A Deep Learning-Based On-Satellite Image Processing Method and System
The remote sensing image on-board processing method based on deep learning utilizes deep sparse coding networks for feature mapping and sparse reconstruction, combined with resource awareness and hierarchical coding, to solve the resource constraint problem in remote sensing image on-board processing and achieve a balance between efficient downlink and high-quality reconstruction.
Patent Information
- Application Number
- CN202611243860.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-17
- Publication Date
- 2026-11-17
AI Technical Summary
Existing on-board processing methods for remote sensing images fail to effectively address the strict constraints of computing, storage, and power consumption resources, leading to processing delays or system crashes. They also lack differentiated resource allocation based on feature importance and awareness of satellite-to-ground link status, making it difficult to achieve efficient downlink transmission and high-quality reconstruction.
A deep learning-based on-board remote sensing image processing method is adopted. Multi-layer sparse feature mapping is performed through a deep sparse coding network, resource occupancy index is calculated, adaptive feature distillation and resource-aware sparse reconstruction are performed, difference-coded data is generated and hierarchical coding and priority arrangement are performed to ensure that key data is transmitted first.
It achieves stable processing under resource-constrained conditions, significantly reduces data volume while retaining key feature information, improves the fidelity of image reconstruction and the transmission efficiency of satellite-to-ground links, and achieves an optimal balance in resource utilization.
Smart Images

Figure CN123123565A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of remote sensing image processing technology, and in particular to a satellite-based remote sensing image processing method and system based on deep learning. Background Technology
[0002] With the development of aerospace remote sensing technology, the resolution of spaceborne optical sensors is constantly improving, and the amount of data in a single remote sensing image is growing exponentially. However, the bandwidth of the space-to-ground link is relatively limited, making it difficult to meet real-time requirements for transmitting all the massive amounts of raw data. Traditional on-board remote sensing image processing methods mostly use fixed transform coding, which has the following limitations: it does not fully consider the strict constraints of on-board computing, storage, and power consumption resources; fixed compression strategies are prone to processing delays or system crashes when resources are scarce; it uses uniform processing for all regions of the image, lacking differentiated resource allocation based on feature importance, and key features are easily lost during compression; and it lacks a perception and adaptive mechanism for the space-to-ground link status, making it difficult to achieve a balance between transmission efficiency and reconstruction quality. Although deep learning methods have significant advantages in the field of image processing, direct deployment on space faces challenges such as large computational load, high storage requirements, and high power consumption.
[0003] Therefore, there is an urgent need for a remote sensing image on-board processing method that can sense the status of on-board resources, adaptively adjust processing strategies, and achieve efficient downlinking and high-quality reconstruction. Summary of the Invention
[0004] This invention proposes a deep learning-based on-board processing method for remote sensing images, comprising: S1. The on-board processing system receives remote sensing images and initiates a deep learning-based on-board remote sensing image processing method. S2. The on-board processing system inputs remote sensing images into a pre-trained deep sparse coding network, performs multi-layer sparse feature mapping, generates a multi-layer sparse feature mapping image set, and calculates the on-board comprehensive resource occupancy index of the image set. S3. Based on the on-board comprehensive resource occupancy index, the on-board processing system performs adaptive feature distillation on the multi-layer sparse feature mapping image set to generate a distilled feature image set. Then, it performs resource-aware sparse reconstruction on the distilled feature image set to generate a reconstructed feature image set and calculates the downlink distortion sensitivity of the reconstructed feature image set. S4. Based on the downlink distortion sensitivity, the on-board processing system performs hierarchical differential encoding on the reconstructed feature image set to generate differential encoded data. Then, it performs on-board encapsulation and priority arrangement on the differential encoded data to generate encapsulated and arranged data packets, and outputs the encapsulated and arranged data packets to the downlink transmission queue. S5, the on-board processing system extracts encapsulated and arranged data packets from the downlink transmission queue and outputs them to the ground receiving station via the satellite-to-ground link to complete the on-board processing of remote sensing images.
[0005] The on-board remote sensing image processing method based on deep learning, as described above, involves an on-board processing system receiving remote sensing images and initiating the deep learning-based on-board remote sensing image processing method, which includes the following sub-steps: The satellite connects to the onboard optical sensor via a high-speed data interface to monitor the sensor's data output status in real time. Once the onboard optical sensor has completed the acquisition of a remote sensing image, the acquired remote sensing image data is transmitted to the data buffer unit of the onboard processing system via a high-speed serial interface. Complete remote sensing image data is read from the data buffer unit, and the frame header identifier and frame tail check code of the image data are verified to confirm the integrity and validity of the data transmission. After confirming that the data is complete and valid, the on-board processing system loads the remote sensing images into the processing memory area, initializes the processing parameters and processing buffer required for the deep learning-based on-board remote sensing image processing method, and begins to execute the on-board processing flow.
[0006] The on-board remote sensing image processing method based on deep learning, as described above, involves the on-board processing system inputting remote sensing images into a pre-trained deep sparse coding network, performing multi-layer sparse feature mapping to generate a multi-layer sparse feature mapping image set, and calculating the on-board comprehensive resource occupancy index of the image set. This includes the following sub-steps: The feature mapping unit of the on-board processing system inputs remote sensing images into a pre-trained deep sparse coding structure. Through multi-layer convolution and sparse activation operations, it extracts multi-scale sparse features from the remote sensing images and generates a multi-layer sparse feature mapping image set containing a preset number of feature layers. During the execution of multi-layer sparse feature mapping, the resource monitoring unit of the on-board processing system collects the usage of computing resources, storage resources, and power consumption resources in real time, and calculates the on-board comprehensive resource occupancy index based on the preset computing resource weights, storage resource weights, and power consumption resource weights. The calculated on-board comprehensive resource occupancy index is compared with preset high occupancy thresholds and low occupancy thresholds to determine the current on-board resource occupancy level, which includes high occupancy level, medium occupancy level and low occupancy level.
[0007] The on-board remote sensing image processing method based on deep learning, as described above, involves the on-board processing system performing adaptive feature distillation on a multi-layer sparse feature mapping image set based on the on-board comprehensive resource occupancy index, generating a distilled feature image set, then performing resource-aware sparse reconstruction on the distilled feature image set to generate a reconstructed feature image set, and calculating the downlink distortion sensitivity of the reconstructed feature image set. This includes the following sub-steps: Based on the determined resource occupancy level, a strategy with corresponding strength is selected from a variety of preset feature distillation strategies. The feature channels in the multi-layer sparse feature mapping image set are discarded or merged according to their importance ranking to generate a distillation feature image set. The resource-aware sparse reconstruction unit of the on-board processing system uses the distillation feature image set as the observation value and a pre-trained sparse dictionary as the sparse basis to solve the reconstruction feature image set through an iterative reconstruction algorithm. For each feature map in the reconstructed feature image set, calculate its mean spatial gradient magnitude and frequency domain energy concentration. Combine the two indicators of mean spatial gradient magnitude and frequency domain energy concentration to calculate the distortion sensitivity of each feature map. Then, take the weighted average of the distortion sensitivity of all feature maps to obtain the downlink distortion sensitivity of the entire reconstructed feature image set.
[0008] The on-board remote sensing image processing method based on deep learning described above includes the following sub-steps: Based on downlink distortion sensitivity, the on-board processing system performs hierarchical differential encoding on the reconstructed feature image set to generate differential encoded data. Then, it performs onboard encapsulation and priority orchestration on the differential encoded data to generate encapsulated and orchestrated data packets, and outputs these packets to the downlink transmission queue. Based on the downlink distortion sensitivity, the on-board processing system determines the number of levels of hierarchical differential coding and the code rate allocation scheme for each level. The feature maps in the reconstructed feature image set are divided into multiple sensitivity levels according to the preset distortion sensitivity threshold, and different coding rates and error correction coding strengths are assigned to different sensitivity levels. Based on the classification results, difference encoding is performed on each feature map in the reconstructed feature image set. For the first feature map in each feature layer, quantization encoding is performed directly. For subsequent feature maps in the same feature layer, the pixel difference between the current feature map and the previous feature map at the same spatial position is calculated, and entropy encoding is performed on the difference matrix. Combine the encoded data of all feature maps according to a preset format, and add hierarchical labels and length information to generate difference encoded data.
[0009] The on-board remote sensing image processing method based on deep learning described above, wherein the on-board processing system extracts encapsulated and orchestrated data packets from the downlink transmission queue and outputs them to the ground receiving station via the satellite-to-ground link to complete the on-board processing of remote sensing images, includes the following sub-steps: The differential encoded data is encapsulated by adding a frame synchronization header, frame length identifier, source identifier, priority label, and frame check sequence according to the spaceborne data encapsulation format standard. The priority label is set according to the sensitivity level of the differential encoded data, with the highest priority label set for high-sensitivity data and the lowest priority label set for low-sensitivity data. The on-board processing system sorts the encapsulated data frames according to priority tags. High-priority data frames are arranged at the front of the transmission queue, and low-priority data frames are arranged at the back. Data frames of the same priority are arranged in order of encapsulation time. The sorted encapsulated data frames are combined into encapsulated and arranged data packets according to the queue order and output to the downlink transmission queue of the on-board processing system.
[0010] The on-board remote sensing image processing method based on deep learning described above, wherein the sorted encapsulated data frames are combined into encapsulated and arranged data packets according to the queue order and output to the downlink transmission queue of the on-board processing system, includes the following sub-steps: The onboard processing system monitors the status of the downlink transmission queue in real time. When the downlink transmission queue is not empty and the satellite-to-ground link transmission resources are available, it extracts, encapsulates, and arranges data packets from the head of the queue and transmits the data packets to the onboard data transmitter. The onboard data transmitter modulates the packaged and arranged data packets onto the radio frequency carrier, which is then radiated into free space in the form of electromagnetic waves through the onboard antenna and transmitted to the ground receiving station via the space-to-ground link. The antenna system of the ground receiving station captures radio frequency signals, demodulates and recovers the encapsulated and arranged data packets, completes the synchronization and integrity verification of data frames according to the synchronization header and check information in the encapsulation format, identifies the data type according to the priority label, and completes decoding and reconstruction according to the hierarchical label to recover the remote sensing image.
[0011] This invention also proposes a deep learning-based on-board remote sensing image processing system, comprising: Image receiving and initiation module: The on-board processing system receives remote sensing images and initiates a deep learning-based on-board remote sensing image processing method; Feature mapping and resource awareness module: The on-board processing system inputs remote sensing images into a pre-trained deep sparse coding network, performs multi-layer sparse feature mapping, generates a multi-layer sparse feature mapping image set, and calculates the on-board comprehensive resource occupancy index of the image set. Feature distillation and reconstruction module: Based on the on-board comprehensive resource occupancy index, the on-board processing system performs adaptive feature distillation on the multi-layer sparse feature mapping image set to generate a distilled feature image set, and then performs resource-aware sparse reconstruction on the distilled feature image set to generate a reconstructed feature image set, and calculates the downlink distortion sensitivity of the reconstructed feature image set. Hierarchical coding and encapsulation module: Based on the downlink distortion sensitivity, the on-board processing system performs hierarchical differential coding on the reconstructed feature image set to generate differential coded data, then performs on-board encapsulation and priority orchestration on the differential coded data to generate encapsulated and orchestrated data packets, and outputs the encapsulated and orchestrated data packets to the downlink transmission queue. Downlink output module: The on-board processing system extracts encapsulated and arranged data packets from the downlink transmission queue and outputs them to the ground receiving station via the satellite-to-ground link to complete the on-board processing of remote sensing images.
[0012] The beneficial effects achieved by this invention are as follows: By monitoring the real-time usage of onboard computing, storage, and power resources, a comprehensive onboard resource occupancy index is calculated, and the feature distillation strategy is adaptively adjusted accordingly. This proactively reduces processing complexity during resource-constrained periods, effectively ensuring the stable operation of the onboard processing system under strict resource constraints. Through multi-layer sparse feature mapping and resource-aware sparse reconstruction, key feature information is preserved while significantly reducing data volume. Combined with quantitative assessment of downlink distortion sensitivity, a scientific basis for differentiated coding is provided, effectively improving the fidelity of ground-based image reconstruction. A hierarchical differential coding method is employed, assigning differentiated coding rates and error correction coding strengths to different feature maps based on feature distortion sensitivity. Encapsulated data frames are sorted and transmitted according to priority tags, ensuring that highly sensitive key data is transmitted first, achieving an optimal balance between satellite-to-ground link transmission efficiency and reconstruction quality. This invention constructs a complete adaptive processing link from onboard reception to ground reconstruction, significantly improving the intelligence level and engineering practical value of onboard remote sensing image processing. Attached Figure Description
[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings.
[0014] Figure 1 This is a flowchart of a satellite-based remote sensing image processing method based on deep learning, provided in an embodiment of this application.
[0015] Figure 2 This is a schematic diagram of a satellite-based remote sensing image processing system based on deep learning, provided in an embodiment of this application. Detailed Implementation
[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0017] Example 1 like Figure 1As shown in the figure, an embodiment of this application provides a satellite-based remote sensing image processing method based on deep learning, comprising: Step S1: The on-board processing system receives remote sensing images and initiates a deep learning-based on-board remote sensing image processing method. Specifically, the onboard processing system connects to the onboard optical sensor via a high-speed data interface, monitors the sensor's data output status in real time, and transmits the image data to the data buffer unit via a high-speed serial interface after the onboard optical sensor completes the acquisition of a remote sensing image. The system reads the complete image data from the buffer unit, verifies the frame header identifier and frame tail checksum to confirm the integrity and validity of the data transmission, and loads the remote sensing image into the processing memory area after confirming the data is correct. It then initializes the processing parameters and buffers required for the processing method and executes the onboard processing flow, which includes the following sub-steps: Step S11: Connect to the onboard optical sensor via a high-speed data interface and monitor the sensor data output status in real time. After the onboard optical sensor completes the acquisition of a remote sensing image, transmit the acquired remote sensing image data to the data buffer unit of the onboard processing system via a high-speed serial interface. The high-speed data interface adopts the SpaceWire interface standard or Camera Link interface standard suitable for the spaceborne environment to ensure the reliability and real-time performance of on-orbit data transmission. The on-board processing system continuously queries the sensor data output status register. Once the valid frame signal is detected, the direct memory access transmission mechanism is activated to write the remote sensing image data into the data buffer unit and simultaneously record the number of rows, columns, bands, and acquisition timestamp information of the image.
[0018] Step S12: Read the complete remote sensing image data from the data buffer unit, verify the frame header identifier and frame tail check code of the image data, and confirm the integrity and validity of the data transmission. The frame header identifier contains a unique image identification code, acquisition timestamp, and image size information. The frame tail check code is generated using a cyclic redundancy check algorithm. After the on-board processing system reads data from the data buffer unit, it first locates the frame header identifier and parses the metadata. Then, it uses the same cyclic redundancy check algorithm to calculate the entire frame data and compares the calculation result with the frame tail check code. If they match, the remote sensing image data transmission is considered complete and valid. If they do not match, a retransmission request is triggered or an error log is recorded to prevent erroneous data from entering subsequent processing stages.
[0019] Step S13: After confirming that the data is complete and valid, the on-board processing system loads the remote sensing image into the processing memory area, initializes the processing parameters and processing buffer required by the deep learning-based on-board remote sensing image processing method, and begins to execute the on-board processing process. The processing memory area is divided into an input buffer, a feature buffer, an encoding buffer, and an output buffer. The size of each buffer is dynamically allocated according to the size of the remote sensing image and the preset maximum number of feature layers. At the same time, the pre-learned deep sparse coding network parameters, including the convolution kernel coefficients of each layer, sparse activation thresholds, maximum number of iterations, and convergence error limits, are loaded from the on-board non-volatile memory. The sparse dictionary parameters are also initialized to configure the initial computing environment for subsequent multi-layer sparse feature mapping, adaptive feature distillation, and resource-aware sparse reconstruction processing.
[0020] Step S2: The on-board processing system inputs the remote sensing images into a pre-trained deep sparse coding network, performs multi-layer sparse feature mapping, generates a multi-layer sparse feature mapping image set, and calculates the on-board comprehensive resource occupancy index of the image set. Specifically, remote sensing images are input into a pre-trained deep sparse coding structure, and multi-scale sparse features are extracted through multi-layer convolution and sparse activation operations to generate a multi-layer sparse feature mapping image set. Simultaneously, the usage of computing resources, storage resources, and power consumption resources is collected in real time, and an on-board comprehensive resource occupancy index is calculated according to preset weights. This index is then compared with preset high, medium, and low occupancy thresholds to determine the current resource occupancy level. The process includes the following sub-steps: Step S21: The feature mapping unit of the on-board processing system inputs the remote sensing image into the pre-trained deep sparse coding structure, and extracts the multi-scale sparse features of the remote sensing image through multi-layer convolution and sparse activation operations to generate a multi-layer sparse feature mapping image set containing a preset number of feature layers. The deep sparse coding network employs a residual sparse autoencoder architecture, consisting of multiple sequentially stacked coding layers. Each coding layer sequentially includes convolution operations, batch normalization operations, and nonlinear sparse activation operations. The remote sensing image is first input into the first coding layer. In the convolution operation, a sliding window convolution is performed on the input image using a convolution kernel of a preset size to generate an initial set of feature maps. Subsequently, batch normalization is used to standardize the feature maps to accelerate convergence and prevent feature distribution drift. Finally, a sparse activation function is used to perform a nonlinear mapping on the feature maps, setting feature responses below a preset sparse threshold to zero. The feature response value after sparse activation is expressed by the following formula: in, Indicates the first In the coding layer, the first The spatial location of each convolutional kernel The characteristic response value output at the location; Represents the hyperbolic tangent activation function; This represents the preset activation slope control factor; This represents the corresponding convolution response value; Indicates the first Sparsity adjustment coefficients for each coding layer Indicates the first The global maximum value among all convolutional kernel response values of the coding layer.
[0021] The sparse feature map output from the previous layer is used as the input to the next encoding layer. After progressive processing, a multi-layer sparse feature map image set covering low-level detail features to high-level semantic features is finally generated. The shallow feature map retains rich spatial texture information, while the deep feature map encodes high-level semantic structure information.
[0022] Step S22: During the execution of multi-layer sparse feature mapping, the resource monitoring unit of the on-board processing system collects the usage of computing resources, storage resources, and power consumption resources in real time, and calculates the on-board comprehensive resource occupancy index based on the preset computing resource weight, storage resource weight, and power consumption resource weight. The resource monitoring unit interacts with the hardware management module of the on-board processing system. It obtains the instantaneous load and computational workload of the processor by reading the performance counter as the computing resource usage, the size of the allocated dynamic storage space by reading the storage controller as the storage resource usage, and the current power consumption value by reading the power management unit as the power resource usage. After obtaining the above three resource usages, the on-board comprehensive resource occupancy index is calculated by weighted summation. The sum of the weights of computing resources, storage resources, and power resources is one, and each weight can be dynamically adjusted according to the current working mode of the satellite, thereby accurately reflecting the overall resource consumption level of the on-board processing tasks.
[0023] Step S23: Compare the calculated on-board comprehensive resource occupancy index with the preset high occupancy threshold and low occupancy threshold to determine the current on-board resource occupancy level, which includes high occupancy level, medium occupancy level and low occupancy level; When the overall on-board resource occupancy index is greater than or equal to the preset high occupancy threshold, the current resource occupancy level is determined to be high, indicating that the on-board computing and storage resources are approaching saturation, and an aggressive feature simplification strategy needs to be adopted to avoid resource overload. When the overall on-board resource occupancy index is less than or equal to the preset low occupancy threshold, the current resource occupancy level is determined to be low, indicating that the on-board resources are sufficient, and more feature information can be retained to improve image reconstruction quality. When the overall on-board resource occupancy index is between the low and high occupancy thresholds, the current resource occupancy level is determined to be medium, and a moderately compromised feature processing strategy is adopted to balance resource consumption and image quality.
[0024] Step S3: Based on the on-board integrated resource occupancy index, the on-board processing system performs adaptive feature distillation on the multi-layer sparse feature mapping image set to generate a distilled feature image set, and then performs resource-aware sparse reconstruction on the distilled feature image set to generate a reconstructed feature image set, and calculates the downlink distortion sensitivity of the reconstructed feature image set. Specifically, based on the resource occupancy level, a feature distillation strategy of corresponding strength is selected. Feature channels in the multi-layer sparse feature mapping image set are discarded or merged according to their importance ranking to generate a distilled feature image set. Using the distilled feature image set as the observation value and a pre-trained sparse dictionary as the sparse basis, an iterative reconstruction algorithm is used to solve for the reconstructed feature image set. The mean spatial gradient magnitude and frequency domain energy concentration are calculated for each feature map in the reconstructed feature image set. The distortion sensitivity of each feature map is calculated by combining both. Then, a weighted average of the distortion sensitivity of all feature maps is taken to obtain the overall downlink distortion sensitivity of the reconstructed feature image set. This process includes the following sub-steps: Step S31: Based on the determined resource occupancy level, select the strategy with the corresponding strength from a variety of preset feature distillation strategies, and discard or merge each feature channel in the multi-layer sparse feature mapping image set according to its importance ranking to generate a distillation feature image set. The preset feature distillation strategies include a lightweight simplification strategy, a moderate compression strategy, and a deep simplification strategy. When resource usage is low, the lightweight simplification strategy is selected, discarding only a small number of redundant feature channels at the bottom of the importance ranking, while retaining all other feature channels. When resource usage is medium, the moderate compression strategy is selected, discarding the last channels and weighted merging of some adjacent and highly similar feature channels to reduce the number of channels while retaining key information. When resource usage is high, the deep simplification strategy is selected, significantly increasing the number of channels discarded and performing more aggressive merging on the retained channels. The importance ranking of each feature channel is obtained by calculating the response energy statistical value of each channel's feature map, expressed by the following formula: Importance score of each feature channel: in, Indicates the first Importance score of each feature channel; and These represent the height and width of the feature channel, respectively. express Each feature channel in spatial location The characteristic response value at that location; This represents the preset peak response weighting coefficient. A larger response energy indicates that the channel contains more information and is more important, thus achieving adaptive feature simplification under different resource constraints.
[0025] Step S32: The resource-aware sparse reconstruction unit of the on-board processing system uses the distillation feature image set as the observation value and the pre-trained sparse dictionary as the sparse basis to solve the reconstruction feature image set through an iterative reconstruction algorithm. The pre-learned sparse dictionary is obtained offline using large-scale sample data with the same distribution as the satellite remote sensing images. It is trained using a dictionary learning algorithm, enabling efficient sparse representation of the feature maps of the remote sensing images. The iterative reconstruction algorithm employs either orthogonal matching pursuit or iterative threshold shrinkage. In each iteration, the correlation coefficient between the current residual and the atoms in the sparse dictionary is calculated, and the best-matching atom is selected to update the support set. The following formula is used to calculate the... In the nth iteration Correlation coefficients between sparse dictionary atoms and the current residual signal: in, Indicates the first In the nth iteration The correlation coefficients between sparse dictionary atoms and the current residual signal; Indicates the first The residual signal after the next iteration; Represents the first in a sparse dictionary One atom; Indicates inner product operation; Denotes the Euclidean norm; This represents a preset small regularization constant used to prevent the denominator from being zero; After atom selection is completed, the reconstruction coefficients are updated using the following formula: in, Indicates the first The reconstructed coefficient vector after the next iteration update; Indicates the first The residual signal after the next iteration; Indicates the current support set The submatrix formed by the dictionary atoms corresponding to the indexes in the middle; This represents the vector of reconstructed coefficients to be solved in the objective function; This represents the preset regularization baseline strength coefficient; This indicates the current overall resource utilization index on the satellite; This represents the nominal resource occupancy index value; This represents the sum of the absolute values of all non-zero elements in the vector.
[0026] Then, the reconstruction coefficients are updated by least squares solution until the residual energy is lower than the preset convergence error limit or the number of iterations reaches the upper limit, and finally the reconstructed feature image set is obtained. During the iteration process, resource awareness is reflected in the dynamic adjustment of the upper limit of the number of iterations and the convergence error limit according to the current resource occupancy level. The higher the resource occupancy level, the lower the upper limit of the number of iterations and the larger the convergence error limit, so as to reduce the computational overhead. Conversely, the reconstruction accuracy is improved, thereby maximizing the preservation of image information while ensuring on-board real-time performance.
[0027] Step S33: Calculate the mean spatial gradient magnitude and frequency domain energy concentration of each feature map in the reconstructed feature image set. Combine the two indicators of mean spatial gradient magnitude and frequency domain energy concentration to calculate the distortion sensitivity of each feature map. Then, take the weighted average of the distortion sensitivity of all feature maps to obtain the downlink distortion sensitivity of the entire reconstructed feature image set. The mean spatial gradient magnitude reflects the spatial activity and edge detail richness of the feature map. A larger mean gradient magnitude indicates a higher requirement for preserving spatial details, and consequently, a higher sensitivity to distortion. Frequency domain energy concentration is obtained by calculating the proportion of low-frequency energy to total energy after performing a Fast Fourier Transform on the feature map. A higher energy concentration indicates that the signal energy is more concentrated in the low-frequency region, and the loss of high-frequency components has a smaller impact on visual quality, resulting in lower distortion sensitivity. The following formula represents the value of the first (i) in the reconstructed feature image set. Downlink distortion sensitivity of each feature map: in, Indicates the first The downlink distortion sensitivity of a feature map. The larger the value, the more sensitive the feature map is to distortion introduced by the transmission link, and the higher the protection priority it should be given in downlink transmission. and These represent spatial texture weights and frequency domain distribution weights, respectively. and They represent the first The height and width of each feature map; Representing the spatial location of the feature map The spatial gradient magnitude at a given location is calculated by taking the square root of the sum of the squares of the horizontal and vertical gradients. This value reflects the degree of change between the pixel at that location and its neighboring pixels. The larger the gradient magnitude, the richer the spatial detail information contained at that location. Indicates the first Low-frequency energy values of each feature map; Indicates the first The total energy value of each feature map; This represents a preset regularization constant used to prevent the denominator from being zero; When combining the two indicators, the mean value of the spatial gradient magnitude is multiplied by a preset first weighting coefficient, and the insufficient amount of frequency domain energy concentration is multiplied by a preset second weighting coefficient. The sum of the two results in the distortion sensitivity of the feature map. When the distortion sensitivity of all feature maps is weighted and averaged, the weight of each feature map is determined by the spatial resolution layer in which it is located. Shallow feature maps are given higher weights because they contain more spatial detail information. This results in the downlink distortion sensitivity, which can objectively characterize the anti-distortion ability of the reconstructed feature image set in downlink transmission.
[0028] Step S4: Based on the downlink distortion sensitivity, the on-board processing system performs hierarchical differential encoding on the reconstructed feature image set to generate differential encoded data. Then, it performs on-board encapsulation and priority orchestration on the differential encoded data to generate encapsulated and orchestrated data packets, and outputs the encapsulated and orchestrated data packets to the downlink transmission queue. Specifically, based on the downlink distortion sensitivity, the number of levels and the bitrate allocation for each level of hierarchical differential encoding are set. Feature maps in the reconstructed feature image set are divided into different sensitivity levels according to a preset threshold, and correspondingly differentiated encoding bitrates and error correction encoding strengths are configured. When performing differential encoding on feature maps, the first feature map within the same feature layer is directly quantized and encoded. Subsequent feature maps calculate the pixel difference at the same spatial location as the previous feature map, and entropy encoding is applied to the difference matrix. The encoded data of all feature maps are summarized in a predetermined format, with hierarchical labels and length information appended, ultimately generating differential encoded data. This process includes the following sub-steps: Step S41: Based on the downlink distortion sensitivity, the on-board processing system determines the number of levels of the hierarchical differential coding and the code rate allocation scheme for each level. The feature maps in the reconstructed feature image set are divided into multiple sensitivity levels according to the preset distortion sensitivity threshold, and different coding rates and error correction coding strengths are allocated to different sensitivity levels. The preset distortion sensitivity thresholds include a first sensitivity threshold and a second sensitivity threshold, where the first sensitivity threshold is greater than the second sensitivity threshold. Feature maps with distortion sensitivity greater than or equal to the first sensitivity threshold are classified as high-sensitivity, and assigned the highest coding rate and the strongest error correction coding strength to maximize the protection of the integrity of key feature information during satellite-to-ground link transmission. Feature maps with distortion sensitivity between the second and first sensitivity thresholds are classified as medium-sensitivity, and assigned a moderate coding rate and error correction coding strength. Feature maps with distortion sensitivity less than or equal to the second sensitivity threshold are classified as low-sensitivity, and assigned a lower coding rate and weaker error correction coding strength. Thus, under the condition of limited total transmission bandwidth, priority is given to ensuring the transmission reliability of high-value data, significantly improving the overall transmission efficiency of the downlink.
[0029] Step S42: Based on the hierarchical results, perform difference encoding on each feature map in the reconstructed feature image set. For the first feature map in each feature layer, perform quantization encoding directly. For subsequent feature maps in the same feature layer, calculate the pixel difference between the current feature map and the previous feature map at the same spatial position, and perform entropy encoding on the difference matrix. Quantization encoding employs either uniform or non-uniform quantization. The quantization step size is determined based on the bitrate allocation scheme corresponding to each sensitivity level. High-sensitivity feature maps use finer quantization step sizes to reduce quantization distortion, while low-sensitivity feature maps use coarser quantization step sizes to compress data volume. For the first feature map within the same feature layer... The subsequent feature map is calculated to compare it with the first... The pixel differences at the same spatial location in each feature map are used to obtain a difference matrix; the difference matrix is then quantized and encoded using the following formula: in, Represents the spatial position of the difference matrix The quantization encoding value at that location; Represents a symbolic function; Indicates sensitivity level The corresponding quantization step size; This indicates the floor function; Since the feature maps are strongly correlated between adjacent channels, the energy of the difference matrix is highly concentrated near the zero value and exhibits a Laplace distribution characteristic. Therefore, the entropy coding method is used to perform lossless compression on the difference matrix to further reduce the amount of encoded data.
[0030] Step S43: Combine the encoded data of all feature maps according to a preset format, and add hierarchical labels and length information to generate difference encoded data; The preset format is arranged according to the feature layer order and the encoding order within each layer. The encoded data of each feature map is preceded by the layer number, channel number, sensitivity level label and data length field of the feature map, forming a structured difference encoded data stream.
[0031] Step S5: The on-board processing system extracts and encapsulates data packets from the downlink transmission queue and outputs them to the ground receiving station via the satellite-to-ground link to complete the on-board processing of remote sensing images. Specifically, the differentially encoded data is encapsulated using a satellite-based encapsulation format, with frame synchronization headers, frame length identifiers, source identifiers, priority tags, and frame check sequences added. Priority tags are set according to sensitivity levels. Encapsulated data frames are sorted by priority and encapsulation time, forming encapsulated and arranged data packets that are then sent to the downlink transmission queue. When the onboard processing system detects that the queue is not empty and link resources are available, it extracts the data packets, modulates them onto an RF carrier via a data transmitter, and radiates them to the ground via the satellite antenna. The ground receiving station captures the signal and demodulates it to recover the data packets. Synchronization and integrity verification are performed based on the synchronization headers and check information. Decoding and reconstruction are then performed based on priority and hierarchical tags to recover the remote sensing image. This process includes the following sub-steps: Step S51: Add frame synchronization header, frame length identifier, source identifier, priority label and frame check sequence to the differential encoded data according to the spaceborne data encapsulation format standard to complete the spaceborne encapsulation; the priority label is set according to the sensitivity level of the differential encoded data, with the highest priority label set for high sensitivity level data and the lowest priority label set for low sensitivity level data. The frame synchronization header uses a specific synchronization code bit sequence for ground receivers to achieve synchronization and positioning of data frames; the frame length identifier indicates the number of payload bytes in the entire data frame; the source identifier indicates the satellite number and sensor number from which the data originates; the frame check sequence is generated using a cyclic redundancy check algorithm and is used by the receiver to detect transmission errors; the priority label is strictly bound to the sensitivity level determined in step S41 to ensure that key feature data with high sensitivity levels receive the highest priority protection during transmission.
[0032] Step S52: The on-board processing system sorts the encapsulated data frames according to the priority tags. High-priority data frames are arranged at the front of the transmission queue, and low-priority data frames are arranged at the back. Data frames of the same priority are arranged in order of encapsulation time. The sorting operation is performed in the downlink transmission queue management module of the on-board processing system. When a new data frame enters the queue, its insertion position in the queue is found according to its priority label to ensure that the queue is always arranged in order from high to low priority. When there is an intermittent interruption or a sudden drop in bandwidth in the satellite-to-ground link, this priority orchestration mechanism can ensure that critical data with high sensitivity level is transmitted first, while data with low sensitivity level can be delayed or discarded, thereby maximizing the effective information transmission volume of the limited downlink.
[0033] Step S53: Combine the sorted encapsulated data frames into encapsulated and arranged data packets according to the queue order, and output them to the downlink transmission queue of the on-board processing system. Step S531: The on-board processing system monitors the status of the downlink transmission queue in real time. When the downlink transmission queue is not empty and the satellite-to-ground link transmission resources are available, it extracts and encapsulates the data packets from the head of the queue and transmits the data packets to the on-board data transmitter. The onboard processing system determines whether the link resources are available by using a handshake signal with the data transmitter. The handshake signal includes a transmitter ready flag and a carrier lock flag. When both flags are valid, the satellite-to-ground link transmission resources are determined to be available. At this time, the highest priority encapsulated and orchestrated data packet is extracted from the head of the downlink transmission queue and transmitted to the input buffer of the onboard data transmitter through the internal high-speed data bus, waiting for modulation and transmission.
[0034] Step S532: The onboard data transmitter modulates the packaged data packets onto the radio frequency carrier, radiates them into free space in the form of electromagnetic waves through the onboard antenna, and transmits them to the ground receiving station via the satellite-to-ground link. The modulation method adopts a phase shift keying modulation system suitable for satellite-to-ground link transmission. The radio frequency carrier frequency is set according to the satellite communication frequency band, and the transmission power is adaptively adjusted according to the link budget and the current space weather conditions. The modulated radio frequency signal is radiated into free space in the form of a directional electromagnetic beam through the onboard antenna, and then transmitted to the ground receiving station via the satellite-to-ground link to complete the physical layer transmission of onboard data to the ground.
[0035] Step S533: The antenna system of the ground receiving station captures the radio frequency signal, demodulates and recovers the encapsulated and arranged data packet, completes the synchronization and integrity verification of the data frame according to the synchronization header and check information in the encapsulation format, identifies the data type according to the priority label, completes the decoding and reconstruction according to the hierarchical label, and recovers the remote sensing image. The ground receiving station first acquires the satellite-to-ground link radio frequency signal through the antenna. After low-noise amplification and down-conversion processing, the baseband signal is obtained. Then, the baseband data stream is recovered through carrier synchronization and timing synchronization. After detecting the frame synchronization header, data frames are received. The frame length identifier is read to determine the data frame boundary. After reception, the check sequence is calculated and compared with the frame check sequence to complete the integrity verification of the data frame. For data frames that pass the verification, the corresponding decoding strategy is selected according to the hierarchical label. Entropy decoding, inverse difference operation and dequantization processing are performed in sequence to restore the reconstructed feature map. Finally, all reconstructed feature maps are input into the pre-configured feature decoder. The feature decoder adopts a deconvolution structure symmetrical with the on-board deep sparse coding network to upsample and fuse the feature maps layer by layer to restore the remote sensing image at the original resolution.
[0036] Example 2 like Figure 2 As shown, Embodiment 2 of this application provides a satellite-based remote sensing image processing system based on deep learning, comprising: Image Reception and Initiation Module 21: The on-board processing system receives remote sensing images and initiates a deep learning-based on-board remote sensing image processing method; it includes the following sub-modules: Data receiving and buffering submodule 211: It connects to the onboard optical sensor through a high-speed data interface, monitors the sensor data output status in real time, and transmits the acquired remote sensing image data to the data buffering unit of the onboard processing system through a high-speed serial interface after the onboard optical sensor completes the acquisition of a remote sensing image. Data verification and validation submodule 212: Reads complete remote sensing image data from the data cache unit, verifies the frame header identifier and frame tail check code of the image data, and confirms the integrity and validity of the data transmission; Initialization Submodule 213: After confirming that the data is complete and valid, the on-board processing system loads the remote sensing images into the processing memory area, initializes the processing parameters and processing buffer required by the deep learning-based on-board remote sensing image processing method, and begins to execute the on-board processing process. Feature Mapping and Resource Awareness Module 22: The on-board processing system inputs remote sensing images into a pre-trained deep sparse coding network, performs multi-layer sparse feature mapping, generates a multi-layer sparse feature mapping image set, and calculates the on-board comprehensive resource occupancy index of the image set; including the following sub-modules: Multi-layer sparse feature mapping submodule 221: The feature mapping unit of the on-board processing system inputs remote sensing images into a pre-trained deep sparse coding structure, and extracts multi-scale sparse features of the remote sensing images through multi-layer convolution and sparse activation operations to generate a multi-layer sparse feature mapping image set containing a preset number of feature layers. Resource occupancy real-time monitoring and index calculation submodule 222: During the execution of multi-layer sparse feature mapping, the resource monitoring unit of the on-board processing system collects the usage of computing resources, storage resources and power consumption resources in real time, and calculates the on-board comprehensive resource occupancy index according to the preset computing resource weight, storage resource weight and power consumption resource weight. Resource occupancy level determination submodule 223: compares the calculated on-board comprehensive resource occupancy index with preset high occupancy threshold and low occupancy threshold to determine the current on-board resource occupancy level, which includes high occupancy level, medium occupancy level and low occupancy level; Feature distillation and reconstruction module 23: Based on the on-board comprehensive resource occupancy index, the on-board processing system performs adaptive feature distillation on the multi-layer sparse feature mapping image set to generate a distilled feature image set. Then, it performs resource-aware sparse reconstruction on the distilled feature image set to generate a reconstructed feature image set and calculates the downlink distortion sensitivity of the reconstructed feature image set. This includes the following sub-modules: Adaptive Feature Distillation Submodule 231: Based on the determined resource occupancy level, selects a strategy with corresponding strength from a variety of preset feature distillation strategies, and discards or merges each feature channel in the multi-layer sparse feature mapping image set according to its importance ranking to generate a distillation feature image set. Resource-aware sparse reconstruction submodule 232: The resource-aware sparse reconstruction unit of the on-board processing system uses the distillation feature image set as the observation value and a pre-trained sparse dictionary as the sparse basis to solve the reconstructed feature image set through an iterative reconstruction algorithm. Distortion sensitivity assessment submodule 233: For each feature map in the reconstructed feature image set, calculate its mean spatial gradient magnitude and frequency domain energy concentration. Combine the two indicators of mean spatial gradient magnitude and frequency domain energy concentration to calculate the distortion sensitivity of each feature map. Then, take the weighted average of the distortion sensitivity of all feature maps to obtain the downlink distortion sensitivity of the entire reconstructed feature image set. Hierarchical coding and encapsulation module 24: Based on downlink distortion sensitivity, the on-board processing system performs hierarchical differential coding on the reconstructed feature image set to generate differential coded data. Then, it performs on-board encapsulation and priority orchestration on the differential coded data to generate encapsulated and orchestrated data packets, and outputs these packets to the downlink transmission queue. This includes the following sub-modules: Hierarchical coding strategy determination submodule 241: Based on the downlink distortion sensitivity, the on-board processing system determines the number of hierarchical differential coding levels and the code rate allocation scheme for each level, divides the feature maps in the reconstructed feature image set into multiple sensitivity levels according to the preset distortion sensitivity threshold, and assigns differentiated coding rates and error correction coding strengths to different sensitivity levels. Differential encoding execution submodule 242: Based on the hierarchical results, differential encoding is performed on each feature map in the reconstructed feature image set. For the first feature map in each feature layer, quantization encoding is performed directly. For subsequent feature maps in the same feature layer, the pixel difference between the current feature map and the previous feature map at the same spatial position is calculated, and entropy encoding is performed on the difference matrix. Encoding data combination submodule 243: Combines the encoding data of all feature maps according to a preset format, and adds hierarchical labels and length information to generate difference encoding data; Downlink Output Module 25: The on-board processing system extracts encapsulated and orchestrated data packets from the downlink transmission queue and outputs them to the ground receiving station via the satellite-to-ground link to complete the on-board processing of remote sensing images; it includes the following sub-modules: Spaceborne data encapsulation submodule 251: Adds frame synchronization header, frame length identifier, source identifier, priority label and frame check sequence to the differential encoded data according to the spaceborne data encapsulation format standard to complete the spaceborne encapsulation; the priority label is set according to the sensitivity level corresponding to the differential encoded data, with high-sensitivity data set with the highest priority label and low-sensitivity data set with the lowest priority label. Priority scheduling and sorting submodule 252: The on-board processing system sorts the encapsulated data frames according to the priority tags. High-priority data frames are arranged at the front of the transmission queue, and low-priority data frames are arranged at the back. Data frames of the same priority are arranged in order of encapsulation time. Queue Management and Output Submodule 253: Combines the sorted encapsulated data frames into encapsulated and arranged data packets according to the queue order, and outputs them to the downlink transmission queue of the on-board processing system.
Claims
1. A satellite-based remote sensing image processing method based on deep learning, characterized in that, include: S1. The on-board processing system receives remote sensing images and initiates a deep learning-based on-board remote sensing image processing method. S2. The on-board processing system inputs remote sensing images into a pre-trained deep sparse coding network, performs multi-layer sparse feature mapping, generates a multi-layer sparse feature mapping image set, and calculates the on-board comprehensive resource occupancy index of the image set. S3. Based on the on-board comprehensive resource occupancy index, the on-board processing system performs adaptive feature distillation on the multi-layer sparse feature mapping image set to generate a distilled feature image set. Then, it performs resource-aware sparse reconstruction on the distilled feature image set to generate a reconstructed feature image set and calculates the downlink distortion sensitivity of the reconstructed feature image set. S4. Based on the downlink distortion sensitivity, the on-board processing system performs hierarchical differential encoding on the reconstructed feature image set to generate differential encoded data. Then, it performs on-board encapsulation and priority arrangement on the differential encoded data to generate encapsulated and arranged data packets, and outputs the encapsulated and arranged data packets to the downlink transmission queue. S5, the on-board processing system extracts encapsulated and arranged data packets from the downlink transmission queue and outputs them to the ground receiving station via the satellite-to-ground link to complete the on-board processing of remote sensing images.
2. The on-board processing method for remote sensing images based on deep learning according to claim 1, characterized in that, The on-board processing system receives remote sensing images and initiates a deep learning-based on-board remote sensing image processing method, including the following sub-steps: The satellite connects to the onboard optical sensor via a high-speed data interface to monitor the sensor's data output status in real time. Once the onboard optical sensor has completed the acquisition of a remote sensing image, the acquired remote sensing image data is transmitted to the data buffer unit of the onboard processing system via a high-speed serial interface. Complete remote sensing image data is read from the data buffer unit, and the frame header identifier and frame tail check code of the image data are verified to confirm the integrity and validity of the data transmission. After confirming that the data is complete and valid, the on-board processing system loads the remote sensing images into the processing memory area, initializes the processing parameters and processing buffer required for the deep learning-based on-board remote sensing image processing method, and begins to execute the on-board processing flow.
3. The on-board processing method for remote sensing images based on deep learning according to claim 1, characterized in that, The on-board processing system inputs remote sensing images into a pre-trained deep sparse coding network, performs multi-layer sparse feature mapping, generates a multi-layer sparse feature mapping image set, and calculates the on-board comprehensive resource occupancy index of the image set, including the following sub-steps: The feature mapping unit of the on-board processing system inputs remote sensing images into a pre-trained deep sparse coding structure. Through multi-layer convolution and sparse activation operations, it extracts multi-scale sparse features from the remote sensing images and generates a multi-layer sparse feature mapping image set containing a preset number of feature layers. During the execution of multi-layer sparse feature mapping, the resource monitoring unit of the on-board processing system collects the usage of computing resources, storage resources, and power consumption resources in real time, and calculates the on-board comprehensive resource occupancy index based on the preset computing resource weights, storage resource weights, and power consumption resource weights. The calculated on-board comprehensive resource occupancy index is compared with preset high occupancy thresholds and low occupancy thresholds to determine the current on-board resource occupancy level, which includes high occupancy level, medium occupancy level and low occupancy level.
4. The on-board processing method for remote sensing images based on deep learning according to claim 1, characterized in that, Based on the onboard integrated resource occupancy index, the onboard processing system performs adaptive feature distillation on the multi-layer sparse feature mapping image set to generate a distilled feature image set. Then, it performs resource-aware sparse reconstruction on the distilled feature image set to generate a reconstructed feature image set and calculates the downlink distortion sensitivity of the reconstructed feature image set. This process includes the following sub-steps: Based on the determined resource occupancy level, a strategy with corresponding strength is selected from a variety of preset feature distillation strategies. The feature channels in the multi-layer sparse feature mapping image set are discarded or merged according to their importance ranking to generate a distillation feature image set. The resource-aware sparse reconstruction unit of the on-board processing system uses the distillation feature image set as the observation value and a pre-trained sparse dictionary as the sparse basis to solve the reconstruction feature image set through an iterative reconstruction algorithm. For each feature map in the reconstructed feature image set, calculate its mean spatial gradient magnitude and frequency domain energy concentration. Combine the two indicators of mean spatial gradient magnitude and frequency domain energy concentration to calculate the distortion sensitivity of each feature map. Then, take the weighted average of the distortion sensitivity of all feature maps to obtain the downlink distortion sensitivity of the entire reconstructed feature image set.
5. The on-board processing method for remote sensing images based on deep learning according to claim 1, characterized in that, Based on downlink distortion sensitivity, the onboard processing system performs hierarchical differential encoding on the reconstructed feature image set to generate differential encoded data. Then, it performs onboard encapsulation and priority orchestration on the differential encoded data to generate encapsulated and orchestrated data packets, and outputs these packets to the downlink transmission queue. This process includes the following sub-steps: Based on the downlink distortion sensitivity, the on-board processing system determines the number of levels of hierarchical differential coding and the code rate allocation scheme for each level. The feature maps in the reconstructed feature image set are divided into multiple sensitivity levels according to the preset distortion sensitivity threshold, and different coding rates and error correction coding strengths are assigned to different sensitivity levels. Based on the classification results, difference encoding is performed on each feature map in the reconstructed feature image set. For the first feature map in each feature layer, quantization encoding is performed directly. For subsequent feature maps in the same feature layer, the pixel difference between the current feature map and the previous feature map at the same spatial position is calculated, and entropy encoding is performed on the difference matrix. Combine the encoded data of all feature maps according to a preset format, and add hierarchical labels and length information to generate difference encoded data.
6. The on-board processing method for remote sensing images based on deep learning according to claim 1, characterized in that, The on-board processing system extracts encapsulated and orchestrated data packets from the downlink transmission queue and outputs them to the ground receiving station via the satellite-to-ground link to complete the on-board processing of remote sensing images, including the following sub-steps: Add frame synchronization header, frame length identifier, source identifier, priority label and frame check sequence to the differential encoded data according to the spaceborne data encapsulation format standard to complete the spaceborne encapsulation; The priority label is set according to the sensitivity level corresponding to the difference-coded data. High-sensitivity data is set with the highest priority label, and low-sensitivity data is set with the lowest priority label. The on-board processing system sorts the encapsulated data frames according to priority tags. High-priority data frames are arranged at the front of the transmission queue, and low-priority data frames are arranged at the back. Data frames of the same priority are arranged in order of encapsulation time. The sorted encapsulated data frames are combined into encapsulated and arranged data packets according to the queue order and output to the downlink transmission queue of the on-board processing system.
7. The on-board processing method for remote sensing images based on deep learning according to claim 6, characterized in that, The sorted encapsulated data frames are combined into encapsulated and orchestrated data packets according to the queue order and output to the downlink transmission queue of the on-board processing system, including the following sub-steps: The onboard processing system monitors the status of the downlink transmission queue in real time. When the downlink transmission queue is not empty and the satellite-to-ground link transmission resources are available, it extracts, encapsulates, and arranges data packets from the head of the queue and transmits the data packets to the onboard data transmitter. The onboard data transmitter modulates the packaged and arranged data packets onto the radio frequency carrier, which is then radiated into free space in the form of electromagnetic waves through the onboard antenna and transmitted to the ground receiving station via the space-to-ground link. The antenna system of the ground receiving station captures radio frequency signals, demodulates and recovers the encapsulated and arranged data packets, completes the synchronization and integrity verification of data frames according to the synchronization header and check information in the encapsulation format, identifies the data type according to the priority label, and completes decoding and reconstruction according to the hierarchical label to recover the remote sensing image.
8. A satellite-based remote sensing image processing system based on deep learning, characterized in that, include: Image receiving and initiation module: The on-board processing system receives remote sensing images and initiates a deep learning-based on-board remote sensing image processing method; Feature mapping and resource awareness module: The on-board processing system inputs remote sensing images into a pre-trained deep sparse coding network, performs multi-layer sparse feature mapping, generates a multi-layer sparse feature mapping image set, and calculates the on-board comprehensive resource occupancy index of the image set. Feature distillation and reconstruction module: Based on the on-board comprehensive resource occupancy index, the on-board processing system performs adaptive feature distillation on the multi-layer sparse feature mapping image set to generate a distilled feature image set, and then performs resource-aware sparse reconstruction on the distilled feature image set to generate a reconstructed feature image set, and calculates the downlink distortion sensitivity of the reconstructed feature image set. Hierarchical coding and encapsulation module: Based on the downlink distortion sensitivity, the on-board processing system performs hierarchical differential coding on the reconstructed feature image set to generate differential coded data, then performs on-board encapsulation and priority orchestration on the differential coded data to generate encapsulated and orchestrated data packets, and outputs the encapsulated and orchestrated data packets to the downlink transmission queue. Downlink output module: The on-board processing system extracts encapsulated and arranged data packets from the downlink transmission queue and outputs them to the ground receiving station via the satellite-to-ground link to complete the on-board processing of remote sensing images.
