A short-term rainfall prediction method fusing satellite cloud image and radar echo

Abstract

The application discloses a kind of short-acting rain prediction methods of fusing satellite cloud picture and radar echo, it is related to short-acting rain weather prediction field, the specific steps of this method are as follows: synchronous acquisition preprocessed satellite cloud picture and radar echo data, respectively with symmetric time series deep learning architecture extraction feature, generate dimension alignment's rainfall probability discrete distribution;Unified distribution is obtained by double mode confrontation fusion;Probability confidence calibration is carried out, and the distribution of confidence interval is generated;According to magnitude interval mapping, the output different magnitude rainfall probability and confidence interval are calculated, and the grid point prediction is completed;The application adopts symmetric time series deep learning architecture to process multi-modal data, realizes macroscopic and microscopic collaborative representation, improves the completeness and consistency of rainfall probability generation;Combined with historical sample calibration deviation, the original logic and features are retained, and the standardized grid point prediction results are output;Optimize each link in whole process, so that the prediction result is more in line with the actual, and the business adaptability and application value are improved.

Description

Technical Field

[0001] This invention relates to the field of short-term precipitation weather forecasting, specifically a method for short-term precipitation forecasting that integrates satellite cloud images and radar echoes. Background Technology

[0002] Short-term precipitation forecasting is a core technological support for meteorological disaster early warning, urban flood control, and agricultural production scheduling, and is crucial for ensuring public safety and socio-economic development. Geostationary meteorological satellite cloud images can capture the spatiotemporal evolution of macroscopic cloud systems across a wide area, but their accuracy in characterizing microscopic precipitation particles is insufficient. Weather radar echoes can accurately invert the characteristics of microscopic precipitation particles, offering high spatial resolution, but they suffer from detection blind spots and limited regional coverage. Current short-term precipitation forecasting largely relies on single-modal data or employs simple weighting and stitching fusion methods, making it difficult to leverage the advantages of dual-modal data. Furthermore, it lacks probability calibration and confidence interval quantification mechanisms, resulting in forecast accuracy and reliability that cannot meet the needs of refined meteorological operations.

[0003] Existing short-term precipitation forecasting technologies have many shortcomings: single-modal forecasting models sever the intrinsic connection between macroscopic cloud systems and microscopic precipitation particles; satellite cloud image forecasts are prone to precipitation intensity deviations; and radar echo forecasts are limited by coverage and cannot achieve full-domain gridded forecasting. Dual-modal data fusion is mostly shallow stitching or fixed-weighting, without using a symmetric temporal deep learning architecture to extract common features, thus failing to preserve the consistency of spatiotemporal evolution between the two modalities. There is a lack of adversarial learning game-theoretic fusion mechanisms, making it easy for feature conflicts and information loss to occur in the dual-modal probability distribution, resulting in poor unified distribution generation. The predicted probabilities are not systematically calibrated, and there is no confidence interval quantification or error propagation calculation, leading to systematic biases in the forecasts. The probability output for rainfall magnitude classification is coarse and cannot provide reliable probabilistic basis for decision-making.

[0004] In summary, existing technologies cannot simultaneously meet the requirements of full coverage, micro-level precision, effective fusion, and probabilistic reliability. They struggle to address a series of technical issues, such as limitations of single-modal data, coarse dual-modal fusion, probabilistic distortion, and a lack of confidence quantification in forecast results. Therefore, there is an urgent need to develop a short-term precipitation forecasting method that fuses satellite cloud images and radar echoes. This method should leverage symmetric temporal deep learning, dual-modal adversarial fusion, probabilistic confidence calibration, and hierarchical error propagation techniques to combine the advantages of dual-modal data, overcome existing technical challenges, and improve the accuracy, comprehensiveness, and reliability of short-term precipitation forecasts. This would provide meteorological operations with an efficient and reliable refined forecasting solution. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method for short-term precipitation prediction that integrates satellite cloud images and radar echoes. This method simultaneously collects preprocessed cloud image and radar echo data, extracts macroscopic cloud system and microscopic precipitation particle features using a symmetric temporal deep learning architecture, and generates a dimension-aligned precipitation probability distribution. After dual-mode adversarial fusion to retain the core features of both modes, a unified distribution is obtained. Probability confidence calibration is performed by combining historical samples. According to a preset magnitude interval mapping, the probability of precipitation occurrence and confidence interval of different magnitudes are calculated and output, realizing gridded short-term precipitation probability forecasting.

[0006] To solve the above-mentioned technical problems, this invention provides the following technical solution: a method for short-term rainfall prediction that integrates satellite cloud images and radar echoes, the specific steps of which are as follows:

[0007] S100, Cloud Image Probability Generation: Collect time-series continuous geostationary meteorological satellite multi-channel cloud image data that has been preprocessed by spatiotemporal alignment and normalization, extract the spatiotemporal evolution characteristics of macro cloud systems through a time-series deep learning architecture, learn the probability mapping relationship between cloud system evolution and rainfall occurrence, and generate and output the discrete distribution of cloud image-driven rainfall probability with corresponding forecast lead time and full grid coverage.

[0008] S200, Echo Probability Generation: Executed synchronously with S100, it collects temporal and spatially aligned and normalized preprocessed temporal continuous weather radar combined reflectivity echo data. It adopts a temporal deep learning architecture that is completely symmetrical with S100 to extract the spatiotemporal evolution characteristics of micro precipitation particle echoes, learn the probabilistic mapping relationship between echo evolution and rainfall intensity, and generate and output the discrete distribution of echo-driven rainfall probability that is completely aligned with the dimensions of S100.

[0009] S300, Dual-Mode Adversarial Fusion: Receives the dual-modal precipitation probability discrete distribution output by S100 and S200, constructs a minimax game constraint through an adversarial learning mechanism, and generates and outputs a unified precipitation probability distribution that simultaneously retains the spatiotemporal evolution characteristics of macroscopic cloud systems and the spatiotemporal evolution characteristics of microscopic precipitation particle echoes.

[0010] S400, Probability Confidence Calibration: Receives the unified rainfall probability distribution output by S300, combines it with a pre-built historical calibration sample library, completes probability distribution calibration and prediction confidence quantification, and generates and outputs the calibrated rainfall probability distribution with a 95% confidence interval;

[0011] S500, graded forecast output: Receives the calibrated rainfall probability distribution with confidence intervals output from S400, completes the rainfall level mapping according to the preset rainfall level intervals, calculates and outputs the probability of occurrence of rainfall of different levels and the corresponding confidence intervals, and completes the gridded short-term rainfall probability forecast.

[0012] Furthermore, the cloud image probability generation adopts a ConvLSTM-based conditional variational autoencoder as a temporal deep learning architecture, including an encoder, a latent variable layer, and a decoder. The input is a temporally continuous sequence of multi-channel cloud images from geostationary meteorological satellites, with the temporal coverage of the sequence perfectly matching that of the radar echo input sequence. The encoder extracts the spatiotemporal features of the input sequence, constructs a constraint space for probability mapping through the latent variable layer, and finally generates a discrete distribution of rainfall probability corresponding to the forecast lead time through the decoder. The output distribution sets multiple continuous rainfall intensity bins, with each bin interval covering the complete rainfall intensity range, and the sum of the probabilities of all bins corresponding to each grid point is 1.

[0013] Furthermore, the echo probability generation employs a conditional variational autoencoder based on ConvLSTM, which is completely symmetrical to S100, as a temporal deep learning architecture. It includes an encoder, a latent variable layer, and a decoder that are completely identical to the structure of S100. The input is a temporally continuous weather radar combined reflectivity sequence, with the sequence time interval perfectly matching the time interval of the satellite cloud image input sequence. The encoder extracts the spatiotemporal features of the input sequence, and the latent variable layer constructs a constraint space that is of the same origin as S100. Finally, the decoder generates a discrete distribution of rainfall probability. The output distribution has the same spatial resolution, forecast lead time dimension, and rainfall intensity binning setting as the distribution output by S100.

[0014] Furthermore, the dual-modal adversarial fusion completes the fusion processing of dual-modal probability distributions through an adversarial learning mechanism, including a generator and a discriminator working in tandem. The generator's input is the channel dimension concatenation data of the dual-modal rainfall probability distributions output by S100 and S200, and its output is a unified rainfall probability distribution. The discriminator's input is the probability distribution to be discriminated, and its output is the grid-level distribution source discrimination result. The generator and discriminator update their parameters through alternating training and complete the fusion processing of dual-modal probability distributions through minimax game constraints. The output unified rainfall probability distribution is completely consistent with the dimensions of the input dual-modal distributions.

[0015] Furthermore, in the dual-mode adversarial fusion, the discriminator's training input includes three types of samples: the discrete distribution of rainfall probability driven by cloud image output from S100, the discrete distribution of rainfall probability driven by echo output from S200, and the unified rainfall probability distribution output by the generator in real time. The training objective of the discriminator is to distinguish the source type of the input samples, and the training objective of the generator is to make the output unified rainfall probability distribution unable to be distinguished by the discriminator. During the generator training process, all weight parameters of the discriminator are frozen, and during the discriminator training process, all weight parameters of the generator are frozen. The two are updated through a fixed alternating training frequency.

[0016] Furthermore, the pre-constructed historical calibration sample library consists of a unified rainfall probability distribution output by S300 for historical time periods and paired samples of actual rainfall observation data for the corresponding time periods; it includes full-scene samples of different seasons and different rainfall types; it sets up a daily incremental update mechanism, adding paired samples of the unified rainfall probability distribution output by S300 and the corresponding actual rainfall observation data for the previous day; and it sets up a fixed-period incremental retraining mechanism to complete feature statistics and data reconstruction of the sample library based on the newly added samples.

[0017] Furthermore, by combining a pre-built historical calibration sample library, probability distribution calibration and prediction reliability quantification are completed:

[0018] Using the uniform rainfall probability value output by S300 in the historical calibration sample library as input and the corresponding real-time rainfall binary label as output, the order-preserving regression calibration mapping function is trained to complete the grid-by-grid and time-by-time calibration of the uniform rainfall probability distribution output by S300 in real time.

[0019] Based on the trained ordinal-preserving regression mapping function f, the calibration residual for each sample in the historical calibration sample library is calculated. The standard deviation of the residuals within each probability interval is calculated to construct the 95% confidence interval of the calibrated probability.

[0020] The training objective formula for the order-preserving regression calibration mapping function is:

[0021]

[0022] in, To calibrate the mapping function for order-preserving regression; The uniform rainfall probability value output by S300 for the i-th sample in the historical calibration sample library; Let be the binary label of the actual rainfall corresponding to the i-th sample, with a value of 1 when rainfall occurs and a value of 0 when rainfall does not occur; The total number of valid samples in the historical calibration sample library; For constraint identification; , For any two original predicted probability values;

[0023] By minimizing the fitting residual between the predicted probability and the actual rainfall, a systematic bias calibration of the fused rainfall probability is achieved. At the same time, the sorting logic of the original probability is preserved with a monotonically non-decreasing constraint to avoid the calibration destroying the distribution characteristics of the bimodal adversarial fusion.

[0024] Furthermore, the preset multiple consecutive non-overlapping rainfall level intervals, the numerical range of each rainfall level interval completely matches the rainfall intensity bin intervals set in S100 and S200; each rainfall level interval corresponds to one or more consecutive rainfall intensity bins preset in S100 and S200, with no overlap or gap between intervals, completely covering all rainfall intensity bins; for each grid point and each forecast lead time, the probability values ​​of all rainfall intensity bins covered by the corresponding rainfall level interval are accumulated to obtain the rainfall probability corresponding to the rainfall level; the probability value corresponding to each rainfall level forms a one-to-one mapping relationship with the confidence interval of the corresponding covered bin.

[0025] Furthermore, the calculation and output of the probability of rainfall at different magnitudes and the corresponding confidence intervals are as follows:

[0026] Extract the upper and lower limits of the probability confidence intervals corresponding to all rainfall intensity bins covered by the target rainfall level range output by S400, calculate the total standard error of the multi-interval synthesis through the error propagation algorithm, and obtain the 95% confidence interval corresponding to the rainfall probability of the rainfall level based on the total standard error;

[0027] The error propagation algorithm formula is as follows:

[0028]

[0029] in, The composite standard error of the total probability of the target rainfall level; The total number of rainfall intensity bins covering the target rainfall level range corresponds exactly to the bin settings in S100 and S200. The probability standard error corresponding to the kth rainfall intensity bin covered by the target rainfall level interval; This represents the lower limit of the 95% confidence interval for the probability of rainfall at the target rainfall level. This represents the upper limit of the 95% confidence interval for the probability of rainfall at the target rainfall level. 1.96 represents the probability of rainfall occurring at the target rainfall level; 1.96 is the quantile corresponding to the 95% confidence level of the standard normal distribution.

[0030] Compared with existing technologies, this method for short-term rainfall prediction that integrates satellite cloud images and radar echoes has the following advantages:

[0031] I. This invention employs a symmetric temporal deep learning architecture to process satellite cloud images and radar echo data separately, simultaneously extracting the spatiotemporal evolution features of macroscopic cloud systems and microscopic precipitation particle echoes. It constructs a dimension-aligned discrete distribution of rainfall probability, effectively compensating for the inherent shortcomings of single-modal data and addressing the issues of insufficient accuracy in satellite cloud image representation and the detection blind spots and coverage limitations of radar echoes. By relying on an adversarial learning mechanism to establish a minimax game constraint, it completes the fusion of dual-modal probabilities, ensuring that the unified rainfall probability distribution fully preserves the core features of both modes, avoiding feature conflicts and information loss, and achieving a collaborative representation of macroscopic cloud system conditions and microscopic precipitation details. This improves the completeness and consistency of rainfall probability generation, ensuring forecast output with full-domain grid coverage, allowing short-term rainfall prediction to simultaneously consider wide-area coverage and high-resolution detection accuracy, adapting to the needs of meteorological monitoring and forecasting in multiple scenarios.

[0032] Second, this invention achieves systematic deviation calibration of the unified rainfall probability distribution by combining a historical calibration sample library. Relying on confidence interval quantification and error propagation rules, it retains the original probability ranking logic and fusion distribution characteristics, effectively improving the accuracy and reliability of the prediction results. It completes probability mapping and hierarchical calculation according to preset rainfall level intervals, outputting the occurrence probability of rainfall of different levels and corresponding confidence intervals, forming standardized and unified gridded probability forecast results. The entire process is optimized from probability generation, fusion, calibration to hierarchical output, making the forecast results more consistent with the actual rainfall situation, providing accurate and reliable quantitative support for meteorological early warning, urban flood control, production scheduling and other work, significantly improving the operational adaptability and practical application value of short-term rainfall forecasts.

[0033] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0035] Figure 1 This is a step-by-step framework diagram of a short-term rainfall prediction method that integrates satellite cloud images and radar echoes;

[0036] Figure 2 A flowchart illustrating a dual-mode adversarial fusion method for short-term rainfall prediction that integrates satellite cloud images and radar echoes;

[0037] Figure 3 This is a flowchart of the probability confidence calibration and graded forecasting process for a short-term rainfall prediction method that integrates satellite cloud images and radar echoes. Detailed Implementation

[0038] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0039] Example 1:

[0040] S100 Cloud Image Probabilistic Generation: This method involves collecting multi-channel, time-series continuous geostationary meteorological satellite cloud image data that has undergone spatiotemporal alignment and normalization preprocessing. Preprocessing ensures complete matching of the cloud image data along the time axis and spatial grid, standardizing and regularizing the numerical range to eliminate dimensional differences and numerical biases between different channels and time periods, providing a clean and standardized input foundation for subsequent deep learning feature extraction. A ConvLSTM-based conditional variational autoencoder is employed as the temporal deep learning architecture. This architecture efficiently extracts the spatiotemporal evolution features of the city's overall macroscopic cloud system through the encoder, fully capturing the entire cycle of cloud system generation, movement, development, and dissipation. The latent variable layer constructs a standardized constraint space for the probability mapping between cloud system evolution and rainfall occurrence, ensuring logical consistency between feature extraction and probability generation. The decoder, based on the extracted macroscopic spatiotemporal features and constraint space, generates and outputs a discrete distribution of cloud image-driven rainfall probability with corresponding forecast lead time and full grid coverage. The output distribution is set with multiple continuous rainfall intensity bins, and the bin intervals completely cover the entire rainfall intensity range. The sum of the probabilities of all bins corresponding to each grid point is 1. The deep extraction of macro cloud features allows the cloud map-driven probability to clearly reflect the overall rainfall trend of the city, making up for the shortcomings of insufficient coverage of ground observation and radar observation in the suburbs and edge areas of the city. It provides a stable and reliable macro probability input with a full-domain view for dual-modal data fusion, ensuring the spatial integrity of the subsequent fusion results.

[0041] S200 echo probability generation: Executed synchronously with S100 cloud image probability generation, it collects temporally and spatially aligned and normalized preprocessed time-series continuous weather radar combined reflectivity echo data. The synchronous execution mode ensures that the radar echo data and satellite cloud image data are completely matched in terms of time span and time interval. The preprocessing operation keeps the spatial resolution and numerical specifications of the radar echo data consistent with the cloud image data, avoiding fusion failure due to spatiotemporal mismatch of dual-mode data. A conditional variational autoencoder based on ConvLSTM, completely symmetrical to the S100, is adopted as the temporal deep learning architecture. The encoder, latent variable layer, and decoder structures are completely identical to those of the S100. The encoder accurately extracts the spatiotemporal evolution features of micro-precipitation particle echoes within the city, meticulously capturing details such as intensity changes, movement speed, and range contraction and expansion in the core precipitation area. The latent variable layer constructs a constraint space homologous to that of the S100, ensuring the logical homology of the bimodal probability generation. The decoder generates and outputs a discrete distribution of echo-driven precipitation probability that is completely aligned with the dimensions of the S100. This distribution has the same spatial resolution, forecast lead time dimension, and precipitation intensity binning settings as the S100 output distribution. The accurate extraction of micro-precipitation particle features allows the echo-driven probability to reflect the precipitation intensity status of local urban areas in real time. The symmetrical architecture and aligned dimensions design give the bimodal data a natural fusion adaptability, with macro trends and micro details forming an efficient complement, laying a solid data foundation for subsequent bimodal adversarial fusion and eliminating the problems of missing details or trend bias in single-modal prediction.

[0042] The S300 dual-modal adversarial fusion system receives dual-modal discrete rainfall probability distributions from the S100 and S200 systems. It constructs a minimax game constraint through adversarial learning and achieves deep fusion of the dual-modal probability distributions through a coordinated generator and discriminator. The generator takes channel-dimensional concatenation of the dual-modal rainfall probability distributions as input, integrating macroscopic cloud features and microscopic echo features to output a unified rainfall probability distribution. The discriminator takes three types of samples as input: cloud-driven discrete rainfall probability distribution, echo-driven discrete rainfall probability distribution, and the unified rainfall probability distribution output by the generator in real time. It outputs a grid-level distribution source discrimination result. The generator and discriminator update their parameters through alternating training, freezing their weights during training. The fusion effect is continuously optimized based on the minimax game constraint, ultimately ensuring that the output unified rainfall probability distribution is completely consistent with the dimensions of the input dual-modal distributions. The adversarial learning mechanism can automatically balance the feature weights of dual-modal data. The generator fully preserves the spatiotemporal evolution features of macroscopic cloud systems and microscopic precipitation particle echoes. The discriminator's accurate discrimination drives the continuous optimization of the fused distribution, enabling the unified rainfall probability distribution to have both the ability to judge macroscopic trends across the entire city and the ability to perceive microscopic precision in local urban areas. This completely solves the problems of insufficient detail in single satellite cloud image prediction and incomplete coverage in single radar echo prediction, providing unbiased and highly fused basic data for subsequent probability confidence calibration.

[0043] S400 Probability Confidence Calibration: Receives the unified rainfall probability distribution output from the S300 and combines it with a pre-built historical calibration sample library to complete probability distribution calibration and prediction confidence quantification. The historical calibration sample library consists of paired unified rainfall probability distributions for historical time periods with corresponding real-time rainfall observation data, covering full-scene samples across different seasons and rainfall types. A daily incremental update mechanism and a fixed-period incremental retraining mechanism are implemented to continuously supplement the latest rainfall samples and complete feature statistics and data reconstruction of the sample library, ensuring that the sample library always reflects the actual changing patterns of urban rainfall. An ordinal-preserving regression calibration mapping function is used to perform grid-by-grid, time-by-time calibration of the unified rainfall probability distribution. Based on the trained ordinal-preserving regression mapping function f, the calibration residual for each sample in the historical calibration sample library is calculated. The standard deviation of the residuals within each probability interval is calculated to construct the 95% confidence interval of the calibrated probability, as shown in the formula: ,in, To calibrate the mapping function for order-preserving regression; The uniform rainfall probability value output by S300 for the i-th sample in the historical calibration sample library; The actual rainfall is represented by a binary label for the i-th sample. The total number of valid samples in the historical calibration sample library; For constraint identification; , Given any two original predicted probability values, the calibration process focuses on minimizing the fitting residual between the predicted probability and the actual rainfall. Simultaneously, it preserves the original probability ranking logic through monotonically non-decreasing constraints, maintaining the distribution characteristics formed by the adversarial fusion of the two modes. Ultimately, it generates and outputs a calibrated rainfall probability distribution with a 95% confidence interval. The calibration process effectively corrects the systematic bias of the fused probability, accurately quantifies the prediction reliability of each grid point and each time point, and the probability distribution with confidence intervals significantly improves the stability and reliability of the forecast results. This allows various urban operational departments to judge the reliability of the forecast based on the confidence intervals, providing accurate and reliable calibration data for subsequent tiered forecast outputs.

[0044] S500 tiered forecast output: Receives the calibrated rainfall probability distribution with confidence intervals from the S400 output, and maps rainfall levels according to preset rainfall level intervals. The numerical ranges of multiple preset consecutive non-overlapping rainfall level intervals perfectly match the rainfall intensity binning intervals set in S100 and S200. Each rainfall level interval corresponds to one or more consecutive rainfall intensity bins, with no overlap or gaps between intervals, completely covering all rainfall intensity bins. For each grid point and each forecast lead time, the probability values ​​of all rainfall intensity bins covered by the corresponding rainfall level interval are accumulated to obtain the rainfall probability corresponding to each rainfall level. The total standard error after multi-interval synthesis is calculated using an error propagation algorithm, as shown in the formula: ,in, The composite standard error of the total probability of the target rainfall level; The total number of rainfall intensity bins covering the target rainfall level range; The probability standard error corresponding to the kth rainfall intensity bin covered by the target rainfall level interval; This represents the lower limit of the 95% confidence interval for the probability of rainfall at the target rainfall level. This represents the upper limit of the 95% confidence interval for the probability of rainfall at the target rainfall level. The probability of rainfall at the target rainfall level is defined as follows: 1.96 represents the quantile corresponding to the 95% confidence level of the standard normal distribution. Based on the total standard error, the 95% confidence intervals corresponding to the rainfall probability at each rainfall level are obtained. Finally, the probability of rainfall at different levels and the corresponding confidence intervals are output, completing the probability forecast of gridded short-term rainfall in the city. The hierarchical mapping logic fully meets the actual needs of urban meteorological operations. The error propagation algorithm ensures the accuracy of confidence interval calculation. The gridded forecast results can accurately cover every observation unit in the city, achieving comprehensive forecasting from the core urban area to the suburbs. This provides refined and highly reliable short-term rainfall forecast support for various tasks such as urban traffic control, flood control scheduling, outdoor work arrangements, and public meteorological services, fully meeting the accurate forecasting needs of modern urban operation and emergency response, and improving the overall efficiency and safety level of urban meteorological support.

[0045] This embodiment relies on a short-term precipitation prediction method that integrates satellite cloud imagery and radar echoes to complete the entire process of gridded short-term precipitation probability forecasting for the entire urban area, fully verifying the practicality and reliability of this method in urban meteorological support scenarios. The process involves simultaneously generating cloud image probability and echo probability, employing a symmetrical ConvLSTM-based conditional variational autoencoder architecture to extract the spatiotemporal features of macroscopic cloud systems and microscopic precipitation particles, generating a dimension-aligned dual-modal precipitation probability distribution to address the issues of incomplete coverage and missing details in single-modal data. The dual-modal adversarial fusion stage leverages adversarial learning and minimax game constraints to achieve deep integration of dual-modal features, outputting a unified precipitation probability distribution that balances overall coverage and local accuracy. The probability confidence calibration stage utilizes a dynamically updated historical calibration sample library, employing an ordinal regression calibration mapping function to complete bias correction and quantitative prediction of confidence, outputting calibration results with a 95% confidence interval. The tiered forecast output stage completes probability mapping according to preset magnitudes, calculating the magnitude probability and confidence interval through an error propagation algorithm, ultimately forming the urban gridded forecast results, such as... Figure 1 As shown, the standardized execution throughout the process effectively improves the accuracy, completeness, and reliability of short-term rainfall forecasts for cities, providing core data support for scenarios such as urban traffic management, flood control scheduling, and public services, and fully adapting to the actual needs of modern urban meteorological support and emergency response.

[0046] Example 2:

[0047] S100 Cloud Image Probabilistic Generation: This method collects multi-channel, time-series continuous geostationary meteorological satellite cloud image data that has undergone spatiotemporal alignment and normalization preprocessing. Preprocessing ensures the cloud image data is uniformly regularized in the spatiotemporal dimension and standardizes the numerical range, eliminating spatiotemporal biases and numerical differences between data from different time periods and channels. This provides a stable and standardized data foundation for extracting macroscopic cloud system features at the watershed scale. A ConvLSTM-based conditional variational autoencoder is used as the temporal deep learning architecture. The encoder extracts the spatiotemporal evolution features of macroscopic cloud systems across the entire watershed, fully capturing the generation, movement, development, and dissipation patterns of cloud systems within the watershed. It comprehensively covers the cloud system change trends in the upstream, midstream, and downstream sections of the watershed. A latent variable layer constructs a constraint space for the probability mapping between cloud system evolution and rainfall occurrence, ensuring the logical rigor of feature extraction and probability generation. The decoder generates and outputs a discrete distribution of cloud image-driven rainfall probability with corresponding forecast lead time and full grid coverage. The output distribution is set with multiple continuous rainfall intensity bins, and the bin intervals completely cover the entire rainfall intensity range. The sum of the probabilities of all bins corresponding to each grid point is 1. The deep extraction of macro cloud features allows the cloud map-driven probability to clearly reflect the overall rainfall development trend of the watershed. This solves the problems of insufficient coverage of observation equipment and difficulty in data acquisition in mountainous areas and remote river sections of the watershed. It provides macro probability input covering the entire watershed for dual-modal data fusion, ensuring the spatial coverage integrity and trend judgment accuracy of the subsequent fusion results.

[0048] S200 echo probability generation: Executed synchronously with S100 cloud image probability generation, it collects temporally and spatially aligned and normalized preprocessed time-series continuous weather radar combined reflectivity echo data. The synchronous execution mode ensures that the time span and time interval of the radar echo data and satellite cloud image data are completely consistent. The preprocessing operation keeps the spatial resolution and numerical specifications of the radar echo data consistent with the cloud image data, avoiding the impact of spatiotemporal misalignment of dual-modal data on the fusion effect. A conditional variational autoencoder based on ConvLSTM, which is completely symmetrical with S100, is adopted as the temporal deep learning architecture. The encoder, latent variable layer, and decoder structure of the architecture are completely consistent with S100. The encoder accurately extracts the spatiotemporal evolution characteristics of micro-precipitation particle echoes within the watershed, and meticulously captures detailed information such as precipitation intensity, movement direction, and duration in key river sections, reservoir areas, and flash flood-prone areas of the watershed. The latent variable layer constructs a constraint space that is homologous with S100 to ensure the logical consistency of bimodal probability generation. The decoder generates and outputs a discrete distribution of echo-driven rainfall probability that is completely aligned with the dimensions of S100. This distribution has the same spatial resolution, forecast lead time dimension, and rainfall intensity binning settings as the output distribution of S100. The precise extraction of micro-precipitation particle characteristics enables the echo-driven probability to reflect the actual precipitation situation in local areas of the watershed in real time. The symmetrical architecture and aligned dimensions design allow for seamless fusion of dual-modal data. The macro-precipitation trend and micro-precipitation details form an efficient complement, providing high-quality dual-modal data adapted to the watershed scenario for subsequent dual-modal adversarial fusion, avoiding the errors and limitations of single-modal prediction.

[0049] The S300 dual-modal adversarial fusion system receives the dual-modal discrete distributions of rainfall probability from the S100 and S200 systems. It constructs a minimax game constraint through an adversarial learning mechanism, and relies on a coordinated generator and discriminator to achieve deep fusion processing of the dual-modal probability distributions. The generator takes the channel-dimensional spliced ​​data of the dual-modal rainfall probability distribution as input, integrates the spatiotemporal characteristics of macroscopic cloud systems in the watershed with the microscopic precipitation particle echo characteristics, and outputs a unified rainfall probability distribution. The discriminator takes three types of samples as input: cloud-driven discrete rainfall probability distribution, echo-driven discrete rainfall probability distribution, and the unified rainfall probability distribution output in real time by the generator, and outputs a grid-level distribution source discrimination result. The generator and discriminator update their parameters through alternating training, freezing their weights during training. The fusion effect is continuously optimized based on the minimax game constraint, and the final output unified rainfall probability distribution is completely consistent with the dimensions of the input dual-modal distribution. The adversarial learning mechanism can reasonably balance the feature weights of dual-modal data. The generator fully preserves the macroscopic cloud system evolution patterns and microscopic precipitation particle change details of the watershed. The discriminator's accurate discrimination continuously optimizes the rationality of the fused distribution, allowing the unified rainfall probability distribution to not only grasp the overall trend of rainfall across the entire watershed but also accurately focus on the precipitation details of key local areas. This completely eliminates the problems of insufficient prediction accuracy from single satellite cloud images and incomplete coverage from single radar echo predictions. It provides highly fused, low-biased basic data adapted to the watershed flood control scenario for subsequent probability confidence calibration, such as... Figure 2 As shown.

[0050] S400 Probability Confidence Calibration: Receives the unified rainfall probability distribution output from S300, and combines it with a pre-built historical calibration sample library to complete probability distribution calibration and prediction confidence quantification. The historical calibration sample library consists of a pairing of the unified rainfall probability distribution for historical periods with corresponding real-time rainfall observation data, covering full-scene samples of different seasons and rainfall types in the watershed. It sets up a daily incremental update mechanism and a fixed-period incremental retraining mechanism to continuously supplement the latest rainfall samples in the watershed and complete the feature statistics and data reconstruction of the sample library, ensuring that the sample library always matches the seasonal changes, topographical influences, and type differences in rainfall in the watershed. A grid-by-grid, time-series-by-time calibration of a unified rainfall probability distribution is performed using an ordinal-preserving regression calibration mapping function. Based on the trained ordinal-preserving regression mapping function, the calibration residuals for each sample in the historical calibration sample library are calculated, and the standard deviation of the residuals within each probability interval is statistically analyzed to construct a 95% confidence interval for the calibrated probability. The calibration process focuses on minimizing the fitting residual between the predicted probability and the actual rainfall, while preserving the original probability ranking logic through monotonically non-decreasing constraints, thus maintaining the distribution characteristics formed by the adversarial fusion of the two modes. Finally, a calibrated rainfall probability distribution with 95% confidence intervals is generated and output. The calibration process effectively corrects the systematic bias in watershed-scale rainfall forecasts, accurately quantifies the prediction reliability of each grid point and each time series, and the probability distribution with confidence intervals significantly improves the stability and accuracy of watershed rainfall forecasts. This allows watershed flood control departments to determine the forecast risk level based on the confidence intervals, providing accurate, reliable, and flood control-adaptable calibration data for subsequent tiered forecast outputs.

[0051] The S500 tiered forecast output receives the calibrated rainfall probability distribution with confidence intervals from the S400. It maps rainfall levels according to preset rainfall level intervals. The numerical ranges of multiple consecutive, non-overlapping rainfall level intervals perfectly match the rainfall intensity binning intervals set in S100 and S200. Each rainfall level interval corresponds to one or more consecutive rainfall intensity bins, with no overlap or gaps between intervals, completely covering all rainfall intensity bins. For each grid point and each forecast lead time, the probability values ​​of all rainfall intensity bins covered by the corresponding rainfall level interval are accumulated to obtain the rainfall probability for each rainfall level. An error propagation algorithm is used to calculate the total standard error after multi-interval synthesis. Based on the total standard error, a 95% confidence interval corresponding to the rainfall probability for each rainfall level is obtained. Finally, the probability of rainfall at different levels and the corresponding confidence intervals are output, completing the probability forecast of gridded short-term rainfall in the watershed. Figure 3As shown, the hierarchical mapping logic perfectly aligns with the operational needs of flood control and drought relief in the basin. The error propagation algorithm ensures the accuracy of confidence interval calculations, and the gridded forecast results can accurately cover all hydrological observation units in the upstream, middle, and downstream areas of the basin, reservoir areas, and riverbanks. It provides comprehensive forecasts from overall trends to local details, offering refined and highly reliable short-term rainfall forecasts to support core tasks such as reservoir scheduling, river flood control, flash flood warnings, and optimal water resource allocation. This effectively prevents the risk of floods in the basin, improves the efficiency of water resource utilization and the accuracy of flood control and drought relief management, and provides solid meteorological support for the ecological security and production and livelihood safety of the basin.

[0052] This embodiment applies a short-term precipitation forecasting method that integrates satellite cloud imagery and radar echoes to a watershed flood control and early warning scenario. It fully realizes gridded, highly reliable short-term precipitation forecasts at the watershed scale, demonstrating the method's adaptability and effectiveness in watershed hydrological and meteorological operations. During implementation, the cloud image probability generation stage relies on satellite data to acquire macroscopic cloud system characteristics across the entire watershed, compensating for observational shortcomings in remote areas. The echo probability generation stage simultaneously extracts microscopic precipitation characteristics from radar echoes, ensuring complete alignment of the two modal data dimensions and efficient feature complementarity. Dual-modal adversarial fusion balances the weights of the two modalities through alternating training of the generator and discriminator, preserving both overall trends and local details while eliminating single-modal prediction errors. Probability confidence calibration relies on a full-scene historical calibration sample library, employing an ordinal regression calibration mapping function to calibrate the distribution and calculate a 95% confidence interval, significantly improving forecast stability and accuracy. The tiered forecast output stage strictly matches rainfall intensity bins, calculating the magnitude probability and confidence interval using an error propagation algorithm to form gridded forecast data for the watershed. It effectively addresses the issues of insufficient basin observation coverage and limited forecast accuracy, providing precise support for reservoir scheduling, river flood control, flash flood warning, and water resource allocation, significantly improving the basin's flood control and drought relief management level, and effectively safeguarding the basin's production, livelihood, and ecological security.

[0053] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for short-term rainfall prediction that integrates satellite cloud images and radar echoes, characterized in that, The specific steps of this method are as follows: S100, Cloud Image Probability Generation: Collect time-series continuous geostationary meteorological satellite multi-channel cloud image data that has been preprocessed by spatiotemporal alignment and normalization, extract the spatiotemporal evolution characteristics of macro cloud systems through a time-series deep learning architecture, learn the probability mapping relationship between cloud system evolution and rainfall occurrence, and generate and output the discrete distribution of cloud image-driven rainfall probability with corresponding forecast lead time and full grid coverage. S200, Echo Probability Generation: Executed synchronously with S100, it collects temporal and spatially aligned and normalized preprocessed temporal continuous weather radar combined reflectivity echo data. It adopts a temporal deep learning architecture that is completely symmetrical with S100 to extract the spatiotemporal evolution characteristics of micro precipitation particle echoes, learn the probabilistic mapping relationship between echo evolution and rainfall intensity, and generate and output the discrete distribution of echo-driven rainfall probability that is completely aligned with the dimensions of S100. S300, Dual-Mode Adversarial Fusion: Receives the dual-modal precipitation probability discrete distribution output by S100 and S200, constructs a minimax game constraint through an adversarial learning mechanism, and generates and outputs a unified precipitation probability distribution that simultaneously retains the spatiotemporal evolution characteristics of macroscopic cloud systems and the spatiotemporal evolution characteristics of microscopic precipitation particle echoes. S400, Probability Confidence Calibration: Receives the unified rainfall probability distribution output by S300, combines it with a pre-built historical calibration sample library, completes probability distribution calibration and prediction confidence quantification, and generates and outputs the calibrated rainfall probability distribution with a 95% confidence interval; S500, graded forecast output: Receives the calibrated rainfall probability distribution with confidence intervals output from S400, completes the rainfall level mapping according to the preset rainfall level intervals, calculates and outputs the probability of occurrence of rainfall of different levels and the corresponding confidence intervals, and completes the gridded short-term rainfall probability forecast.

2. The method for short-term rainfall prediction by fusing satellite cloud images and radar echoes according to claim 1, characterized in that, In step S100, the cloud image probability generation adopts a conditional variational autoencoder based on ConvLSTM as a temporal deep learning architecture, including an encoder, a latent variable layer, and a decoder. The input is a temporally continuous sequence of multi-channel cloud images from geostationary meteorological satellites, with the temporal coverage of the sequence completely matching that of the radar echo input sequence. The encoder extracts the spatiotemporal features of the input sequence, constructs a constraint space for probability mapping through the latent variable layer, and finally generates a discrete distribution of rainfall probability corresponding to the forecast lead time through the decoder. The output distribution sets multiple continuous rainfall intensity bins, with the bin interval covering the complete rainfall intensity range, and the sum of the probabilities of all bins corresponding to each grid point is 1.

3. The method for short-term rainfall prediction by fusing satellite cloud images and radar echoes according to claim 1, characterized in that, In step S200, the echo probability generation adopts a conditional variational autoencoder based on ConvLSTM, which is completely symmetrical to S100, as a temporal deep learning architecture. It includes an encoder, a latent variable layer, and a decoder that are completely identical to the structure of S100. The input is a temporally continuous weather radar combined reflectivity sequence, and the sequence time interval is completely matched with the time interval of the satellite cloud image input sequence. The spatiotemporal features of the input sequence are extracted by the encoder, and a constraint space of the same origin as S100 is constructed through the latent variable layer. Finally, the discrete distribution of rainfall probability is generated by the decoder. The output distribution has the same spatial resolution, forecast lead time dimension, and rainfall intensity binning setting as the distribution output by S100.

4. The method for short-term rainfall prediction by fusing satellite cloud images and radar echoes according to claim 1, characterized in that, In step S300, the dual-modal adversarial fusion is performed by an adversarial learning mechanism to fuse the dual-modal probability distributions. This includes a generator and a discriminator working together. The generator's input is the channel dimension concatenation data of the dual-modal rainfall probability distributions output from S100 and S200, and its output is a unified rainfall probability distribution. The discriminator's input is the probability distribution to be discriminated, and its output is a grid-level distribution source discrimination result. The generator and discriminator update their parameters through alternating training and complete the fusion of the dual-modal probability distributions through minimax game constraints. The output unified rainfall probability distribution is completely consistent with the dimensions of the input dual-modal distribution.

5. A method for short-term rainfall prediction by fusing satellite cloud images and radar echoes according to claim 1 or 4, characterized in that, In step S300, the dual-mode adversarial fusion involves three types of samples being input to the discriminator: a discrete distribution of rainfall probability driven by cloud imagery output from S100, a discrete distribution of rainfall probability driven by echo output from S200, and a unified rainfall probability distribution output by the generator in real time. The discriminator is trained to distinguish the source type of the input samples, while the generator is trained to ensure that the output unified rainfall probability distribution cannot be distinguished by the discriminator. During generator training, all weight parameters of the discriminator are frozen, and during discriminator training, all weight parameters of the generator are frozen. The two are updated through a fixed alternating training frequency.

6. The method for short-term rainfall prediction by fusing satellite cloud images and radar echoes according to claim 1, characterized in that, In step S400, the pre-constructed historical calibration sample library consists of a unified rainfall probability distribution output by historical time period S300 and paired samples of actual rainfall observation data for the corresponding time period; it includes full-scene samples of different seasons and different rainfall types; a daily incremental update mechanism is set up, adding paired samples of the unified rainfall probability distribution output by S300 from the previous day and the corresponding actual rainfall observation data every day; a fixed-period incremental retraining mechanism is set up, completing feature statistics and data reconstruction of the sample library based on the newly added samples.

7. The method for short-term rainfall prediction by fusing satellite cloud images and radar echoes according to claim 1, characterized in that, In step S400, the probability distribution calibration and prediction confidence quantification are completed by combining the pre-built historical calibration sample library: Using the uniform rainfall probability value output by S300 in the historical calibration sample library as input and the corresponding real-time rainfall binary label as output, the order-preserving regression calibration mapping function is trained to complete the grid-by-grid and time-by-time calibration of the uniform rainfall probability distribution output by S300 in real time. Based on the trained ordinal-preserving regression mapping function f, the calibration residual for each sample in the historical calibration sample library is calculated. The standard deviation of the residuals within each probability interval is calculated to construct the 95% confidence interval of the calibrated probability. The training objective formula for the order-preserving regression calibration mapping function is: in, To calibrate the mapping function for order-preserving regression; The uniform rainfall probability value output by S300 for the i-th sample in the historical calibration sample library; The actual rainfall is represented by a binary label for the i-th sample. The total number of valid samples in the historical calibration sample library; For constraint identification; , Let be any two original predicted probability values.

8. The method for short-term rainfall prediction by fusing satellite cloud images and radar echoes according to claim 1, characterized in that, In step S500, the preset multiple consecutive non-overlapping rainfall level intervals are provided, and the numerical range of each rainfall level interval is completely matched with the rainfall intensity sub-bin intervals set in S100 and S200. Each rainfall level interval corresponds to one or more consecutive rainfall intensity sub-bins preset in S100 and S200, and there is no overlap or gap between the intervals, completely covering all rainfall intensity sub-bins. For each grid point and each forecast lead time, the probability values ​​of all rainfall intensity bins covered by the corresponding rainfall level interval are accumulated to obtain the probability of rainfall occurrence corresponding to the rainfall level; the probability value corresponding to each rainfall level forms a one-to-one mapping relationship with the confidence interval of the corresponding coverage bin.

9. A method for short-term rainfall prediction by fusing satellite cloud images and radar echoes according to claim 1 or 8, characterized in that, In step S500, the probability of rainfall of different magnitudes and their corresponding confidence intervals are calculated and output: Extract the upper and lower limits of the probability confidence intervals corresponding to all rainfall intensity bins covered by the target rainfall level range output by S400, calculate the total standard error of the multi-interval synthesis through the error propagation algorithm, and obtain the 95% confidence interval corresponding to the rainfall probability of the rainfall level based on the total standard error; The error propagation algorithm formula is as follows: in, The composite standard error of the total probability of the target rainfall level; The total number of rainfall intensity bins covering the target rainfall level range; The probability standard error corresponding to the kth rainfall intensity bin covered by the target rainfall level interval; This represents the lower limit of the 95% confidence interval for the probability of rainfall at the target rainfall level. This represents the upper limit of the 95% confidence interval for the probability of rainfall at the target rainfall level. 1.96 represents the probability of rainfall occurring at the target rainfall level; 1.96 is the quantile corresponding to the 95% confidence level of the standard normal distribution.

Images