) implementation module, a dynamic scale adaptation system, a multimodal uncertainty quantification module, a direct distribution communication system, and a real-time data assimilation engine. The satellite system is characterized in that the satellite-mounted CM implementation module generates high-resolution data from low-resolution ESM data in a single step, and the real-time data assimilation engine immediately integrates the high-resolution data with observational data from the observation sensor array.

2. In the satellite system described in claim 1, The satellite system is characterized by the dynamic scale adaptation system automatically adjusting to the optimal spatial scale according to observation conditions, the multimodal uncertainty quantification module generating multiple high-resolution realized values ​​from a single ESM output to perform a comprehensive uncertainty assessment, and the direct distribution communication system directly distributing high-resolution climate data to end users in areas with limited communication infrastructure.

3. In the satellite system according to claim 1 or 2, A satellite system implemented as a constellation of multiple satellites, in which each satellite cooperates to generate high-resolution climate data for the entire Earth, and in which the real-time data assimilation engine integrates observational data and ESM predictions using at least one of variational data assimilation, ensemble Kalman filtering, and hybrid assimilation.