High-resolution remote sensing image processing method based on superpixel cloud edge collaboration

Abstract

The application discloses a high-resolution remote sensing image processing method based on super-pixel cloud edge cooperation, belongs to the technical field of remote sensing image processing, and is used for high-resolution remote sensing image processing. The method comprises the following steps: acquiring remote sensing images and performing data preprocessing, constructing a neural network to output a final fusion feature vector; setting a clustering center on the remote sensing image, calculating an adaptive step and constructing a comprehensive distance, and obtaining a super-pixel segmentation result by iteratively clustering the center; decoupling the processing flow into atomic sub-tasks, constructing a state vector by a cloud server, and executing tasks by an edge node; establishing a distributed storage pool, setting a scheduling decision, constructing a final result map according to the processing result output after the task of the edge node. Through multi-scale feature fusion, adaptive super-pixel segmentation and reinforcement learning scheduling, the application realizes the collaborative optimization of the efficiency, real-time performance and system resource utilization rate of high-resolution remote sensing image processing.

Description

Technical Field

[0001] This invention discloses a high-resolution remote sensing image processing method based on superpixel cloud-edge collaboration, belonging to the field of remote sensing image processing technology. Background Technology

[0002] With the rapid development of Earth observation technology, the amount of high-resolution remote sensing imagery data is growing exponentially. This data is widely used in fields such as digital cities, environmental monitoring, disaster assessment, and agricultural censuses. However, traditional remote sensing image processing methods face bottlenecks in both computational efficiency and storage I / O when processing massive amounts of data, ranging from TB to PB. First, the computational efficiency bottleneck is significant. The large size of individual images makes conventional single-machine serial processing extremely time-consuming, which is difficult to meet the needs of emergency response and real-time analysis. Second, storage and I / O bottlenecks are significant. Centralizing massive datasets in a single data center and calling them as a whole during processing not only places high demands on network bandwidth but also results in substantial I / O latency, leading to low actual utilization of high-performance computing resources.

[0003] Existing technologies typically employ simple rule-based segmentation (such as uniform grid partitioning) combined with distributed computing frameworks (such as Hadoop and Spark) for processing. However, this approach has inherent drawbacks: rule-based segmentation ignores the intrinsic relationships within image content, easily fragmenting complete ground features (such as a road or a field), leading to decreased accuracy in subsequent feature extraction and target recognition tasks. Furthermore, existing cloud-edge collaborative computing models often employ static or simple polling strategies for task scheduling, failing to fully consider the heterogeneity and dynamism of edge node computing power, as well as network state fluctuations. This prevents optimal matching between computing tasks and computing resources, significantly reducing overall processing efficiency. In addition, existing methods lack native support for multi-temporal data collaborative analysis and fail to achieve deep binding between computing tasks and storage locations, resulting in the migration of a large amount of redundant data.

[0004] Therefore, there is an urgent need in this field for a new method that can understand image content, perceive ground features, support temporal analysis, and enable intelligent, efficient collaboration and in-memory computing in a distributed environment, so as to fundamentally improve the automation level and execution efficiency of massive remote sensing image processing. Summary of the Invention

[0005] The purpose of this invention is to provide a high-resolution remote sensing image processing method based on superpixel cloud-edge collaboration, so as to solve the problems of insufficient real-time performance and interpretation accuracy caused by large computational load, high transmission delay and limited edge resources in the cloud-edge collaborative processing of high-resolution remote sensing images in the prior art.

[0006] High-resolution remote sensing image processing methods based on superpixel cloud-edge collaboration include: S1. Acquire the remote sensing image data to be processed and perform data preprocessing. Construct a multi-scale deep feature extraction network to extract spectral features, texture features and temporal variation features, and stitch them together in the channel dimension. Perform adaptive weighting and dimensionality reduction through a gated fusion unit to obtain the final fused feature vector of each pixel. S2. Based on the final fused feature vector, cluster centers are uniformly set on the remote sensing image. The search range is constructed by calculating the adaptive step size based on the cluster centers. The comprehensive distance between the pixels and the cluster centers within the search range is calculated using the adaptive distance metric formula. The cluster centers are iteratively updated until convergence is achieved to obtain the superpixel segmentation result. A unique identifier is assigned to each superpixel block, and an adjacency graph between superpixel blocks is constructed. S3. Decouple the remote sensing image processing flow into atomic subtasks, construct a task dependency graph based on the adjacency graph, obtain the state information of each edge node and construct a state vector, select the edge node and model version to execute the task according to the state vector, calculate the reward value and update the decision model parameters after the task is executed, and establish a caching mechanism based on cloud server storage and edge node storage. S4. Based on cloud server storage and edge node storage, a distributed storage pool is established. Each superpixel block in the distributed storage pool corresponds to a data container. A distributed hash table is used to store the metadata of the data container. The scheduling decision prioritizes the edge nodes that have cached the data required by the task to execute the task. After the edge node completes the task, it outputs the processing result. Based on the unique identifier and the adjacency relationship graph, the processing results of all edge nodes are spliced ​​together to obtain the final result graph.

[0007] S1 includes S1.1, data preprocessing including radiometric calibration, atmospheric correction, geometric fine correction and multi-temporal registration, to obtain a standardized data cube; S1 includes S1.2, constructing a multi-scale deep feature extraction network, using a pre-trained ResNet-Transformer hybrid network as the backbone layer to extract spectral features, texture features and temporal variation features; The output of deep convolutional layers is used as spectral features. ; The gray-level co-occurrence matrix is ​​calculated based on the shallow feature map, and the contrast ratio is extracted from it. Entropy : ; ; In the formula, The coordinates in the gray-level co-occurrence matrix are The probability value; Will and Merge into texture features Calculate the time series variance of each pixel on the spectral index. As a characteristic of temporal change : ; In the formula, For the coordinates Pixels at different times The variance of the sequence formed by a certain spectral index value, For the current pixel in the th... The index value of each phase; S1 includes, S1.3, and will , and The splicing is performed along the channel dimension and then fused using a gated fusion unit. Adaptive weighting and dimensionality reduction are performed to generate the final fused feature vector for each pixel. : ; In the formula, This is for channel splicing operations.

[0008] S2 includes, S2.1, uniformly setting on the remote sensing image. Each cluster center is assigned a unique label. The center point of each cluster center is used as the feature vector. : ; In the formula, These are CIELAB color space values. For pixel coordinates, It is a normalized time sequence identifier. for Index; S2 includes S2.2, and the initial step size. : ; In the formula, This represents the total number of pixels. based on Calculate adaptive step size : ; In the formula, The image entropy is the local region centered on the current cluster center. For the global maximum entropy, This refers to the importance weight scalar value of the task expected to be processed in the current region, given by the task parser. An adjustment coefficient is used to control the intensity of the influence of image content complexity on the granularity of the segments. An adjustment coefficient for controlling the intensity of the impact of task requirements on the granularity of the blocks.

[0009] S2 includes, S2.3, taking the center point of the cluster center as the center, and... Using a radius, construct a pixel search range, and use an adaptive distance metric formula to calculate the comprehensive distance between pixels within the search range and cluster centers. : ; In the formula, For pixels and Euclidean distance in the CIELAB color space For pixels and In coordinate space Euclidean distance in for The cosine distance between them For pixels and In the time series dimension The absolute difference on For the normalized weights of color distance, The normalized weights are the feature distances. The normalized weights for time-series distances.

[0010] S2 includes S2.4, repeating steps S2.2 and S2.3, to calculate the relationship between each pixel and all cluster centers. Value, then select Minimum cluster center The current pixel is used as the label, and the position of the cluster center is recalculated until the cluster center converges to the maximum number of iterations. S2 includes, S2.5, and allocation based on each pixel. The initial superpixel segmentation results are obtained. A forced connectivity post-processing method is used to minimize these results. Mathematical morphology operators are then used to remove isolated noise and incomplete boundaries, resulting in the final superpixel blocks. A globally unique identifier is assigned to each superpixel block. And construct an enhanced adjacency graph. ,vertex For each superpixel block, the edges Spatial adjacency relationships between superpixel blocks, attribute Record the similarity in feature space and temporal correlation between the two superpixel blocks connected by each edge.

[0011] S3 includes, S3.1, decoupling the remote sensing image processing workflow into atomic subtasks, and based on... Construct a directed acyclic task dependency graph Each task is described as: ; In the formula, For the first Atom sub-tasks, For atomic subtask indexing, , This is an identifier for the type of algorithm that the task needs to execute. The address for storing the input data required for task execution. Weighted priority for tasks. A set of precursor tasks for the main task; S3 includes, S3.2, assuming the data processing system consists of a cloud server and The system consists of several edge nodes. At each decision moment, the cloud server collects the state information of all edge nodes and constructs a state vector. ; In the formula, This is a set of CPU and memory utilization rates for each edge node. The set of available memory for each edge node. For network bandwidth collection, This is the set of task queue lengths for each edge node. This is a set of energy consumption states for edge nodes. This is a set of data block version identifiers locally cached by each edge node; Cloud servers are based on the current Table, output action : ; In the formula, Number the selected edge nodes. Number the selected lightweight model version; The table is a two-dimensional data table, with rows corresponding to... , column corresponding The value of each element In order to be in Execute The expected value of cumulative rewards; implement The reward function is calculated afterward: ; In the formula, For the task In the selected node Use the selected model version The estimated completion time, If The data transfer time required without holding the data blocks needed for the task. For the task At the node The energy consumption estimate for the above execution, To estimate the accuracy gain of a specific lightweight model version compared to the baseline model, , , , These are the weighting coefficients. As an indicator function, when the data required by the task is already available... The value is 1 if it is in the local cache, and 0 otherwise.

[0012] S3 includes S3.3, which uses the Q-learning algorithm for updating. value: ; In the formula, For learning rate, As a discount factor, For the next state All possible actions The largest value; S3 includes S3.4, establishing a three-level caching mechanism, including edge node local memory L1, edge node local solid-state storage L2, and cloud storage L3; based on Divide upstream and downstream tasks, and obtain the scheduling probability of each downstream task after the upstream task is completed by statistically analyzing the historical task execution sequence. When the current upstream task is about to be completed, load the data block with the highest scheduling probability from L3 or L2 into L1.

[0013] S4 includes S4.1, establishing a distributed storage pool based on L2 and L3. Each superpixel block in the distributed storage pool corresponds to a data container. The fields inside the container include data header, pixel data, intermediate results, and a set of operation identifiers. S4 includes S4.2, which uses distributed hash table technology to manage the data index, distributing metadata across edge nodes. The metadata is: ; In the formula, To store a list of edge node numbers for the current container replica. This represents the historical access frequency of the current container. For container data size, For checksums used for data integrity verification, This is a copy of the container header; S4 includes S4.3, where scheduling decisions follow the data affinity principle, which prioritizes assigning tasks to edge nodes that have cached the data containers required by the current task; During task execution, edge nodes calculate intermediate states as snapshots and write them into the intermediate result field of the data container of the current superpixel block.

[0014] S4 includes S4.4, periodically scanning the metadata of each data container, setting a frequency threshold, data above the frequency threshold is hot data, and data below or equal to the frequency threshold is cold data; For hot data, increase the number of replicas in the L2 cache of edge nodes, and store it in L1 when L2 is insufficient; For cold data, compress and archive it to reduce the number of copies; Superpixel block containers with high spatiotemporal correlation are stored contiguously in physical storage or merged into a single archive file. High spatiotemporal correlation is... middle Records the relationships between different time phases of the same land feature.

[0015] S4 includes S4.5, after all edge nodes have completed their respective superpixel block tasks, they write the processing results back to the corresponding data container; according to and Following the order of the superpixel blocks, the processing results of the data containers are read sequentially, and the built-in splicing operator of the data containers is called to splice the output results of all containers into the final result image.

[0016] Compared with existing technologies, the present invention has the following advantages: By introducing intelligent segmentation with multi-temporal feature fusion, cloud-edge collaborative reasoning based on federated scheduling, and a data container integrating storage and computing, the present invention improves the traditional rule-based segmentation method into an adaptive segmentation method with spatiotemporal semantic awareness. Furthermore, by reducing data migration overhead through a tight storage-computing coupling design, the present invention improves the accuracy, efficiency, and robustness of massive remote sensing image processing, and can meet the wide range of application needs from real-time analysis to long-term monitoring. Attached Figure Description

[0017] Figure 1 This is a flowchart of the technology of this invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0019] High-resolution remote sensing image processing methods based on superpixel cloud-edge collaboration include: S1. Acquire the remote sensing image data to be processed and perform data preprocessing. Construct a multi-scale deep feature extraction network to extract spectral features, texture features and temporal variation features, and stitch them together in the channel dimension. Perform adaptive weighting and dimensionality reduction through a gated fusion unit to obtain the final fused feature vector of each pixel. S2. Based on the final fused feature vector, cluster centers are uniformly set on the remote sensing image. The search range is constructed by calculating the adaptive step size based on the cluster centers. The comprehensive distance between the pixels and the cluster centers within the search range is calculated using the adaptive distance metric formula. The cluster centers are iteratively updated until convergence is achieved to obtain the superpixel segmentation result. A unique identifier is assigned to each superpixel block, and an adjacency graph between superpixel blocks is constructed. S3. Decouple the remote sensing image processing flow into atomic subtasks, construct a task dependency graph based on the adjacency graph, obtain the state information of each edge node and construct a state vector, select the edge node and model version to execute the task according to the state vector, calculate the reward value and update the decision model parameters after the task is executed, and establish a caching mechanism based on cloud server storage and edge node storage. S4. Based on cloud server storage and edge node storage, a distributed storage pool is established. Each superpixel block in the distributed storage pool corresponds to a data container. A distributed hash table is used to store the metadata of the data container. The scheduling decision prioritizes the edge nodes that have cached the data required by the task to execute the task. After the edge node completes the task, it outputs the processing result. Based on the unique identifier and the adjacency relationship graph, the processing results of all edge nodes are spliced ​​together to obtain the final result graph.

[0020] S1 includes S1.1, data preprocessing including radiometric calibration, atmospheric correction, geometric fine correction and multi-temporal registration, to obtain a standardized data cube; S1 includes S1.2, constructing a multi-scale deep feature extraction network, using a pre-trained ResNet-Transformer hybrid network as the backbone layer to extract spectral features, texture features and temporal variation features; The output of deep convolutional layers is used as spectral features. ; The gray-level co-occurrence matrix is ​​calculated based on the shallow feature map, and the contrast ratio is extracted from it. Entropy : ; ; In the formula, The coordinates in the gray-level co-occurrence matrix are The probability value; Will and Merge into texture features Calculate the time series variance of each pixel on the spectral index. As a characteristic of temporal change : ; In the formula, For the coordinates Pixels at different times The variance of the sequence formed by a certain spectral index value, For the current pixel in the th... The index value of each phase; S1 includes, S1.3, and will , and The splicing is performed along the channel dimension and then fused using a gated fusion unit. Adaptive weighting and dimensionality reduction are performed to generate the final fused feature vector for each pixel. : ; In the formula, This is for channel splicing operations.

[0021] S2 includes, S2.1, uniformly setting on the remote sensing image. Each cluster center is assigned a unique label. The center point of each cluster center is used as the feature vector. : ; In the formula, These are CIELAB color space values. For pixel coordinates, It is a normalized time sequence identifier. for Index; S2 includes S2.2, and the initial step size. : ; In the formula, This represents the total number of pixels. based on Calculate adaptive step size : ; In the formula, The image entropy is the local region centered on the current cluster center. For the global maximum entropy, This refers to the importance weight scalar value of the task expected to be processed in the current region, given by the task parser. An adjustment coefficient is used to control the intensity of the influence of image content complexity on the granularity of the segments. An adjustment coefficient for controlling the intensity of the impact of task requirements on the granularity of the blocks.

[0022] S2 includes, S2.3, taking the center point of the cluster center as the center, and... Using a radius, construct a pixel search range, and use an adaptive distance metric formula to calculate the comprehensive distance between pixels within the search range and cluster centers. : ; In the formula, For pixels and Euclidean distance in the CIELAB color space For pixels and In coordinate space Euclidean distance in for The cosine distance between them For pixels and In the time series dimension The absolute difference on For the normalized weights of color distance, The normalized weights are the feature distances. The normalized weights for time-series distances.

[0023] S2 includes S2.4, repeating steps S2.2 and S2.3, to calculate the relationship between each pixel and all cluster centers. Value, then select Minimum cluster center The current pixel is used as the label, and the position of the cluster center is recalculated until the cluster center converges to the maximum number of iterations. S2 includes, S2.5, and allocation based on each pixel. The initial superpixel segmentation results are obtained. A forced connectivity post-processing method is used to minimize these results. Mathematical morphology operators are then used to remove isolated noise and incomplete boundaries, resulting in the final superpixel blocks. A globally unique identifier is assigned to each superpixel block. And construct an enhanced adjacency graph. ,vertex For each superpixel block, the edges Spatial adjacency relationships between superpixel blocks, attribute Record the similarity in feature space and temporal correlation between the two superpixel blocks connected by each edge.

[0024] S3 includes, S3.1, decoupling the remote sensing image processing workflow into atomic subtasks, and based on... Construct a directed acyclic task dependency graph Each task is described as: ; In the formula, For the first Atom sub-tasks, For atomic subtask indexing, , This is an identifier for the type of algorithm that the task needs to execute. The address for storing the input data required for task execution. Weighted priority for tasks. A set of precursor tasks for the main task; S3 includes, S3.2, assuming the data processing system consists of a cloud server and The system consists of several edge nodes. At each decision moment, the cloud server collects the state information of all edge nodes and constructs a state vector. ; In the formula, This is a set of CPU and memory utilization rates for each edge node. The set of available memory for each edge node. For network bandwidth collection, This is the set of task queue lengths for each edge node. This is a set of energy consumption states for edge nodes. This is a set of data block version identifiers locally cached by each edge node; Cloud servers are based on the current Table, output action : ; In the formula, Number the selected edge nodes. Number the selected lightweight model version; The table is a two-dimensional data table, with rows corresponding to... , column corresponding The value of each element In order to be in Execute The expected value of cumulative rewards; implement The reward function is calculated afterward: ; In the formula, For the task In the selected node Use the selected model version The estimated completion time, If The data transfer time required without holding the data blocks needed for the task. For the task At the node The energy consumption estimate for the above execution, To estimate the accuracy gain of a specific lightweight model version compared to the baseline model, , , , These are the weighting coefficients. As an indicator function, when the data required by the task is already available... The value is 1 if it is in the local cache, and 0 otherwise.

[0025] S3 includes S3.3, which uses the Q-learning algorithm for updating. value: ; In the formula, For learning rate, As a discount factor, For the next state All possible actions The largest value; S3 includes S3.4, establishing a three-level caching mechanism, including edge node local memory L1, edge node local solid-state storage L2, and cloud storage L3; based on Divide upstream and downstream tasks, and obtain the scheduling probability of each downstream task after the upstream task is completed by statistically analyzing the historical task execution sequence. When the current upstream task is about to be completed, load the data block with the highest scheduling probability from L3 or L2 into L1.

[0026] S4 includes S4.1, establishing a distributed storage pool based on L2 and L3. Each superpixel block in the distributed storage pool corresponds to a data container. The fields inside the container include data header, pixel data, intermediate results, and a set of operation identifiers. S4 includes S4.2, which uses distributed hash table technology to manage the data index, distributing metadata across edge nodes. The metadata is: ; In the formula, To store a list of edge node numbers for the current container replica. This represents the historical access frequency of the current container. For container data size, For checksums used for data integrity verification, This is a copy of the container header; S4 includes S4.3, where scheduling decisions follow the data affinity principle, which prioritizes assigning tasks to edge nodes that have cached the data containers required by the current task; During task execution, edge nodes calculate intermediate states as snapshots and write them into the intermediate result field of the data container of the current superpixel block.

[0027] S4 includes S4.4, periodically scanning the metadata of each data container, setting a frequency threshold, data above the frequency threshold is hot data, and data below or equal to the frequency threshold is cold data; For hot data, increase the number of replicas in the L2 cache of edge nodes, and store it in L1 when L2 is insufficient; For cold data, compress and archive it to reduce the number of copies; Superpixel block containers with high spatiotemporal correlation are stored contiguously in physical storage or merged into a single archive file. High spatiotemporal correlation is... middle Records the relationships between different time phases of the same land feature.

[0028] S4 includes S4.5, after all edge nodes have completed their respective superpixel block tasks, they write the processing results back to the corresponding data container; according to and Following the order of the superpixel blocks, the processing results of the data containers are read sequentially, and the built-in splicing operator of the data containers is called to splice the output results of all containers into the final result image.

[0029] The initialization of cluster centers in this invention refers to uniformly distributing cluster centers across the entire remote sensing image at fixed grid intervals before the algorithm iteration begins. There are 10 seed points. Each seed point serves as an initial cluster center, possessing coordinates and a multi-dimensional feature vector at that location. Its purpose is to provide an initial reference for subsequent pixel allocation and center iteration updates, ensuring the spatial uniformity and feature representativeness of the segmentation results.

[0030] The present invention performs minimization processing on the preliminary segmentation results in order to merge superpixel blocks with excessively small areas (such as only a few dozen pixels or a few filaments) into adjacent superpixel blocks and eliminate isolated sporadic pixels.

[0031] The scheduling decision in this invention follows the data affinity principle. Specifically, when there are superpixel block processing tasks to be scheduled, the scheduler first queries the metadata. The system obtains a list of all edge nodes that have cached the data containers required for the task. If the list is not empty, it selects the edge node with the lowest load (lowest CPU utilization or shortest task queue) as the target node. If the list is empty, it selects the target node according to the reinforcement learning scheduling strategy and transfers the required data containers from L3 or other nodes to the L2 cache of the target node before task execution. During task execution, the edge node calculates the intermediate state as a snapshot and writes it to the intermediate result field in the data container of the current superpixel block.

[0032] The process of this invention is as follows Figure 1 As shown, the system comprises four modules: data preprocessing and feature analysis, superpixel intelligent segmentation, cloud-edge collaborative processing, and distributed storage and output. After inputting high-resolution remote sensing image data, the data preprocessing and feature analysis module performs data preprocessing (radiometric calibration, atmospheric correction, and geometric correction), multi-scale feature extraction (spectral features, texture features), and feature fusion (weighted fusion and dimensionality reduction). The data is then input into the superpixel intelligent module, where cluster centers are initialized, and adaptive distance measurement and adaptive granularity control are sequentially applied to iteratively optimize the cluster centers (pixel assignment and center point update). Convergence is then determined (the maximum number of iterations is set to 10). If convergence fails, it returns to iterative optimization; if convergence occurs, post-processing and adjacency relationship construction are performed. The superpixel intelligent segmentation results are input into the cloud-edge collaborative processing module for task decoupling and atomization, followed by reinforcement learning scheduling and Q-value updates. Finally, a three-level caching mechanism is established. Based on cloud-edge collaborative processing, the distributed storage module constructs a distributed storage pool, performs metadata management and tightly coupled storage-computation scheduling, and then performs data lifecycle management. The collaborative results are aggregated, and the processed results are output.

[0033] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A high-resolution remote sensing image processing method based on superpixel cloud-edge collaboration, characterized in that, include: S1. Acquire the remote sensing image data to be processed and perform data preprocessing. Construct a multi-scale deep feature extraction network to extract spectral features, texture features and temporal variation features, and stitch them together in the channel dimension. Perform adaptive weighting and dimensionality reduction through a gated fusion unit to obtain the final fused feature vector of each pixel. S2. Based on the final fused feature vector, cluster centers are uniformly set on the remote sensing image. The search range is constructed by calculating the adaptive step size based on the cluster centers. The comprehensive distance between the pixels and the cluster centers within the search range is calculated using the adaptive distance metric formula. The cluster centers are iteratively updated until convergence is achieved to obtain the superpixel segmentation result. A unique identifier is assigned to each superpixel block, and an adjacency graph between superpixel blocks is constructed. S3. Decouple the remote sensing image processing flow into atomic subtasks, construct a task dependency graph based on the adjacency graph, obtain the state information of each edge node and construct a state vector, select the edge node and model version to execute the task according to the state vector, calculate the reward value and update the decision model parameters after the task is executed, and establish a caching mechanism based on cloud server storage and edge node storage. S4. Based on cloud server storage and edge node storage, a distributed storage pool is established. Each superpixel block in the distributed storage pool corresponds to a data container. A distributed hash table is used to store the metadata of the data container. The scheduling decision prioritizes the edge nodes that have cached the data required by the task to execute the task. After the edge node completes the task, it outputs the processing result. Based on the unique identifier and the adjacency relationship graph, the processing results of all edge nodes are spliced ​​together to obtain the final result graph.

2. The high-resolution remote sensing image processing method based on superpixel cloud-edge collaboration according to claim 1, characterized in that, S1 includes S1.1, data preprocessing including radiometric calibration, atmospheric correction, geometric fine correction and multi-temporal registration, to obtain a standardized data cube; S1 includes S1.2, constructing a multi-scale deep feature extraction network, using a pre-trained ResNet-Transformer hybrid network as the backbone layer to extract spectral features, texture features and temporal variation features; The output of deep convolutional layers is used as spectral features. ; The gray-level co-occurrence matrix is ​​calculated based on the shallow feature map, and the contrast ratio is extracted from it. Entropy : ； ； In the formula, The coordinates in the gray-level co-occurrence matrix are The probability value; Will and Merge into texture features Calculate the time series variance of each pixel on the spectral index. As a characteristic of temporal change : ； In the formula, For the coordinates Pixels at different times The variance of the sequence formed by a certain spectral index value, For the current pixel in the th... The index value of each phase; S1 includes, S1.3, and will , and The splicing is performed along the channel dimension and then fused using a gated fusion unit. Adaptive weighting and dimensionality reduction are performed to generate the final fused feature vector for each pixel. : ； In the formula, This is for channel splicing operations.

3. The high-resolution remote sensing image processing method based on superpixel cloud-edge collaboration according to claim 2, characterized in that, S2 includes, S2.1, uniformly setting on the remote sensing image. Each cluster center is assigned a unique label. The center point of each cluster center is used as the feature vector. : ； In the formula, These are CIELAB color space values. For pixel coordinates, It is a normalized time sequence identifier. for The index; S2 includes S2.2, and the initial step size. : ； In the formula, This represents the total number of pixels. based on Calculate adaptive step size : ； In the formula, The image entropy is the local region centered on the current cluster center. For the global maximum entropy, This refers to the importance weight scalar value of the task expected to be processed in the current region, given by the task parser. An adjustment coefficient is used to control the intensity of the influence of image content complexity on the granularity of the segments. An adjustment coefficient for controlling the intensity of the impact of task requirements on the granularity of the blocks.

4. The high-resolution remote sensing image processing method based on superpixel cloud-edge collaboration according to claim 3, characterized in that, S2 includes, S2.3, taking the center point of the cluster center as the center, and... Using a radius, construct a pixel search range, and use an adaptive distance metric formula to calculate the comprehensive distance between pixels within the search range and cluster centers. : ； In the formula, For pixels and Euclidean distance in the CIELAB color space For pixels and In coordinate space Euclidean distance in for The cosine distance between them For pixels and In the time series dimension The absolute difference on For the normalized weights of color distance, The normalized weights are the feature distances. The normalized weights for time-series distances.

5. The high-resolution remote sensing image processing method based on superpixel cloud-edge collaboration according to claim 4, characterized in that, S2 includes S2.4, repeating steps S2.2 and S2.3, to calculate the relationship between each pixel and all cluster centers. Value, then select Minimum cluster center The current pixel is used as the label, and the position of the cluster center is recalculated until the cluster center converges to the maximum number of iterations. S2 includes, S2.5, and allocation based on each pixel. The preliminary superpixel segmentation results are obtained. The forced connectivity post-processing method is used to minimize the preliminary segmentation results. Mathematical morphology operators are used to remove isolated noise and incomplete boundaries to obtain the final superpixel blocks. Assign a globally unique identifier to each superpixel block And construct an enhanced adjacency graph. ,vertex For each superpixel block, the edges Spatial adjacency between superpixel blocks, attribute Record the similarity in feature space and temporal correlation between the two superpixel blocks connected by each edge.

6. The high-resolution remote sensing image processing method based on superpixel cloud-edge collaboration according to claim 5, characterized in that, S3 includes, S3.1, decoupling the remote sensing image processing workflow into atomic subtasks, and based on... Construct a directed acyclic task dependency graph Each task is described as: ； In the formula, For the first Atom sub-tasks, For atomic subtask indexing, , This is an identifier for the type of algorithm that the task needs to execute. The address for storing the input data required for task execution. Weighted priority for tasks. A set of precursor tasks for the task; S3 includes, S3.2, assuming the data processing system consists of a cloud server and The system consists of several edge nodes. At each decision moment, the cloud server collects the state information of all edge nodes and constructs a state vector. ； In the formula, This is a set of CPU and memory utilization rates for each edge node. The set of available memory for each edge node. For network bandwidth collection, This is the set of task queue lengths for each edge node. This is a set of energy consumption states for edge nodes. This is a set of data block version identifiers locally cached by each edge node; Cloud servers are based on the current Table, output action : ； In the formula, Number the selected edge nodes. Number the selected lightweight model version; The table is a two-dimensional data table, with rows corresponding to... , column corresponding The value of each element In order to be in Execute The expected value of cumulative rewards; implement The reward function is calculated afterward: ； In the formula, For the task In the selected node Use the selected model version The estimated completion time, If The data transfer time required without holding the data blocks needed for the task. For the task At the node The energy consumption estimate for the above execution, To estimate the accuracy gain of a specific lightweight model version compared to the baseline model, , , , These are the weighting coefficients. As an indicator function, when the data required by the task is already available... The value is 1 if it is in the local cache, and 0 otherwise.

7. The high-resolution remote sensing image processing method based on superpixel cloud-edge collaboration according to claim 6, characterized in that, S3 includes S3.3, which uses the Q-learning algorithm for updating. value: ； In the formula, For learning rate, As a discount factor, For the next state All possible actions The largest value; S3 includes S3.4, establishing a three-level caching mechanism, including edge node local memory L1, edge node local solid-state storage L2, and cloud storage L3; based on Divide upstream and downstream tasks, and obtain the scheduling probability of each downstream task after the upstream task is completed by statistically analyzing the historical task execution sequence. When the current upstream task is about to be completed, load the data block with the highest scheduling probability from L3 or L2 into L1.

8. The high-resolution remote sensing image processing method based on superpixel cloud-edge collaboration according to claim 7, characterized in that, S4 includes S4.1, establishing a distributed storage pool based on L2 and L3. Each superpixel block in the distributed storage pool corresponds to a data container. The fields inside the container include data header, pixel data, intermediate results, and a set of operation identifiers. S4 includes S4.2, which uses distributed hash table technology to manage the data index, distributing metadata across edge nodes. The metadata is: ； In the formula, To store a list of edge node numbers for the current container replica. This represents the historical access frequency of the current container. For container data size, For checksums used for data integrity verification, This is a copy of the container header; S4 includes S4.3, where scheduling decisions follow the data affinity principle, which prioritizes assigning tasks to edge nodes that have cached the data containers required by the current task; During task execution, edge nodes calculate intermediate states as snapshots and write them into the intermediate result field of the data container of the current superpixel block.

9. The high-resolution remote sensing image processing method based on superpixel cloud-edge collaboration according to claim 8, characterized in that, S4 includes S4.4, periodically scanning the metadata of each data container, setting a frequency threshold, data above the frequency threshold is hot data, and data below or equal to the frequency threshold is cold data; For hot data, increase the number of replicas in the L2 cache of edge nodes, and store it in L1 when L2 is insufficient; For cold data, compress and archive it to reduce the number of copies; Superpixel block containers with high spatiotemporal correlation are stored contiguously in physical storage or merged into a single archive file. High spatiotemporal correlation is... middle Records the relationships between different time phases of the same land feature.

10. The high-resolution remote sensing image processing method based on superpixel cloud-edge collaboration according to claim 9, characterized in that, S4 includes S4.5, after all edge nodes have completed their respective superpixel block tasks, they write the processing results back to the corresponding data container; according to and Following the order of the superpixel blocks, the processing results of the data containers are read sequentially, and the built-in splicing operator of the data containers is called to splice the output results of all containers into the final result image.

Images