Satellite edge hyperspectral image processing method and system

Abstract

The invention discloses a satellite edge hyperspectral image processing method and system, relates to the technical field of communication, and aims to solve the problems of resource limitation, communication bottleneck and unreasonable task scheduling faced by hyperspectral image processing in satellite edge calculation. The method comprises the following steps: taking a future data transmission rate of a satellite as a communication feature vector, and constructing a task unloading matrix in combination with a graph attention mechanism to realize accurate matching of tasks and satellite resources; the task unloading matrix is converted into task features, the task features and the hyperspectral image are fused to form a fusion tensor, a high-dimensional feature vector is generated based on the fusion tensor, key wavebands are screened, and a waveband selection matrix is obtained; and constructing a joint loss function containing a task unloading loss function and a band selection loss function, and iteratively optimizing the double matrixes until the loss is minimum to obtain an optimal matrix. Through collaborative optimization of task scheduling and image processing, system energy consumption, communication constraint and processing precision are balanced, the satellite resource utilization rate and image processing real-time performance are improved, and the method is suitable for low-orbit satellite edge calculation scenes.

Description

Technical Field

[0001] This invention relates to the field of communication technology, specifically to a satellite edge hyperspectral image processing method and system. Background Technology

[0002] With the continuous increase in the number of low-Earth orbit satellites, their data acquisition capabilities in fields such as communication, remote sensing, and environmental monitoring have been significantly enhanced. In particular, hyperspectral imaging (HSI) missions with high spatial and spectral resolution provide rich information support for applications such as agricultural monitoring, geological surveys, and ecological assessments. However, hyperspectral data is massive in volume, and traditional satellites rely on limited space-to-ground communication windows and transmission bandwidth to transmit raw images to ground centers for processing. This results in severe transmission bottlenecks and latency issues, limiting the real-time performance of missions and the efficiency of system response.

[0003] To alleviate these problems, Satellite Edge Computing (SEC) emerged. By deploying lightweight computing modules on satellites, SEC enables local offloading and processing of tasks, effectively improving the real-time performance and processing efficiency of image analysis. However, limited by onboard computing resources, power supply, and the complex orbital environment, SEC still faces challenges such as uneven task allocation, high energy consumption, and low resource utilization when dealing with large-scale, multi-task hyperspectral image processing.

[0004] At the same time, the complexity of inter-satellite resource coordination and mission scheduling is constantly increasing, and there is an urgent need to design reasonable mission offloading strategies and band selection mechanisms to make full use of the multi-satellite collaborative processing capabilities, optimize the allocation of computing and communication resources, and achieve the optimal balance between mission execution efficiency and system energy consumption.

[0005] However, existing research mostly focuses on optimizing single modules, lacking systematic research on the collaborative optimization of multiple modules such as "task offloading and band selection" in hyperspectral image processing tasks. This makes it difficult to meet the resource coordination and performance trade-off requirements in multi-satellite collaborative processing scenarios. Therefore, constructing a joint optimization method that integrates task allocation strategies and band selection mechanisms has become crucial for improving the performance of satellite edge computing systems. Summary of the Invention

[0006] To address the problems of resource constraints, communication bottlenecks, and unreasonable task scheduling faced by hyperspectral image processing in satellite edge computing in existing technologies, this invention provides a satellite edge hyperspectral image processing method and system. This method is designed for hyperspectral remote sensing missions in low-Earth orbit satellite networks and constructs a multi-objective scheduling framework that combines task offloading and band selection optimization to improve the overall system efficiency and energy utilization.

[0007] This invention is achieved through the following technical solution: A method for processing hyperspectral images of satellite edges includes the following steps: Step 1: Assign the data transmission rate of the satellite in the future time as a communication feature vector to the satellite node, determine the attention weight of each task node and all satellite nodes, construct the offloading probability distribution vector of each task based on the attention weight, and construct the task offloading matrix based on all the offloading probability distribution vectors. Step 2: Convert the task unloading matrix into task features, fuse the hyperspectral image with the task features to form a fusion tensor, generate a high-dimensional feature vector for each band region of the hyperspectral image based on the fusion tensor, and perform band selection of the hyperspectral image based on the high-dimensional feature vector to obtain the band selection matrix. Step 3: Construct a joint loss function based on the task unloading loss function and the band selection loss function. Use the joint loss function to iteratively optimize the task unloading matrix and the band selection matrix until the joint loss function is minimized, thus obtaining the optimal task unloading matrix and band selection matrix.

[0008] Preferably, the method for determining the data transmission rate of the satellite in future times is as follows: Based on the historical communication channel characteristics of the satellite, a historical communication channel sequence is constructed. A pre-trained time-series network model is used to predict the data transmission rate of the satellite in the future.

[0009] Preferably, the step of assigning the future data transmission rate of the satellite as a communication feature vector to the satellite node and determining the attention weight of each task node with all satellite nodes includes: A graph is constructed based on task nodes and satellite nodes. Data transmission rate is assigned as a communication feature vector to satellite nodes. The attention weight of each task with all satellites is determined based on a graph attention mechanism.

[0010] Preferably, the step of constructing the unloading probability distribution vector for each task based on the attention weights, and constructing the task unloading matrix based on all the unloading probability distribution vectors, includes: The attention weights of each task are converted into an unloading probability distribution vector. The attention weights of each task node and all satellite nodes are regarded as the unloading intention of the task for each candidate satellite, thus forming an unloading probability distribution vector. The maximum value of the unloading probability distribution vector is determined by the maximum confidence selection rule. The satellite corresponding to the maximum value is taken as the target unloading node of the mission. The mission unloading matrix is ​​obtained by discretizing all the target unloading nodes.

[0011] Preferably, the step of converting the task unloading matrix into task features and fusing the hyperspectral image with the task features to form a fusion tensor includes: The task unloading matrix is ​​converted into task features that match the spatial size of the hyperspectral image. Then, through channel-dimensional stitching operations, the hyperspectral image and task features are fused to form a fusion tensor.

[0012] Preferably, the step of generating a high-dimensional feature vector for each band region of the hyperspectral image based on the fusion tensor, and performing band selection of the hyperspectral image based on the high-dimensional feature vector to obtain a band selection matrix includes: A preliminary feature map of the fused tensor is obtained, and channel attention and spatial attention are used to enhance the band channels and target spatial regions of the preliminary feature map to obtain a multi-dimensional attention feature map. Based on the magnitude of the high-dimensional feature vector and a set threshold, band selection is performed on the hyperspectral image, and a band selection matrix is ​​constructed based on the selected bands.

[0013] Preferably, the multidimensional attention feature map is obtained as follows: The initial feature map is input into the convolutional block attention mechanism module. Through the concatenation mechanism of channel attention and spatial attention, the band channels and target spatial regions of the multidimensional attention feature map are enhanced respectively to obtain the multidimensional attention feature map.

[0014] Preferably, the band selection of the hyperspectral image based on the modulus of the high-dimensional feature vector and a set threshold includes: The multidimensional attention feature map is input into the capsule network. The capsule network generates a high-dimensional feature vector for each band region of the hyperspectral image through convolutional capsule layers. The magnitude of the high-dimensional feature vector is compared with a threshold. Based on the comparison result, the bands of the hyperspectral image are selected, and a band selection matrix is ​​constructed based on the selected bands.

[0015] Preferably, the expression for the joint loss function is as follows:

[0016] in, As a balance factor, Unload the loss function for the task; Choose a loss function for the band; The task unloading loss function is as follows:

[0017] in, The total energy consumption after the mission is allocated to the satellite. For the first z Penalty function for each constraint term; These are weighting coefficients; Z The number of constraint terms; The band selection loss function as follows: . in, For the true label of the task, The classification confidence level of this task in the current band combination; Indicates whether the l-th band is selected; To select the maximum number of tapes, This is the regularization coefficient used to control the number of bands.

[0018] A satellite edge hyperspectral image processing system, comprising: The task offloading module is used to assign the satellite's future data transmission rate as a communication feature vector to the satellite node, determine the attention weight of each task node and all satellite nodes, construct the offloading probability distribution vector of each task based on the attention weight, and construct the task offloading matrix based on all the offloading probability distribution vectors. The band selection module is used to convert the task unloading matrix into task features, fuse the hyperspectral image with the task features to form a fusion tensor, generate a high-dimensional feature vector for each band region of the hyperspectral image based on the fusion tensor, and perform band selection of the hyperspectral image based on the high-dimensional feature vector to obtain the band selection matrix. The iterative module is used to construct a joint loss function based on the task unloading loss function and the band selection loss function. The joint loss function is then used to iteratively optimize the task unloading matrix and the band selection matrix until the joint loss function is minimized, thus obtaining the optimal task unloading matrix and band selection matrix.

[0019] Compared with the prior art, the present invention has the following beneficial technical effects: This application provides a satellite edge hyperspectral image processing method. The scheme first uses the future data transmission rate of the satellite as a communication feature, and combines it with a graph attention mechanism to generate a task offload matrix. This accurately matches tasks with satellite resources, avoiding node overload or resource idleness caused by uneven allocation in traditional methods. Then, through feature fusion of the task offload matrix and hyperspectral images, attention enhancement, and high-dimensional vector band selection, redundant bands are eliminated, alleviating the bandwidth bottleneck of satellite-to-ground communication while ensuring that key image information is not lost. Finally, a joint loss function is used to achieve iterative optimization of the dual-matrix, balancing task energy consumption and image processing accuracy, avoiding the limitations of single-module optimization. The overall scheme is suitable for real-world scenarios with limited low-Earth orbit satellite resources and time-varying channels. It effectively balances system energy consumption, communication constraints, and processing efficiency, significantly improving the real-time performance, resource utilization, and accuracy stability of satellite edge hyperspectral image processing, and solving the core problem of insufficient multi-module collaboration in existing technologies.

[0020] This application also proposes a satellite edge hyperspectral image processing system, an electronic device, and a computer storage medium, which possess all the advantages of the aforementioned satellite edge hyperspectral image processing methods. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is an application scenario diagram of the resource optimization method for satellite edge computing hyperspectral image processing provided in the embodiments of the present invention; Figure 2 This is a flowchart of the satellite edge hyperspectral image processing method provided in the embodiments of the present invention; Figure 3 This is a simulation diagram demonstrating the effectiveness of the JFIO algorithm provided in this embodiment of the invention; Figure 4 This is a schematic diagram illustrating the hyperspectral image classification accuracy provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the average energy consumption of hyperspectral image processing provided in the embodiments of the present invention; Figure 6 This is a schematic diagram illustrating the time consumption of hyperspectral image processing provided in an embodiment of the present invention; Figure 7 This is a comparative schematic diagram of the changes in hyperspectral image processing utility under different weight parameters provided in the embodiments of the present invention; Figure 8 This is a schematic diagram comparing the average utility of hyperspectral image processing under different memory buffer configurations provided in the embodiments of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0024] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0025] A method for processing hyperspectral images of satellite edges includes the following steps: Step 1: Use the data transmission rate of the satellite in the future as the communication feature vector of the satellite node. Construct a graph based on the task nodes and satellite nodes, determine the attention weight of each task node with all satellite nodes, construct the offloading probability distribution vector of each task based on the attention weight, and construct the task offloading matrix based on all the offloading probability distribution vectors. The task offloading matrix is ​​the core carrier of the allocation relationship between satellites and tasks. Its core function is to clarify which satellite will perform each hyperspectral image processing task, achieving precise task-satellite matching. This effectively balances the computational load of multiple satellites, avoiding situations where some satellites are overloaded and others have idle resources, reducing processing delays and energy waste caused by unreasonable task allocation, and improving the overall system's computational resource utilization.

[0026] This step uses the satellite's future data transmission rate as the core communication characteristic, endowing satellite nodes with dynamic and realistic communication capabilities. Combined with features such as task data volume and computational load of task nodes, a graph structure is constructed, enabling accurate modeling of resource coupling relationships between multiple tasks and multiple satellites. Calculating attention weights based on a graph attention mechanism quantifies the scheduling adaptability of each task with different satellites, avoiding blind task allocation. Transforming the weights into an offload probability distribution vector and constructing a task offload matrix ensures that each task is assigned to a unique satellite with the highest resource matching degree, effectively solving the problem of uneven task allocation in traditional methods, improving satellite resource utilization, and providing a suitable task scheduling basis for subsequent band selection, reducing energy waste in subsequent processing caused by task-satellite mismatch.

[0027] Step 2: Convert the task unloading matrix into task features, fuse the hyperspectral image with the task features to form a fusion tensor, generate a high-dimensional feature vector for each band region of the hyperspectral image based on the fusion tensor, and perform band selection of the hyperspectral image based on the high-dimensional feature vector to obtain the band selection matrix. The band selection matrix is ​​used to filter out key bands with high discriminative power for image processing (such as classification tasks) from hundreds of consecutive bands with highly overlapping information in hyperspectral images, while eliminating redundant bands. This significantly reduces the amount of data to be transmitted and processed while ensuring image processing accuracy, effectively alleviating the bottleneck of limited satellite-to-ground communication bandwidth, reducing data transmission energy consumption and satellite computing burden, and preventing redundant bands from occupying excessive communication and computing resources.

[0028] This step transforms the task offloading matrix into task features that match the spatial dimensions of the hyperspectral image, breaking down the information barrier between task scheduling and image processing. This allows task scheduling results to directly guide image band selection. By stitching together the channel dimensions, the hyperspectral image and task features are fused. The resulting fusion tensor contains both image information and task adaptation information, providing comprehensive data support for accurate selection of key bands. Based on the fusion tensor, high-dimensional feature vectors are generated and band selection is performed, automatically filtering out redundant bands. While ensuring no loss of key image information, this significantly reduces the amount of data to be transmitted and processed, alleviating the bandwidth bottleneck of satellite-to-ground communication and reducing data transmission energy consumption and computational burden.

[0029] Step 3: Construct a joint loss function based on the task unloading loss function and the band selection loss function. Use the joint loss function to iteratively optimize the task unloading matrix and the band selection matrix until the joint loss function is minimized, thus obtaining the optimal task unloading matrix and band selection matrix.

[0030] The constructed joint loss function, by adjusting the weights of task offloading loss and band selection loss through a balancing factor, can simultaneously consider energy consumption optimization in task scheduling and classification accuracy assurance in band selection, avoiding overall system performance imbalance caused by optimizing a single module. Iterative optimization of the task offloading matrix and band selection matrix using this joint loss function enables end-to-end collaborative optimization, allowing the task scheduling scheme and band selection results to adapt and dynamically adjust to each other, ultimately yielding the optimal matrix. This process effectively balances system energy consumption, communication constraints, and image processing accuracy, solving the problem that traditional single-module optimization cannot meet the needs of multi-satellite collaborative processing. It significantly improves the overall efficiency and resource utilization of the satellite edge hyperspectral image processing system, ensuring stable and efficient system operation under complex orbital environments and resource-constrained conditions.

[0031] In some embodiments, the method for determining the data transmission rate of a satellite in the future is as follows: Based on the historical communication channel characteristics of the satellite, a historical communication channel sequence is constructed. A pre-trained time-series network model is used to predict the data transmission rate of the satellite in the future.

[0032] Optionally, the temporal network model can be a long short-term memory network model, a gated recurrent unit model, or a temporal convolutional network model.

[0033] In some embodiments, the method for constructing the band selection matrix is ​​as follows: First, a graph is constructed based on task nodes and satellite nodes. The data transmission rate is assigned as the communication feature vector of the satellite node, and the attention weight between each task and all satellites is determined based on the graph attention mechanism to measure the scheduling adaptability between task nodes and satellite nodes. Then, the attention weights of each task are converted into an unloading probability distribution vector. In order to convert the attention weights into unloading decisions, the attention weights of each task node and all satellite nodes are regarded as the unloading intention of the task for each candidate satellite, forming an unloading probability distribution vector: an unloading probability distribution vector is generated for each task.

[0034] The maximum value of the unloading probability distribution vector is determined by the maximum confidence selection rule. The satellite corresponding to the maximum value is taken as the target unloading node of the mission. The mission unloading matrix is ​​obtained by discretizing all the target unloading nodes.

[0035] A graph G containing task nodes and satellite nodes is constructed, and the future data transmission rate of the satellites predicted by LSTM is assigned to the satellite nodes as the core communication feature, while also considering the data volume, computational load, and other characteristics of the task nodes. Based on a graph attention mechanism, the attention weights of each task and all satellites are determined. This quantifies the scheduling adaptability of different tasks and satellites in terms of communication capabilities and resource status, avoiding the adaptation bias caused by relying on experience or single indicators in traditional task allocation. The weight results are made more consistent with the actual scenario of "resource-constrained and time-varying channels" in satellite edge computing, providing a scientific quantitative basis for subsequent offloading decisions and ensuring that tasks can initially be matched with satellites with better communication and resource conditions.

[0036] Transforming attention weights into an offloading probability distribution vector essentially converts the quantification of mission-satellite compatibility into a probabilistic expression with "preference bias," clearly presenting the strength of an individual mission's offloading intention for different satellites and making the candidate allocation schemes for missions more intuitive. Furthermore, using the maximum confidence selection rule to determine the target offloading node and discretizing it to obtain the mission offloading matrix transforms probabilistic biases into explicit "0 / 1" allocation results, strictly adhering to the constraint of "each mission being assigned to only one satellite," fundamentally avoiding the problems of duplicate mission allocation or lack of processing, and ensuring the determinism and executability of mission allocation. This process, through the combination of probabilistic transformation and discrete decision-making, retains the compatibility information reflected by the attention weights while forming a feasible mission scheduling scheme, providing a suitable mission execution platform for subsequent band selection and reducing the energy consumption and latency waste in subsequent processing caused by mission-satellite mismatch.

[0037] In some embodiments, the method for constructing the band selection matrix is ​​as follows: The task unloading matrix is ​​converted into task features that match the spatial dimensions of the hyperspectral image. The hyperspectral image and task features are concatenated along the channel dimension to obtain a fusion tensor. A preliminary feature map of the fusion tensor is obtained, and channel attention and spatial attention are applied to enhance the band channels and target spatial regions of the preliminary feature map to obtain a multidimensional attention feature map. A high-dimensional feature vector is generated for each band region of the hyperspectral image based on the multidimensional attention feature map. Band selection of the hyperspectral image is performed based on the high-dimensional feature vector to obtain a band selection matrix.

[0038] This method for constructing a band selection matrix first converts the task offloading matrix into task features that match the spatial dimensions of the hyperspectral image, breaking down the information barriers between task scheduling and image processing and ensuring that satellite resource adaptability information (such as transmission capacity and computational load) can be incorporated into band selection. Then, a fusion tensor is formed by concatenating the channel dimensions, allowing the input to simultaneously include image spectral details and resource features of the task execution carrier, preventing band selection from being divorced from actual execution conditions. Subsequently, a preliminary feature map is extracted and enhanced by channel attention (strengthening high-value band channels) and spatial attention (locating key target regions) to generate a multi-dimensional attention feature map, focusing on core information and suppressing redundant interference. Next, a capsule network is used to generate a high-dimensional feature vector for each band region, quantifying the "activation degree" of the band for the classification task with the vector magnitude, thus more comprehensively characterizing the value of the band. Finally, bands are selected based on the magnitude and threshold to form a discrete band selection matrix. This method ensures that the selected bands are compatible with satellite resources by incorporating mission information, thus avoiding transmission bottlenecks and energy waste. It also ensures that key bands are not overlooked by using attention enhancement and high-dimensional vector evaluation, thus solving the problem that traditional band selection relies solely on image information and is disconnected from mission execution.

[0039] Example 1 Please see Figure 1 , Figure 1 This is an application scenario diagram of the resource optimization method for satellite edge computing hyperspectral image processing provided in this embodiment of the invention. In the application scenario, there is a local satellite with hyperspectral imaging capabilities and several nearby low-Earth orbit satellites, forming a satellite ensemble. , among which number Indicates local satellite.

[0040] Each satellite carries an onboard edge computing module and communication system for band selection and image data transmission. The local satellite acquires image data via a hyperspectral imager and represents each hyperspectral image as a task set. .

[0041] The pre-trained band selection model on the ground is deployed in the onboard computing system for performing hyperspectral image processing tasks in orbit. The local satellite dynamically formulates a task offloading strategy based on image content characteristics and the current satellite computing resource status, allocating some raw images to neighboring satellites for collaborative processing to improve overall processing efficiency and achieve computational load balancing. During the satellite-to-ground visibility window, the satellite transmits the selected bands to the ground receiving station via downlink, where the ground service center continues subsequent analysis and processing.

[0042] The Joint Feature Iterative Optimization (JFIO) method employed in this embodiment boasts scalability and module collaboration capabilities, making it suitable for efficient processing of hyperspectral imagery tasks in resource-constrained low-Earth orbit satellite edge computing environments. Consider a method consisting of... M A network collaborative processing scenario consisting of 10 satellites, in which local satellites carry hyperspectral imagers to acquire image sets and can coordinate with neighboring satellites to offload processing tasks.

[0043] Once the local satellite acquires hyperspectral images, a joint optimization strategy can be employed to make decisions on task offloading and band selection based on the current computing resource status and image content features. The JFIO method proposed in this invention unifies the modeling of task allocation, band selection, and resource optimization processes. Through feature fusion mechanisms and multi-module collaborative feedback, it jointly optimizes the offloading strategy and classification performance, thereby effectively balancing computational load and system energy consumption. During execution, the joint model can dynamically adjust the offloading target based on image features and satellite resources, selecting the satellite with the optimal resource matching to execute the sub-task.

[0044] Furthermore, the proposed model possesses end-to-end deployability. The lightweight feature classification model pre-trained on the ground can be deployed on satellite edge computing platforms, exhibiting characteristics of fast inference and low resource consumption. Simultaneously, the system utilizes a joint loss function to dynamically balance classification accuracy and resource overhead, achieving optimal overall performance for task processing. After processing is complete, each satellite can download selected band data to the ground receiving center during the satellite-to-ground communication window.

[0045] In view of this, this application provides a satellite edge hyperspectral image processing method that improves the system's mission response capability and resource utilization efficiency through joint optimization of task offloading and band selection. This method adapts to the needs of dynamic collaborative processing of multiple satellites, integrates a lightweight classification model and a feature feedback mechanism, effectively reduces energy consumption and latency, and ensures the quality of hyperspectral image processing.

[0046] Please see Figure 2 , Figure 2 This is a flowchart of the satellite edge hyperspectral image processing method provided in this embodiment. The satellite edge hyperspectral image processing method provided in this embodiment includes the following steps: Step 1: During the system initialization phase, set the maximum number of iterations. Initialize joint scheduling variables, including the task unloading matrix. Band selection matrix Joint model parameters and iterate index t Initialize to 0. Collect and preprocess the set of tasks to be processed and the set of available satellites, as well as the basic characteristic information of each task and each satellite.

[0047] Step 2: Predict the satellite data transmission rate based on the satellite's historical communication channel sequence.

[0048] During the satellite link prediction phase, the time-varying characteristics of low-Earth orbit (LEO) satellite communication links—that is, the strong dynamic changes in the link between LEO satellites and the ground—are considered. Based on a Long Short-Term Memory (LSTM) network, the historical communication channel sequence of each satellite is used to predict the future data transmission rate. ; Let the first m The historical channel state sequence of the satellite is Each state includes features such as signal-to-noise ratio, azimuth, and elevation. An LSTM model is used to perform time-series modeling of this historical channel state sequence to predict the data transmission rate at future times. The calculation formula is as follows:

[0049] in, For satellite m Communication bandwidth, Its transmission power, For gain, For the predicted satellite m With ground station g Channel gain, Indicates noise power.

[0050] Step 3: Use the data transmission rate as the communication feature vector of the satellite node, construct a graph based on the task nodes and satellite nodes, determine the attention weight of each task node and all satellite nodes based on the attention mechanism, construct the offloading probability distribution vector of each task based on the attention weight, and construct the task offloading matrix based on all the offloading probability distribution vectors.

[0051] The task offloading matrix is ​​used to assign each task to a single target satellite for execution, i.e., the task offloading scheduling phase. In this embodiment, a Graph Attention Network (GAT) is constructed to model the resource coupling relationship between multiple tasks and multiple satellites, thereby determining the task offloading matrix. The task offloading scheduling method is described in detail below: S3.1 Constructing a graph based on task nodes and satellite nodes The data transmission rate is assigned to the satellite node as a communication characteristic vector. picture , where the set of nodes Task Node Satellite nodes To avoid confusion in numbering, a unified numbering method is adopted: the numbering range of all nodes in the diagram is set to... Among them, task nodes The number is Satellite nodes The number is .

[0052] The predicted data transmission rate is used as the communication feature vector of the satellite node for each graph node. Assigned input communication feature vector The characteristics of a task node include the amount of task data, computational load, and real-time requirements, while the characteristics of a satellite node include remaining computing resources and predicted communication rate.

[0053] S3.2. Determine the attention weights of each task and all satellites based on the graph attention mechanism to measure the scheduling adaptability between task nodes and satellite nodes.

[0054] Attention weight The calculation formula is: , in, a For trainable vectors, W The feature mapping matrix, and These are the feature vectors of the task node and the satellite node, respectively. express i The set of neighboring nodes, This is a vector concatenation operation.

[0055] S3.3. Convert the attention weights of each task into an unloading probability distribution vector. In order to convert the attention weights into unloading decisions, the attention weights of each task node and all satellite nodes are regarded as the unloading intention of the task for each candidate satellite, forming an unloading probability distribution vector: generate an unloading probability distribution vector for each task.

[0056] . This unloading probability distribution vector represents the mission's "preference score" or unloading tendency distribution for all satellites.

[0057] S3.4. The maximum value of the unloading probability distribution vector is determined by the maximum confidence selection rule. The satellite corresponding to the maximum value is taken as the target unloading node of the mission. The mission unloading matrix is ​​obtained by discretizing all the target unloading nodes. The target unloading node is the satellite that performs the mission.

[0058] The maximum confidence selection rule selects the satellite corresponding to the maximum value in the distribution as the mission. n The target unloading node is then used to obtain the discretized task unloading matrix. Its definition is

[0059] The target unloading node satisfies the following constraints:

[0060] This means that each task is assigned to only one target satellite, ensuring the determinism and executability of the scheduling results.

[0061] Step 4: After the task unloading is completed, the task unloading matrix is ​​converted into task features that match the spatial size of the hyperspectral image. Then, through channel-dimensional stitching, the hyperspectral image and task features are fused to form a fusion tensor. A high-dimensional feature vector is generated for each band region of the hyperspectral image based on the fusion tensor. Band selection of the hyperspectral image is performed based on the high-dimensional feature vector to obtain the band selection matrix.

[0062] After completing the task offloading scheme, since hyperspectral images typically contain hundreds of consecutive bands, but most band information highly overlaps, directly transmitting all bands would significantly increase communication burden, computational load, and system energy consumption. Therefore, a Multi-feature Attention Capsule Network (MACN) structure based on multi-feature fusion is designed to automatically select the most discriminative key bands from the original image, achieving the dual goals of effective compression and maintaining recognition performance. In this embodiment, to reduce the bandwidth and energy consumption required for hyperspectral image transmission, a band selection module is introduced to automatically select key bands to reduce computational and communication burden. The method is as follows: S4.1. Convert the task unloading matrix into task features that match the spatial dimensions of the hyperspectral image; Task The corresponding hyperspectral image is a tensor. ; in, H , W These represent the image space dimensions, B This indicates the number of original bands.

[0063] To incorporate the task unloading matrix into the band selection process, the task unloading matrix is ​​mapped to the same spatial scale as the image tensor, and task features are obtained through linear mapping.

[0064] in, This is the unloading vector for the task. Let be the mapping matrix, and d be the number of fusion channels.

[0065] S4.2. Concatenate the hyperspectral image and task features along the channel dimension to obtain the fusion tensor:

[0066] S4.3 Obtain the preliminary feature map of the fused tensor, and use channel attention and spatial attention to enhance the band channels and target spatial regions of the preliminary feature map to obtain a multidimensional attention feature map.

[0067] The fused tensor first passes through a convolutional module to extract low-level spatial features, resulting in a preliminary feature map. To further enhance feature representation capabilities, a Convolutional Block Attention Module (CBAM) is introduced. CBAM employs a cascaded modeling mechanism combining channel attention and spatial attention to significantly enhance key band channels and the target spatial region, respectively. The channel attention module performs max pooling and average pooling operations on each channel to extract global response features, which are then processed by a shared multilayer perceptron to generate channel attention weights. This is used to emphasize important bands; the spatial attention module performs max pooling and average pooling on the channel-weighted feature map along the channel dimension, and then... Convolution to obtain spatial attention map This is used to locate the spatial region of interest, and the final output is a multidimensional attention feature map, expressed as follows: , in, This indicates an element-wise multiplication operation.

[0068] S4.4. Generate a high-dimensional feature vector for each band region of the hyperspectral image based on the multi-dimensional attention feature map, and perform band selection of the hyperspectral image based on the high-dimensional feature vector to obtain the band selection matrix.

[0069] CBAM-enhanced feature map The data is fed into the capsule network module. Unlike traditional neural networks that only use scalar neurons (i.e., each neuron outputs only a single value), capsule networks employ a vector neuron structure. Each capsule unit represents the presence and spatial attributes of a certain image feature through a vector, thus more comprehensively depicting the complex relationship between bands and classification targets in the image. The capsule network generates a high-dimensional feature vector for each band region of the hyperspectral image through convolutional capsule layers. ,in This represents the vectorized representation of the l-th band in the task; The capsule dimension represents the amount of information contained in the vector.

[0070] Capsule networks generate a high-dimensional feature vector for each band region through convolutional capsule layers. This vector contains key information about that band. The magnitude of the vector measures the importance of that band to the current classification task; a larger magnitude indicates a higher activation level. The system sets a threshold; if the magnitude of the vector corresponding to a band exceeds this threshold, it is selected for transmission and processing; otherwise, it is filtered out. This achieves the screening of redundant bands, saving bandwidth and reducing energy consumption.

[0071] S4.5. Select bands based on the modulus of the high-dimensional feature vector and a set threshold, and construct a band selection matrix based on the selected bands.

[0072] This high-dimensional feature vector The Euclidean norm (i.e., modulus) is used to represent the "activation level" of the band for the target classification task. A decision threshold is set. Band selection is performed according to the following rules:

[0073] in, This represents the capsule vector magnitude of the l-th band. Indicates whether task n selects this band.

[0074] Step 5: Construct a joint loss function based on the task unloading loss function and the band selection loss function. Repeat steps S3-S7 to iteratively optimize the task unloading matrix and the band selection matrix until the joint loss function is minimized, thus obtaining the optimal task unloading matrix and band selection matrix.

[0075] To achieve joint optimization of task offloading and band selection, this embodiment designs a joint loss function to guide the model to simultaneously optimize the task scheduling scheme and hyperspectral image compression strategy during end-to-end training, taking into account resource constraints, classification performance, and band number control.

[0076] S5.1 Construct the unloading optimization objective function and the band selection optimization objective function according to the task unloading objective and the band selection objective respectively, and then obtain the joint loss function.

[0077] To achieve end-to-end joint optimization training, let the overall model parameters be... The joint loss function is constructed as follows:

[0078] in, This is a balancing factor used to adjust the weights of the task offloading optimization loss and the band selection loss during joint training. This represents the objective function for task unloading optimization. This represents the objective function for optimizing band selection.

[0079] The optimization objective for the task offloading part is to minimize the total energy consumption after task allocation, while simultaneously satisfying satellite resource constraints. The task offloading loss function is defined as follows:

[0080] in, The total energy consumption after the mission is allocated to the satellite, including data transmission energy consumption and edge computing energy consumption; Indicates the first z Penalty function for each constraint term; Its corresponding weighting coefficient;Z This represents the number of constraints. Resource constraints may include: satellite bandwidth limits, computing power limits, maximum concurrent task limits, etc.

[0081] For example, when satellite The available communication bandwidth is ,Task The amount of data is When the bandwidth constraint term is used, it can be written as , This is used to avoid task allocation exceeding satellite resource capacity and to guide the model to learn feasible offloading schemes.

[0082] In the band selection section, the band selection loss function Taking into account both classification accuracy and band number control, the following definition is provided. . in, For the true label of the task, The classification confidence score of this task in the current band combination (output by MACN); Indicates whether the l-th band is selected; This represents the maximum number of tapes allowed by the system. This is the regularization coefficient used to control the number of bands.

[0083] The first term, cross-entropy loss, guides the model to learn discriminative band combinations. The second term limits the total number of bands selected for each task, reducing communication pressure and controlling energy consumption. Notably, to further improve the stability and training efficiency of the joint loss function, a memory buffer mechanism is introduced in practical deployment. This mechanism preserves historical task features, scheduling states, and gradient information during model training, and uses them together with the batch samples of the current task for loss function calculation and parameter updates.

[0084] S5.2 Repeat steps 3 and 4 to iteratively optimize the task unloading matrix and band selection matrix until the joint loss function is minimized, thus obtaining the optimal task unloading matrix and band selection matrix.

[0085] By minimizing the joint loss function This enables end-to-end joint optimization of the task offloading matrix and the band selection matrix, allowing the model to balance offloading energy consumption, communication constraints, and recognition accuracy under dynamic network resource conditions, thereby improving the overall system efficiency.

[0086] The joint loss function is used to update the model parameters after each iteration. Task Unloading Solution and band selection results Finally, determine whether the current joint scheduling result meets the convergence condition. If it does, output the final task offloading matrix and band selection result; otherwise, set t=t+1, return to the link prediction step, and continue the next round of optimization iteration until the joint optimal solution converges.

[0087] Please see Figure 3 The image shown is a simulation diagram of the effectiveness of the satellite edge hyperspectral image processing method (JFIO algorithm) provided in the embodiments of the present invention. The JFIO algorithm is significantly better than other existing comparative methods in terms of average normalized hyperspectral image utility, and its advantages become more prominent as the number of low-orbit satellites increases.

[0088] Specifically, when the number of LEO satellites increases from 3 to 15, the average normalized total utility of JFIO improves by 104%, compared to relatively limited increases of 102% and 90% for the comparative algorithms MaHSI and EGSSR. In scenarios with 6 LEO satellites, JFIO's utility is about 20% higher than MaHSI, demonstrating superior task scheduling and band compression capabilities. When the number of LEO satellites expands to 9, JFIO's utility significantly exceeds the average level of other methods, exhibiting good scalability and system adaptability. In large-scale scenarios (such as with 15 LEO satellites), JFIO improves by 4% compared to MaHSI and 12% compared to EGSSR, while still maintaining high resource utilization efficiency and task processing performance. This performance improvement is attributed to the joint optimization strategy proposed in this invention, which allows JFIO to simultaneously consider the coupling relationship between task offloading and band selection, constructing an end-to-end joint optimization framework.

[0089] Please see Figure 4 This diagram illustrates the hyperspectral image classification accuracy provided in this embodiment of the invention. With a fixed number of LEO satellites (15), the JFIO algorithm demonstrates a significant classification performance advantage over other band selection methods across multiple datasets. Specifically, JFIO improves classification accuracy by an average of approximately 3%–6% across all experimental datasets. In five sets of repeated experiments, the JFIO algorithm consistently outperforms the comparison methods in classification accuracy, with an average accuracy improvement of 3.2%–3.6%. The shaded areas of the curves in the diagram represent the range of variance variation for different methods across multiple experiments. As can be seen from the diagram, the shaded area for JFIO is significantly smaller than that for other methods, indicating lower volatility, stronger robustness, and better generalization ability and stability in classification tasks.

[0090] Please see Figure 5This diagram illustrates the average energy consumption of hyperspectral image processing provided in this embodiment of the invention. The JFIO algorithm demonstrates superior energy consumption control capabilities across multiple datasets, showing a significant advantage in energy efficiency compared to traditional methods. Specifically, compared to traditional fixed-bandwidth band selection methods, JFIO can significantly reduce the total energy consumption during hyperspectral image processing, with an average energy saving of 15% to 53%. In scenarios with poor network conditions, the dynamic task offloading and band compression joint optimization mechanism proposed in this invention performs particularly well, intelligently avoiding inefficient communication paths and overloaded nodes, effectively suppressing energy consumption spikes caused by redundant transmission. Taking the FullBand strategy as an example, JFIO can achieve nearly 50% energy consumption reduction in extreme cases. Furthermore, compared to methods with some adaptive capabilities such as Greedy and OEC-TA, JFIO still exhibits 15% to 35% energy consumption optimization potential.

[0091] Please see Figure 6 This diagram illustrates the time consumption of hyperspectral image processing according to an embodiment of the present invention. JFIO also demonstrates a significant advantage in terms of time overhead, effectively reducing the total time required for hyperspectral image processing under multiple datasets and varying network loads. Specifically, compared to the traditional FullBand strategy, JFIO achieves a time overhead reduction of approximately 40% in various test environments, with a maximum reduction of up to 45%. This advantage is mainly attributed to its joint optimization strategy employed in task offloading and band selection, resulting in shorter data transmission paths and more balanced computing node loads, significantly reducing latency caused by network congestion or node overload. Even compared to other advanced algorithms with bandwidth adaptive capabilities (such as MaHSI and EGSSR), JFIO maintains a leading position in time consumption, reducing processing time by an average of 15% to 30%.

[0092] Please see Figure 7 This is a schematic diagram comparing the changes in hyperspectral image processing utility under different weight parameters, provided by an embodiment of the present invention. The parameters in the joint loss function are... The value of has a significant impact on model performance, specifically in adjusting the optimization focus between task offloading and band selection, thus affecting the final hyperspectral image processing efficiency. With smaller weight parameters... =0.3 or With a weighting factor of 0.4, the model prioritizes band selection and classification accuracy optimization, thus converging quickly in the early stages of training and achieving high utility values ​​in a short time. This setting is suitable for scenarios requiring high classification accuracy but with limited computational resources. However, because it ignores the energy balance issue during task unloading, long-term operation may lead to decreased resource utilization efficiency. When the weight parameters are set to a moderate value... When the weight ratio is 0.5, the model achieves a good balance between task scheduling and band selection, ensuring high classification accuracy while effectively reducing energy consumption and achieving optimal overall utility. As can be observed from the figure, this setting results in the highest final convergence value, demonstrating optimal performance. For larger weight parameters... =0.7 or With a convergence value of 0.8, the model places greater emphasis on energy consumption control and task offloading optimization, making it suitable for energy-sensitive tasks or long-term sustainable operation scenarios. Although its convergence speed is relatively slow, it is more practical in systems with high resource pressure.

[0093] Please see Figure 8 This diagram illustrates the comparison of average utility for hyperspectral image processing under different memory buffer configurations provided in this embodiment. The size of the memory buffer significantly impacts resource scheduling and utility performance during model training. Specifically, as the memory buffer configuration gradually increases from 0.5 times the batch size to 1 times, the model utility significantly improves. At this stage, the model can fully utilize existing batch data for feature extraction and feedback updates, thereby improving the collaborative efficiency of task offloading and band selection. When the memory buffer is configured to 1 times the batch size, meaning the number of historical task state information retained in each training round is equal to the number of current batch tasks, system resource utilization reaches its optimal level, and the model's processing utility on multiple datasets reaches its peak. This setting not only avoids information redundancy and resource waste but also improves cache hit rate and feature replay quality, making training more stable and efficient. However, when the memory configuration is further expanded to 2 times or more, the rate of utility growth begins to slow down. Although a larger memory buffer can record more task features and intermediate states, helping to build a more comprehensive state representation, the resulting system overhead offsets some of the performance gains, leading to increased model convergence fluctuations.

[0094] This embodiment of the satellite edge hyperspectral image processing method adopts the JFIO framework, which can effectively alleviate the system load problem caused by limited satellite computing resources and communication bandwidth bottlenecks. By introducing an LSTM-based link quality prediction mechanism, the communication status is dynamically perceived, improving the foresight of task scheduling. Graph Attention Network (GAT) is used to adaptively model the resource coupling relationship between tasks and satellites, generating task offloading schemes that meet resource constraints. Attention Capsule Network (MACN) combined with CBAM module achieves fine-grained band selection, suppresses redundant data, and significantly improves image classification performance and compression efficiency. In addition, a joint loss function is designed to uniformly model task energy consumption, communication limitations, and image discrimination performance, achieving end-to-end collaborative optimization of task offloading and band selection. Under the premise of meeting system resource constraints, image processing accuracy is maximized while overall system energy consumption and data transmission volume are minimized. Finally, the optimized task offloading scheme and band selection scheme are output, corresponding to the task scheduling matrix and band selection matrix, respectively, which can directly guide multiple satellites to collaboratively complete hyperspectral image perception, processing, and compression tasks in communication-constrained environments.

[0095] Example 2 A satellite edge hyperspectral image processing system, comprising: The task offloading module is used to assign the satellite's future data transmission rate as a communication feature vector to the satellite node, determine the attention weight of each task node and all satellite nodes, construct the offloading probability distribution vector of each task based on the attention weight, and construct the task offloading matrix based on all the offloading probability distribution vectors. The band selection module is used to convert the task unloading matrix into task features, fuse the hyperspectral image with the task features to form a fusion tensor, generate a high-dimensional feature vector for each band region of the hyperspectral image based on the fusion tensor, and perform band selection of the hyperspectral image based on the high-dimensional feature vector to obtain the band selection matrix. The iterative module is used to construct a joint loss function based on the task unloading loss function and the band selection loss function. The joint loss function is then used to iteratively optimize the task unloading matrix and the band selection matrix until the joint loss function is minimized, thus obtaining the optimal task unloading matrix and band selection matrix.

[0096] It should be noted that, in the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another device, or some features may be ignored or not executed. The modules described as separate components may or may not be physically separated. The components shown as modules may be one or more physical units, that is, they may be located in one place or distributed in multiple different places. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment according to actual needs.

[0097] Furthermore, in the various embodiments of the present invention, the modules can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The integrated unit described above can be implemented in hardware or as a software functional unit.

[0098] An electronic device provided in this application includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the satellite edge hyperspectral image processing method described in any of the above embodiments.

[0099] Another electronic device provided in this application embodiment may further include: an input port connected to a processor for transmitting multimodal data collected by an external acquisition device to the processor; a display unit connected to the processor for displaying the processor's processing results to the outside world; and a communication module connected to the processor for enabling communication between the electronic device and the outside world. The display unit may be a display panel, a laser scanning display, etc.; the communication method adopted by the communication module includes, but is not limited to, Mobile High Definition Link (HML), Universal Serial Bus (USB), High Definition Multimedia Interface (HDMI), and wireless connection (including Wi-Fi, Bluetooth, Bluetooth Low Energy, and IEEE 802.11s-based communication technology).

[0100] This application provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the steps of the satellite edge hyperspectral image processing method described in any of the above embodiments.

[0101] For descriptions of relevant parts of the satellite edge hyperspectral image processing method system, electronic device, and computer-readable storage medium provided in this application's embodiments, please refer to the detailed descriptions of the corresponding parts in the satellite edge hyperspectral image processing method provided in this application's embodiments; they will not be repeated here. Furthermore, parts of the technical solutions provided in this application that are consistent with the implementation principles of corresponding technical solutions in the prior art have not been described in detail to avoid excessive elaboration.

[0102] The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A method for processing satellite edge hyper-spectral images, characterized in that, The method comprises the following steps: Step 1: The data transmission rate of the satellite at the future time is assigned to the satellite node as a communication feature vector, the attention weight of each task node and all satellite nodes is determined, the unloading probability distribution vector of each task is constructed according to the attention weight, and the task unloading matrix is constructed according to all the unloading probability distribution vectors; Step 2: The task unloading matrix is converted into task features, the hyperspectral image is fused with the task features to form a fusion tensor, a high-dimensional feature vector is generated for each band region of the hyperspectral image according to the fusion tensor, and the band selection of the hyperspectral image is performed according to the high-dimensional feature vector to obtain a band selection matrix; Step 3: A joint loss function is constructed based on a task unloading loss function and a band selection loss function, and the task unloading matrix and the band selection matrix are iteratively optimized by using the joint loss function until the joint loss function is minimized, so that the optimal task unloading matrix and the band selection matrix are obtained.

2. The method of claim 1, wherein, The method for determining the data transmission rate of the satellite at the future time is as follows: According to the historical communication channel sequence constructed according to the historical communication channel characteristics of the satellite, a pre-trained time sequence network model is used to predict the historical communication channel sequence to obtain the data transmission rate of the satellite at the future time.

3. The method of claim 1, wherein, The method for assigning the data transmission rate of the satellite at the future time to the satellite node as a communication feature vector and determining the attention weight of each task node and all satellite nodes comprises the following steps: A graph is constructed based on the task node and the satellite node, the data transmission rate is assigned to the satellite node as a communication feature vector, and the attention weight of each task and all satellites is determined based on a graph attention mechanism.

4. The method of claim 1, wherein, The method for constructing the unloading probability distribution vector of each task according to the attention weight and constructing the task unloading matrix according to all the unloading probability distribution vectors comprises the following steps: The attention weight of each task is converted into an unloading probability distribution vector, the attention weight of each task node and all satellite nodes is regarded as the unloading intention of the task to each candidate satellite, and an unloading probability distribution vector is formed; The maximum value of the unloading probability distribution vector is determined by using the maximum confidence selection rule, the satellite corresponding to the maximum value is the target unloading node of the task, and the task unloading matrix is obtained by discretizing all the target unloading nodes.

5. The method of claim 1, wherein, The method for converting the task unloading matrix into task features and fusing the hyperspectral image with the task features to form a fusion tensor comprises the following steps: The task unloading matrix is converted into task features matching the spatial size of the hyperspectral image, and the hyperspectral image and the task features are fused to form a fusion tensor through a channel dimension splicing operation.

6. The method of claim 1 or 5, wherein, The method for generating a high-dimensional feature vector for each band region of the hyperspectral image according to the fusion tensor and performing band selection of the hyperspectral image according to the high-dimensional feature vector to obtain a band selection matrix comprises the following steps: A preliminary feature map of the fusion tensor is obtained, and the band channel and the target spatial region of the preliminary feature map are enhanced by using channel attention and spatial attention to obtain a multi-dimensional attention feature map; The band selection of the hyperspectral image is performed according to the length of the high-dimensional feature vector and in combination with a set threshold value, and a band selection matrix is constructed according to the selected band.

7. The method of claim 6, wherein the satellite edge high-spectral image is processed by: The method for obtaining the multi-dimensional attention feature map comprises the following steps: The preliminary feature map is input into a convolution block attention mechanism module, and through a series connection mechanism of channel attention and spatial attention, a wave band channel and a target spatial region of the multi-dimensional attention feature map are enhanced respectively to obtain the multi-dimensional attention feature map.

8. The method of claim 6, wherein the satellite edge high-spectral image is processed by: The band selection of the hyperspectral image according to the length of the high-dimensional feature vector and in combination with the set threshold value comprises: The multi-dimensional attention feature map is input into a capsule network, the capsule network generates a high-dimensional feature vector for each wave band region of the hyperspectral image through a convolution capsule layer, the length of the high-dimensional feature vector is compared with a threshold value, the band selection of the hyperspectral image is performed according to a comparison result, and a band selection matrix is constructed according to the selected wave band.

9. The method of claim 6, wherein the satellite edge high-spectral image is processed by: An expression of the joint loss function is as follows: wherein, is a balancing factor, is a task offloading loss function; is a band selection loss function; The task offloading loss function is as follows: in, The total energy consumption after the mission is allocated to the satellite. For the first z Penalty function for each constraint term; These are weighting coefficients; Z The number of constraint terms; The band selection loss function As follows: . wherein, is the true label of the task, is the classification confidence of the task under the current band combination; denotes whether the lth band is selected or not; is the maximum number of selected bands, is the regularization coefficient for band number control.

10. A satellite edge high spectral image processing system, characterized in that, The task offloading loss function comprises: The task offloading module is configured to assign a data transmission rate of a satellite at a future time as a communication feature vector to the satellite node, determine an attention weight of each task node and all satellite nodes, construct an offloading probability distribution vector of each task according to the attention weight, and construct a task offloading matrix according to all offloading probability distribution vectors. The band selection module is configured to convert the task offloading matrix into a task feature, fuse the hyperspectral image and the task feature to form a fusion tensor, generate a high-dimensional feature vector for each wave band region of the hyperspectral image according to the fusion tensor, and perform the band selection of the hyperspectral image according to the high-dimensional feature vector to obtain a band selection matrix. The iteration module is configured to construct a joint loss function based on the task offloading loss function and the band selection loss function, perform iterative optimization on the task offloading matrix and the band selection matrix by using the joint loss function, and obtain optimal task offloading matrix and band selection matrix until the joint loss function is minimized.

Images