A method and device for on-orbit screening and compression of remote sensing data based on intelligent computing on satellite

Abstract

The application discloses a kind of on-orbit screening and compression method and device of remote sensing data based on intelligent computing on satellite, belong to remote sensing satellite data processing and data compression technical field.The method is by deploying lightweight AI model on satellite, first, the image quality of complete one scene remote sensing data is evaluated, and the data of which quality is not available is removed according to it;For the scene data of which quality is qualified, further content understanding and value evaluation are carried out, and the whole scene data of which content is not valuable is directly removed according to it;For the effective scene data of which content is valuable, then part block is carried out in scene, and different data blocks are implemented differentiated compression according to value distribution information;Finally, all compression blocks and scene level index table recording its position and compression parameters are packaged as complete data product and downloaded.The application realizes intelligent hierarchical processing of on-orbit data through configurable quality and value standards, and maximizes satellite-ground transmission efficiency under the premise of guaranteeing data availability and value density.

Description

Technical Field

[0001] This invention relates to the field of remote sensing satellite data processing and data compression technology, and in particular to a method and apparatus for on-orbit screening and compression of remote sensing data based on on-board intelligent computing. Background Technology

[0002] With the rapid development of remote sensing satellite technology, satellite observation resolution has advanced from meter-level to sub-meter-level, and the observation frequency has increased from once a day to hourly revisits, resulting in an exponential increase in the amount of raw remote sensing data. However, the satellite-to-ground data transmission link is limited by bandwidth (typically 50Mbps to 200Mbps) and transit time (only 10 to 20 minutes of transmission time per transit), resulting in a long-standing contradiction between a large amount of data generated and a small amount that can be transmitted. Traditional solutions transmit all satellite observation data (including invalid data such as cloud cover, no target, and low-quality data) to the ground processing center without requiring complex onboard processing, but this approach suffers from two major problems: significant bandwidth waste (invalid data can account for 30% to 70%) and heavy pressure on ground processing. To solve the problem of full data transmission, existing technologies have proposed different intelligent processing schemes, but all have limitations:

[0003] Compression optimization-based schemes, such as Chinese patent CN115131673A, improve compression efficiency by identifying regions of interest in orbit and optimizing bitrate allocation. However, this method lacks a pre-screening mechanism, failing to identify and remove invalid whole-scene data before compression, resulting in a large amount of invalid data still occupying on-board resources.

[0004] Target extraction-based schemes, such as Chinese patent CN113486883A, directly extract and only transmit slices of the target of interest through on-orbit intelligent processing. While this method boasts extremely high transmission efficiency, it results in the permanent loss of the original image's geographic integrity, failing to meet the application requirements for macroscopic analysis and change detection of complete geographic space.

[0005] Channel modulation-based schemes, such as Chinese patent CN111478741A, focus on adaptively adjusting data transmission parameters (such as code rate, encoding, modulation scheme, and transmit power) according to the satellite-to-ground link status. Their core lies in optimizing the transmission channel, which is a communication-level optimization. However, they do not fundamentally solve the problem of low data source value density. Their data processing methods remain crude. For example, in data selection logic, they only determine whether to transmit raw or compressed data based on whether the region of interest reaches a preset value, lacking a refined value assessment and hierarchical compression mechanism for the data content itself.

[0006] Other regional compression schemes, such as Chinese patent CN114839631A, use block compression, but their value assessment dimensions are relatively simple (e.g., mainly targeting raw SAR data, or relying on classification accuracy and prediction score to calculate importance).

[0007] In summary, existing technologies have failed to achieve a balance in four key dimensions: data transmission efficiency, on-board resource consumption, data geographic integrity, and application flexibility. They suffer from issues such as full-data downlink ensuring integrity but with extremely low efficiency, target extraction ensuring efficiency but sacrificing integrity, and compression optimization and channel modulation being limited to local improvements or communication-level optimizations, respectively, failing to provide an optimal solution at the system level. Summary of the Invention

[0008] Purpose of the invention: The purpose of this invention is to provide a method and apparatus for on-orbit screening and compression of remote sensing data based on on-board intelligent computing, which aims to solve the contradiction between the surge in remote sensing satellite data volume and the limited satellite-to-ground data transmission resources. It can realize intelligent hierarchical processing of on-orbit data and maximize satellite-to-ground transmission efficiency while ensuring data availability and value density.

[0009] Technical Solution: An on-orbit screening and compression method for remote sensing data based on on-board intelligent computing. After acquiring remote sensing data from the satellite platform, the following steps are performed:

[0010] D1, preprocessing the acquired whole-scene remote sensing data;

[0011] D2 inputs the pre-processed whole-scene remote sensing data into a lightweight AI model deployed on the on-board computing platform;

[0012] D3. The lightweight AI model is used to evaluate the image quality of the entire remote sensing data. If the quality parameters do not meet the preset usability threshold, the data is determined to be of unusable quality.

[0013] For usable whole-scene remote sensing data, the lightweight AI model is used to further understand the content and evaluate its value. If the content value does not meet the preset requirements, the content is deemed invalid.

[0014] Full-view remote sensing data that is of unusable quality or invalid content will be directly deleted from the on-board storage and will not be downloaded; for full-view remote sensing data with valid content, the subsequent compression and download process will be performed.

[0015] D4 divides the whole-scene remote sensing data with valid content into blocks, and a lightweight AI model performs a rapid value assessment on each data block, generating a quantitative value label for each data block; by calling the on-board adaptive compression engine, it dynamically selects compression strategies and parameters for each data block and implements differentiated compression.

[0016] D5 generates a scene-level index table by taking the pixel position of each data block in the original scene, the assigned value tag, and the corresponding compression parameters; then, all compressed data blocks are packaged together with this scene-level index table to form a complete download data product with self-describing capabilities.

[0017] D6 sends the packaged downlink data product to the ground station via the satellite-to-ground data transmission interface.

[0018] Furthermore, the lightweight AI model is a convolutional neural network model with fewer than 5 million parameters, which has undergone structured pruning and 8-bit quantization processing. Moreover, the power consumption for inference of standard-size data of a single scene on the onboard embedded AI processor is less than 10 watts, and the inference latency is less than 1 second.

[0019] Furthermore, the construction and optimization of the lightweight AI model are completed through ground training, and the specific steps are as follows:

[0020] D21, on the ground, uses a combination of large-scale public remote sensing datasets and a self-built proprietary dataset, and performs augmentation processing on the combined data to construct a training sample set covering various land cover types, imaging conditions and geographical environments; the self-built proprietary dataset is a self-built remote sensing image sample library containing buildings, road networks, water bodies, vegetation cover types, crops, clouds and shadows, ships, sea ice, disaster areas and specified mineral or rock types.

[0021] D22 employs a pre-training-fine-tuning approach: First, a standard convolutional neural network is pre-trained using a large-scale public remote sensing dataset to enable it to extract general visual features. Then, for a predefined on-orbit screening task, a self-built proprietary dataset is used to fine-tune the standard convolutional neural network, resulting in a high-performance model after fine-tuning.

[0022] D23, based on the importance score of neurons or channels, performs structured pruning on the fine-tuned high-performance model, removes redundant parts in the convolutional neural network that contribute little to the output, and generates a sparse model.

[0023] The 32-bit floating-point numbers and activation values ​​of the trained sparse model are mapped and converted into 8-bit integer representations to obtain the optimized sparse model.

[0024] D24 utilizes the inference engine provided by the onboard target processor to convert the optimized sparse model into a hardware-friendly format, further optimizing the computation graph and accelerating the inference process.

[0025] Furthermore, the image quality assessment parameters include at least the average cloud coverage ratio of the entire scene, the average signal-to-noise ratio of the entire scene, and image blur. The definition and adjustment of the availability threshold follow a preset task strategy rule library: during the ground mission planning stage, threshold templates for different scenarios are pre-set according to the application objectives; the onboard autonomous decision-making and update module dynamically adjusts the thresholds according to the rules fixed in the state machine based on the real-time resource status; all thresholds are encapsulated in the task strategy configuration file in the form of structured parameters, supporting remote injection and updates from the ground.

[0026] The content understanding and value assessment includes at least the identification and statistical analysis of one or more preset targets of interest, such as buildings, road networks, water bodies, vegetation cover types, crops, clouds and shadows, ships, sea ice, disaster areas, and specified mineral or rock types; the preset requirements for content value are defined and adjusted through the currently effective task strategy configuration.

[0027] Furthermore, the steps to achieve differentiated compression include:

[0028] D41 divides the valid whole-view remote sensing data into multiple data blocks;

[0029] D42, the lightweight AI model performs a rapid value assessment on each data block and generates a quantified value tag for each data block. The set of these value tags constitutes value distribution information.

[0030] D43, by calling the on-board adaptive compression engine, dynamically selects compression strategies and parameters for each data block based on the value distribution information, and implements differentiated compression.

[0031] Furthermore, the differentiated compression includes: for high-value data blocks determined to contain key targets or features, lossless compression or lossy compression with a low compression ratio of 1:2 to 1:5 is used; for low-value data blocks determined to have no key targets, lossy compression with a high compression ratio of 1:10 to 1:25 is used.

[0032] Furthermore, the mission strategy configuration file is received from the ground via the satellite-to-ground telemetry and control link. After verification, the on-orbit strategy configuration is updated and switched. When the remote sensing satellite observation mission mode changes, the on-orbit processing strategy is quickly reconstructed by updating the image quality assessment threshold, target of interest type and value judgment criteria, and differentiated compression parameters in the configuration file.

[0033] An on-orbit remote sensing data screening and compression device based on on-board intelligent computing is used to implement any of the above-mentioned on-orbit remote sensing data screening and compression methods based on on-board intelligent computing, and to complete the on-orbit processing of remote sensing data. It is based on an embedded AI computing platform and integrates the following modules:

[0034] The data preprocessing module is used to carry lightweight on-orbit radiometric calibration and system geometry correction algorithms, providing input data that conforms to radiometric and geometric specifications;

[0035] The data storage module is used to cache the entire scene data to be processed, the compressed data, and store the lightweight AI model, on-orbit screening strategy parameters, and compression strategy parameters.

[0036] An embedded AI computing module is used to carry and execute lightweight AI models that meet predetermined power consumption and performance indicators;

[0037] The data segmentation and compression module is used to perform data block segmentation and value-oriented adaptive compression based on the quantization compression ratio range;

[0038] The data encapsulation module is responsible for generating scene-level index tables containing value tags and packaging data products.

[0039] The data transmission scheduling module manages the scheduling and transmission of downlink data products, sending priority data streams to the ground station.

[0040] The autonomous decision-making and update module enables status monitoring, policy adjustment, and remote update functions based on the policy update instructions sent by the ground station and the status parameters of the satellite platform.

[0041] Compared with the prior art, the significant advantages of this invention are as follows:

[0042] 1. The method of this invention employs a lightweight AI model for dual screening of quality and value, eliminating invalid whole-scene data from the source. The evaluation dimensions are more comprehensive (such as cloud coverage, signal-to-noise ratio, ambiguity, and identification of multiple preset targets of interest), thereby achieving more accurate invalid data removal and more refined value grading. Combined with in-scene tiered compression based on quantization compression ratio (1:2-1:5; 1:10-1:25), the total amount of satellite-to-ground data transmission is reduced by 50%–80%, directly alleviating data backlog problems caused by bandwidth and transit time limitations, and improving the timeliness of satellite data services. Simultaneously, the use of a lightweight AI model with fewer than 5 million parameters, power consumption less than 10 watts, and latency less than 1 second, combined with pipelined processing and real-time removal mechanisms, significantly reduces the demand for onboard computing and storage resources, making it suitable for micro-nano satellite platforms.

[0043] 2. The method of this invention, while achieving high-efficiency compression, preserves the geographical integrity of the imagery through a scene-level index table, meeting the needs of analyzing the entire region and overcoming the shortcomings of traditional target extraction schemes that sacrifice integrity for efficiency, resulting in permanent information loss. The differentiated compression strategy ensures high-value density of the transmitted data, meeting the dual requirements of data integrity and accuracy for advanced applications such as macroscopic analysis and change detection.

[0044] 3. The device of this invention achieves accurate identification of data value and adaptive allocation of resources through on-orbit intelligent computing, dynamically adjusts the screening threshold and compression strategy according to the real-time status of the satellite (such as bandwidth and storage), and supports remote updates of lightweight AI models and task configurations (such as target type and threshold parameters) from the ground. This enables the satellite to flexibly switch between multiple task modes such as disaster emergency response, national land survey, agricultural monitoring, and marine observation, realizing "one satellite for multiple uses" and rapid task response, and enhancing the adaptability and practicality of on-orbit processing. Attached Figure Description

[0045] Figure 1 This is a flowchart of the remote sensing data on-orbit screening and compression method of the present invention;

[0046] Figure 2 This is a schematic diagram of the remote sensing data on-orbit screening and compression device of the present invention. Detailed Implementation

[0047] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0048] like Figure 1 As shown, this invention provides a systematic, adaptive, and configurable on-orbit screening and compression method for remote sensing data. Its overall technical solution is as follows: A satellite-based intelligent processing closed loop is constructed, centered on a lightweight AI model and a scene-level index table. This closed loop uses the entire scene as the basic processing unit, sequentially performing three core operations: dual screening based on quality and value, value-oriented hierarchical compression within the scene, and product encapsulation. Ultimately, it outputs a compressed data package with complete geographic information and extremely high value density for download, thereby solving the problems of data redundancy and transmission bottlenecks at the source. The method includes the following steps:

[0049] Step S0 (Autonomous Decision-Making): Real-time monitoring of key status parameters such as satellite-to-ground data transmission bandwidth, remaining onboard storage capacity, and remaining transit time;

[0050] When any of the monitoring parameters, such as satellite-to-ground data transmission bandwidth, remaining onboard storage capacity, and remaining transit time, are lower than their preset thresholds, the quality availability threshold in step S3 and / or the content value requirement in step S4 are dynamically adjusted. Specifically, this means tightening the quality or value judgment criteria to improve the data rejection rate or increase the compression ratio of low-value data.

[0051] The autonomous decision-making and update module monitors status parameters such as data transmission bandwidth and on-board storage capacity in real time. The threshold values ​​for each judgment are set based on matching the typical operating parameters and mission requirements of the satellite platform. For example, the trigger threshold for bandwidth-constrained mode (80Mbps) is set in the upper-middle range of the typical bandwidth range (50-200Mbps) of the satellite-to-ground data transmission link, aiming to intervene early when a perceptible decline in link quality occurs, balancing transmission efficiency and data integrity. The trigger threshold for emergency storage mode (15%) provides a safe operational margin for the on-board storage system, preventing mission interruptions due to storage exhaustion. Its dynamic adjustment strategy is implemented through a state machine, as follows:

[0052] When the autonomous decision-making and update module detects that the data transmission bandwidth is consistently below 50Mbps, it will automatically switch from "normal mode" to "bandwidth strain mode". In this bandwidth strain mode, the cloud coverage threshold in the "quality availability" judgment is dynamically tightened from 30% to 15%, and the compression ratio of low-value data blocks is automatically increased from 1:15 to 1:18. This prioritizes the transmission of the highest-value data when resources are scarce, achieving adaptive optimization of the processing strategy.

[0053] In addition, the on-orbit screening and compression device for remote sensing data supports multi-parameter linkage decision-making. For example, when both conditions are met, such as "remaining storage capacity is less than 15%" and "remaining transit time is less than 8 minutes", the "end-of-transit mode" strategy will be prioritized, which means that only the highest value data is retained and all low-value data processing is suspended in order to maximize the utilization of the downlink capacity in the final stage.

[0054] The autonomous decision-making and updating module is implemented as follows:

[0055] Status monitoring: The system reads the real-time bandwidth of the data transmission link layer and the remaining storage capacity returned by the SSD (Solid State Drive Controller) at a frequency of 1Hz through the satellite system bus, and calculates the remaining transit time in combination with the onboard clock and orbit forecast.

[0056] Decision state machine: Four preset modes:

[0057] Normal mode (bandwidth ≥ 80Mbps, storage ≥ 30%): cloud coverage threshold = 30%, low value compression ratio = 1:15.

[0058] Bandwidth strained mode (bandwidth <80Mbps): Cloud coverage threshold tightened to 20%, low-value compression ratio increased to 1:18.

[0059] Storage Emergency Mode (Storage <15%): Suspend low-value data processing procedures.

[0060] Terminal transit mode (time < 8 minutes): Discard all low-value data in the queue.

[0061] Update Mechanism: The update module resides in memory and listens to the RS422 telemetry and control interface. After receiving the policy configuration file from the ground, it performs an MD5 check. If the check passes, the new policy is written to the designated configuration area of ​​the SSD controller, and the autonomous decision-making module is notified to load the new configuration when processing the next scene of data.

[0062] Step S1, Data Acquisition; Acquire a complete raw remote sensing data or pre-processed data (such as radiometric correction and denoising) output by the remote sensing satellite observation payload;

[0063] On-orbit preprocessing of raw remote sensing data, including at least radiometric calibration and system geometric correction, is essential to generate standardized preprocessed data. This provides standardized data input with consistent radiometric and geometric references for subsequent intelligent quality assessment, content understanding, and geographic location-based data block indexing. It is a necessary prerequisite for ensuring the accuracy of the entire system's judgments and the usability of the product.

[0064] Step S2, Model Inference; The whole-view remote sensing data output after the preprocessing in Step S1 is input into the lightweight AI model deployed on the satellite.

[0065] The lightweight AI model is a convolutional neural network model with fewer than 5 million parameters, which has undergone structured pruning and 8-bit quantization. Furthermore, the power consumption for inference on standard-size single-scene data on the onboard embedded AI processor is less than 10W, and the inference latency is less than 1 second.

[0066] The lightweight AI model was built and optimized through ground training, and the specific process is as follows:

[0067] Step 21, Training Dataset Construction: On the ground, large-scale public remote sensing datasets (such as ImageNet, MS-COCO) are combined with self-built proprietary datasets (such as a self-built remote sensing image sample library containing targets such as buildings, road networks, water bodies, vegetation cover types, crops, clouds and shadows, ships, sea ice, disaster areas, and specified mineral or rock types). The combined data is augmented to construct a training sample set covering various land cover types, imaging conditions, and geographical environments. Data augmentation includes random rotation, flipping, color jittering, and simulating cloud cover to improve the model's generalization ability.

[0068] Step 22, Training Method and Process: The standard convolutional neural network is trained using a pre-training-fine-tuning approach. First, the standard convolutional neural network (such as MobileNetV3 or ResNet) is pre-trained on a large-scale public remote sensing dataset to acquire general visual feature extraction capabilities. Then, for predefined on-orbit screening tasks (such as quality assessment and crop identification), the standard convolutional neural network is fine-tuned using labeled data from the corresponding themes (i.e., a self-built proprietary dataset) to optimize its performance in the target scenario, resulting in a fine-tuned high-performance model.

[0069] Step 23, Model Optimization for On-board Environment: To meet the strict resource constraints of the on-board computing platform, the fine-tuned high-performance model undergoes deep optimization for embedded deployment.

[0070] Step 231, Structured Pruning: Based on the importance score of neurons or channels, remove redundant parts in the convolutional neural network that contribute little to the output, generating a sparse model, thereby significantly reducing the number of parameters and computational cost.

[0071] 8-bit integer quantization: The full-precision weights (32-bit floating-point numbers) and activation values ​​of the trained sparse model are mapped and converted into 8-bit integer (INT8) representations. This process involves quantization-aware training after conversion to compensate for precision loss and ensure that the sparse model maintains high accuracy even under low-precision computation.

[0072] Step 24, Hardware Co-optimization: Utilize the inference engine (such as TensorRT) provided by the on-board target processor (such as the NVIDIA Jetson series) to convert the optimized sparse model into a hardware-friendly format, further optimize the computation graph and accelerate the inference process.

[0073] Through the above systematic training and optimization process, a lightweight AI model with 3.8 million parameters was finally obtained, which meets the strict power consumption and latency requirements for on-board deployment.

[0074] The lightweight AI model used in this implementation, based on the MobileNetV3 architecture, has 3.8 million parameters after structured pruning and 8-bit integer quantization optimization. The measured average power consumption for inference on a single GF-2 scene on a Jetson AGX Xavier processor is 8.5W, with an average execution time of 0.9 seconds. This lightweight AI model achieves a value label prediction accuracy of 94.2% on the test set, fully meeting the lightweight and real-time requirements for on-board deployment.

[0075] Step S3: The lightweight AI model performs image quality assessment on the entire remote sensing data. If the quality parameters of the entire remote sensing data do not meet the preset availability threshold, it is determined to be "quality unavailable".

[0076] Image quality assessment parameters include at least the average cloud coverage percentage, average signal-to-noise ratio, and image blur. Availability thresholds are defined and adjusted through the currently active task strategy configuration, following a pre-defined task strategy rule base: during the ground mission planning phase, threshold templates for different scenarios are pre-set based on application objectives (e.g., agricultural monitoring, disaster emergency response); the onboard autonomous decision-making and update module dynamically adjusts the thresholds according to real-time resource status and rules embedded in the state machine. All thresholds are encapsulated as structured parameters in the task strategy configuration file, supporting remote injection and updates from the ground, ensuring image quality assessment can be completed under different task requirements and resource constraints.

[0077] Step S4: For whole-view remote sensing data of usable quality, further content understanding and value assessment are performed through a lightweight AI model. If the content value does not meet the preset requirements, it is determined to be "invalid content".

[0078] Content understanding and value assessment include, at a minimum, the identification and statistical analysis of one or more predefined targets of interest, such as buildings, road networks, water bodies, vegetation cover types, crops, clouds and shadows, ships, sea ice, disaster areas (e.g., fires, floods, landslides), and specified mineral or rock types; the predefined requirements for content value are defined and adjusted through the currently effective task strategy configuration.

[0079] Step S5 (Decision and Removal): Based on the above determination, the whole-view remote sensing data that is "unusable in quality" and "invalid in content" is directly deleted from the on-board storage, and only the whole-view remote sensing data that is "valid in content" is retained for subsequent processes.

[0080] Step S6 (Intra-scene segmentation and adaptive compression) divides the "content valid" whole-scene remote sensing data into multiple standard-sized data blocks;

[0081] For valid whole-view remote sensing data, it is first divided into multiple data blocks. Then, a lightweight AI model performs a rapid value assessment on each data block, generating a quantitative value label (such as "high value" or "low value") for each data block. The set of these value labels constitutes the value distribution information. Finally, by calling the on-board adaptive compression engine, based on the value distribution information provided by the lightweight AI model, a compression strategy and parameters are dynamically selected for each data block to implement differentiated compression: for high-value data blocks determined by the lightweight AI model to contain key targets or features, lossless compression or low compression ratio lossy compression with a compression ratio of 1:2 to 1:5 is used; for low-value data blocks determined by the lightweight AI model to lack key targets, high compression ratio lossy compression with a compression ratio of 1:10 to 1:25 is used.

[0082] This invention selects JPEG2000 as the core compression algorithm because its wavelet transform-based characteristics make it particularly suitable for remote sensing image processing: on the one hand, it supports a continuous compression spectrum from lossy to lossless, meeting the needs of differentiated compression; on the other hand, it supports region of interest encoding, which can directly realize the "region adaptive compression" function; at the same time, as an open international standard, there are already mature implementation schemes in satellite embedded systems.

[0083] The compression strategy is configured as follows:

[0084] High-value blocks: JPEG2000 lossy compression is used, wavelet transform level is set to 5, and target compression ratio is 1:2 to 1:5.

[0085] Low-value blocks: JPEG2000 lossy compression is used, wavelet transform level is set to 3, and target compression ratio is 1:10 to 1:25.

[0086] Region-adaptive compression: For low-value blocks containing a small number of targets, the ROI (Region of Interest Encoding) function of JPEG2000 is invoked. The target box is extended by 100 pixels as the region of interest. The region of interest is compressed using high-value parameters, while the background area is compressed using low-value parameters.

[0087] Quality verification: After compression, the engine decodes the data in memory and calculates the PSNR (Peak Signal-to-Noise Ratio) with the original data blocks. If the PSNR of a high-value block is less than 35dB or the PSNR of a low-value block is less than 28dB, the data is automatically recompressed with more conservative parameters (such as reducing the compression ratio by one level).

[0088] Step S7, Product Packaging: Generate a scene-level index table. This table records the pixel position of each data block in the original scene, its assigned value tag (e.g., "high value," "low value"), and its corresponding compression parameters (e.g., compression algorithm, compression ratio). Package all compressed data blocks together with this scene-level index table to form a complete, self-describing download data product.

[0089] Step S8 (data transmission scheduling) sends the packaged downlink data product to the ground station through the satellite-to-ground data transmission interface.

[0090] Deep integration of steps S0 with steps S1 to S8: When any status parameter (such as data transmission bandwidth below 50Mbps or remaining storage capacity below 20%) is detected to be below its preset threshold, the quality availability threshold in step S3 and / or the content value requirement in step S4 will be dynamically adjusted. For example, the judgment criteria will be tightened to improve the data rejection rate or to increase the compression ratio of low-value data.

[0091] A reserved interface (i.e., strategy update) receives mission strategy configuration files from the ground via the satellite-to-ground telemetry and control link. After verification, the on-orbit strategy configuration is updated and switched. When the remote sensing satellite observation mission mode changes (such as switching from agricultural monitoring to disaster emergency response), the on-orbit processing strategy can be quickly reconstructed by updating the image quality assessment threshold, target of interest type and value judgment criteria, and differentiated compression parameters in the configuration file, so as to adapt to the differentiated data screening and compression requirements of different application scenarios.

[0092] An on-orbit remote sensing data screening and compression device based on on-board intelligent computing is used to implement the on-orbit processing of the aforementioned on-orbit remote sensing data screening and compression methods. Its hardware foundation is a low-power embedded AI computing platform, with the specific architecture as follows: Figure 2 As shown, it integrates the following modules:

[0093] The data preprocessing module is used to carry lightweight on-orbit radiometric calibration and system geometry correction algorithms, providing input data that conforms to radiometric and geometric specifications for subsequent processing.

[0094] The data storage module is used to cache the entire scene remote sensing data to be processed, the compressed remote sensing data, and store on-orbit screening strategy parameters, compression strategy parameters, etc., such as lightweight AI models, image quality assessment, and content value assessment.

[0095] An embedded AI computing module is used to carry a lightweight AI model that meets predetermined power consumption and performance indicators, and to complete the inference of the lightweight AI model.

[0096] The data segmentation and compression module performs data block segmentation and value-oriented adaptive compression based on the quantization compression ratio range.

[0097] The data encapsulation module is responsible for generating scene-level index tables (see Table 1) containing key information such as value tags and completing the packaging of data products.

[0098] Table 1. Description of the structure of the scene-level index table

[0099]

[0100] The data transmission scheduling module manages the scheduling and transmission of downlink data products, sending priority data streams to the ground station.

[0101] The autonomous decision-making and update module, based on the strategy update instructions sent by the ground station and the status parameters of the satellite platform, realizes the monitoring of status parameters such as data transmission bandwidth and on-board storage capacity, image quality assessment, and content value assessment strategies.

[0102] To verify the practical effectiveness of this invention, three representative raw observation data from the GF-2 satellite were selected: the GF-2 satellite has a panchromatic resolution of 0.8 meters and a multispectral resolution of 3.2 meters. To simulate on-orbit processing conditions, this embodiment simulates the necessary on-orbit preprocessing, which employs standard methods in the field of satellite remote sensing.

[0103] In-orbit radiometric calibration: Based on the radiative transfer model and the calibration coefficients of the satellite payload, the original DN (Digital Number) values ​​are converted into physically meaningful apparent radiance or apparent reflectance to eliminate the influence of differences in sensor response. The algorithm implementation can refer to remote sensing data calibration protocols recommended by organizations such as CCSDS (Consultative Committee for Space Data Systems).

[0104] Simplified geometric correction in orbit: Based on rigorous geometric models such as collinearity equations, the satellite orbital parameters, attitude parameters and sensor geometric intrinsic parameters stored on the satellite are used to perform preliminary system-level geometric correction on the data, mainly to eliminate systematic distortions caused by factors such as Earth's rotation and changes in satellite attitude.

[0105] After the above preprocessing, the data can be used for subsequent intelligent filtering and compression, but it does not achieve the precise geometric accuracy required after ground processing. Specific information about the three selected scenes is as follows:

[0106] Scene A (No Crops): The imaging area is a suburban area of ​​the city, with features mainly consisting of buildings, roads, and bare soil, and no large-scale crop planting areas. This scene is considered "invalid content" or "low-value" data.

[0107] Scene B (partial crop area): The imaging area is a peri-urban area, including residential areas, roads, and scattered cornfields and vegetable gardens. This scene is used to test the system's ability to identify and classify areas of mixed value.

[0108] Scene C (Large-area crops): The imaging area is a large-scale farm, mainly consisting of contiguous corn and rice planting areas with a regular structure. This scene is considered "high-value" data.

[0109] The tests for this invention were conducted on a ground-based hardware platform simulating an onboard environment. The core configuration and module structure of this platform are as follows:

[0110] Hardware core: NVIDIA Jetson AGX Xavier (32 TOPS computing power, 20W power consumption), equipped with 512GB NVMe SSD.

[0111] Software environment: Ubuntu 18.04, PyTorch 1.10, OpenCV 4.5, OpenJPEG 2.4.

[0112] System Modules: The hardware platform and the software running on it together constitute the remote sensing data on-orbit screening and compression device. Its software functional modules (data reading, lightweight AI model, screening, compression, packaging, scheduling, decision-making and updating) correspond one-to-one with the various modules of the device.

[0113] The processing flow is as follows:

[0114] Step 1: Complete the overall scene screening according to steps S1-S5 above;

[0115] Input the three GF-2 scene data (approximately 2GB / scene) into the system sequentially.

[0116] Quality assessment: The lightweight AI model calculated the average cloud cover (all three scenes <5%) and signal-to-noise ratio (all >30dB) for each scene. Therefore, all three scenes were judged to be of "usable quality".

[0117] Content value assessment: A lightweight AI model performs content understanding on three sets of data.

[0118] Scene A (No crops): No large-scale crop targets were identified, and the content was deemed "invalid" according to the "Agricultural Monitoring" strategy.

[0119] Scene B (partial crops) and Scene C (large area of ​​crops): Both successfully identified crop targets and were judged as "valid content".

[0120] Decision and Removal: Based on the judgment result, the system directly deletes scene A from storage without further processing. Scenes B and C proceed to the next compression step.

[0121] Step 2: In-scene segmentation and adaptive compression;

[0122] Blocking: Divide scene B and scene C into data blocks of 512×512 pixels each.

[0123] Value judgment and differentiated compression: The adaptive compression engine calls the crop distribution map provided by the lightweight AI model to perform compression on each data block:

[0124] For data blocks where crop pixels account for more than 40%, low-compression lossy compression (JPEG2000, compression ratio ~1:5) is used.

[0125] For data blocks where crop pixels account for ≤40%, high compression ratio lossy compression (JPEG2000, compression ratio approximately 1:15) is used.

[0126] Step 3, Product packaging;

[0127] The system generates a scene-level index table for scene B and scene C respectively, and its data structure is shown in Table 2. Table 2 records the spatial location, value category, and actual compression ratio of each compressed data block in the original scene. Finally, all compressed data blocks of each scene are packaged with the scene-level index table into a complete data product.

[0128] Table 2. Examples of Scene-Level Indexes

[0129]

[0130] Step 4, data transmission scheduling;

[0131] The packaged data products are output through the data transmission scheduling module, completing the data product download to the ground station.

[0132] A comparative analysis of the data before and after processing is shown in Table 3;

[0133] Table 3 Data compression effect

[0134]

[0135] As shown in Table 3, by removing the entire scene and compressing the data in different scenes, the total transmission requirement of the three test data was reduced from 6.15GB to 1.38GB, and the total compression ratio reached 77.56%, which significantly improved the efficiency of satellite-to-ground data transmission.

[0136] To verify whether compression impairs the application value of the data, the downloaded Scene B and Scene C products were decompressed and restored at the ground station, and crop information extraction was performed. Using the same deep learning model (U-Net), pixel-level classification of crops (corn and rice) was performed on both the original uncompressed data and the data processed and decompressed according to this invention. The Intersection over Union (IoU) ratio was used as the core evaluation metric for classification accuracy.

[0137] For Scene B (some crops), the average IoU for crop classification before and after compression were 0.853 and 0.847, respectively, with a precision decrease of only 0.006.

[0138] For scenario C (large-area crops), the average IoU for crop classification before and after compression were 0.892 and 0.885, respectively, with a precision decrease of only 0.007.

[0139] In summary, the data processed by this invention achieves a reduction of over 62.5% in data volume while maintaining its most crucial application value—crop classification accuracy—almost perfectly (accuracy loss is less than 0.01%). This fully demonstrates that the adaptive compression strategy employed in this invention achieves an excellent balance between "transmission efficiency" and "data value".

[0140] The processing flow and effectiveness of this invention were fully demonstrated and verified using three representative GF-2 satellite data sets. The test results clearly show that:

[0141] This invention can intelligently filter out 100% invalid scene data, saving resources from the source.

[0142] This invention can achieve significant data reduction (72.68% for scene B and 61.32% for scene C) through in-scene hierarchical compression, with a comprehensive reduction rate of 77.56%.

[0143] This invention, under extreme compression, can effectively ensure the accuracy of downstream core applications (such as crop classification), achieving a balance between data quantity and quality.

Claims

1. A method for on-orbit screening and compression of remote sensing data based on on-board intelligent computing, characterized in that, After acquiring remote sensing data from the satellite platform, perform the following steps: D1, preprocessing the acquired whole-scene remote sensing data; D2 inputs the pre-processed whole-scene remote sensing data into a lightweight AI model deployed on the on-board computing platform; D3. The lightweight AI model is used to evaluate the image quality of the entire remote sensing data. If the quality parameters of the entire remote sensing data do not meet the preset availability threshold, it is determined that the quality is unusable. For usable whole-scene remote sensing data, the lightweight AI model is used to further understand the content and evaluate its value. If the content value does not meet the preset requirements, the content is deemed invalid. Full-view remote sensing data that is of unusable quality or invalid content will be directly deleted from the on-board storage and will not be downloaded; for full-view remote sensing data with valid content, the subsequent compression and download process will be performed. D4. For the whole-scene remote sensing data with valid content, the data is divided into blocks. The lightweight AI model performs a rapid value assessment on each data block and generates a quantitative value label for each data block. By calling the on-board adaptive compression engine, the compression strategy and parameters are dynamically selected for each data block to implement differentiated compression. D5 generates a scene-level index table by taking the pixel position of each data block in the original scene, the assigned value label, and the corresponding compression parameters. Then, all the compressed data blocks are packaged together with the scene-level index table to form a complete download data product with self-describing capabilities. D6 sends the packaged downlink data product to the ground station via the satellite-to-ground data transmission interface; The lightweight AI model was built and optimized through ground training, and the specific steps are as follows: D21, on the ground, uses a combination of large-scale public remote sensing datasets and a self-built proprietary dataset, and performs augmentation processing on the combined data to construct a training sample set covering various land cover types, imaging conditions and geographical environments; the self-built proprietary dataset is a self-built remote sensing image sample library containing buildings, road networks, water bodies, vegetation cover types, crops, clouds and shadows, ships, sea ice, disaster areas and specified mineral or rock types. D22 employs a pre-training-fine-tuning approach: First, a standard convolutional neural network is pre-trained using a large-scale public remote sensing dataset to enable it to extract general visual features. Then, for a predefined on-orbit screening task, a self-built proprietary dataset is used to fine-tune the standard convolutional neural network, resulting in a high-performance model after fine-tuning. D23, based on the importance score of neurons or channels, performs structured pruning on the fine-tuned high-performance model, removes redundant parts in the convolutional neural network that contribute little to the output, and generates a sparse model. The 32-bit floating-point numbers and activation values ​​of the trained sparse model are mapped and converted into 8-bit integer representations to obtain the optimized sparse model. D24 utilizes the inference engine provided by the onboard target processor to convert the optimized sparse model into a hardware-friendly format, further optimizing the computation graph and accelerating the inference process.

2. The method for on-orbit screening and compression of remote sensing data based on on-board intelligent computing according to claim 1, characterized in that, The lightweight AI model is a convolutional neural network model with fewer than 5 million parameters, which has undergone structured pruning and 8-bit quantization. Furthermore, the power consumption for inference on standard-size data of a single scene on the onboard embedded AI processor is less than 10 watts, and the inference latency is less than 1 second.

3. The method for on-orbit screening and compression of remote sensing data based on on-board intelligent computing according to claim 1, characterized in that, The evaluation parameters for image quality assessment include at least the average cloud coverage ratio of the entire scene, the average signal-to-noise ratio of the entire scene, and image blur. The definition and adjustment of the availability threshold follow the preset task strategy rule library: the ground mission planning stage pre-sets threshold templates for different scenarios according to the application objectives; the on-board autonomous decision-making and update module dynamically adjusts the threshold according to the rules fixed in the state machine based on the real-time resource status. All thresholds are encapsulated in the form of structured parameters in the task policy configuration file, supporting remote injection and updates from the ground. The content understanding and value assessment includes at least the identification and statistical analysis of one or more preset targets of interest, such as buildings, road networks, water bodies, vegetation cover types, crops, clouds and shadows, ships, sea ice, disaster areas, and specified mineral or rock types; the preset requirements for content value are defined and adjusted through the currently effective task strategy configuration.

4. The method for on-orbit screening and compression of remote sensing data based on on-board intelligent computing according to claim 1, characterized in that, The steps to achieve differentiated compression include: D41 divides the full-view remote sensing data with valid content into multiple data blocks; D42, the lightweight AI model performs a rapid value assessment on each data block and generates a quantified value tag for each data block. The set of these value tags constitutes value distribution information. D43, by calling the on-board adaptive compression engine, dynamically selects compression strategies and parameters for each data block based on the value distribution information, and implements differentiated compression.

5. The method for on-orbit screening and compression of remote sensing data based on on-board intelligent computing according to claim 4, characterized in that, The differentiated compression includes: for high-value data blocks determined to contain key targets or features, lossless compression or lossy compression with a low compression ratio of 1:2 to 1:5 is used; for low-value data blocks determined to have no key targets, lossy compression with a high compression ratio of 1:10 to 1:25 is used.

6. The method for on-orbit screening and compression of remote sensing data based on on-board intelligent computing according to any one of claims 1-5, characterized in that, The satellite receives mission strategy configuration files from the ground via the satellite-to-ground telemetry and control link. After verification, the satellite updates and switches its on-orbit strategy configuration. When the remote sensing satellite observation mission mode changes, the satellite quickly reconstructs its on-orbit processing strategy by updating the image quality assessment threshold, target of interest type and value judgment criteria, and differentiated compression parameters in the configuration file.

7. A remote sensing data on-orbit screening and compression device based on on-board intelligent computing, used to implement the remote sensing data on-orbit screening and compression method based on on-board intelligent computing as described in any one of claims 1-6, and to complete the on-orbit processing of remote sensing data, characterized in that, Based on an embedded AI computing platform, the following modules are integrated: The data preprocessing module is used to carry lightweight on-orbit radiometric calibration and system geometry correction algorithms, providing input data that conforms to radiometric and geometric specifications; The data storage module is used to cache the whole-scene remote sensing data to be processed, the compressed remote sensing data, and store the lightweight AI model, on-orbit screening strategy parameters and compression strategy parameters. An embedded AI computing module is used to carry and execute lightweight AI models that meet predetermined power consumption and performance indicators; The data segmentation and compression module is used to perform data block segmentation and value-oriented adaptive compression based on the quantization compression ratio range; The data encapsulation module is responsible for generating scene-level index tables containing value tags and packaging data products. The data transmission scheduling module manages the scheduling and transmission of downlink data products, sending priority data streams to the ground station. The autonomous decision-making and update module enables status monitoring, policy adjustment, and remote update functions based on the policy update instructions sent by the ground station and the status parameters of the satellite platform.

Images