Remote sensing image clustering parallel processing method and system

Abstract

The invention discloses a remote sensing image clustering parallel processing method and system. The method comprises the following steps: performing space-time registration processing on an original multi-source data set; carrying out ground feature complexity identification on the to-be-processed multi-source image data set, and carrying out ground feature classification task priority marking to obtain a ground feature classification task set; collecting computing power characteristics of each node by using a heterogeneous computing cluster; respectively distributing each ground feature classification task in the ground feature classification task set to a target node by utilizing a task node matching model based on the computing power features; and outputting a target precision ground feature classification map based on the sub-region ground feature classification result output by each target node. According to the scheme of the invention, efficient and accurate remote sensing image ground object classification result output is realized.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image data processing, and specifically relates to a method and system for clustered parallel processing of remote sensing images. Background Technology

[0002] With the rapid development of remote sensing technology, the spatial and temporal resolutions of multi-source remote sensing equipment such as optical remote sensing, synthetic aperture radar, and interferometric synthetic aperture radar have been continuously improved. Remote sensing image data exhibits significant characteristics of being massive, multi-source, and heterogeneous. Statistics show that a single high-resolution optical satellite can generate tens of terabytes of image data per day, while SAR satellites, with their all-weather, all-day observation advantages, further exacerbate the contradiction between data acquisition and processing. As a core link connecting data and applications, the efficiency and accuracy of remote sensing image processing technology directly determine the reliability of downstream applications such as land cover classification, environmental monitoring, and land use planning.

[0003] Currently, remote sensing image processing has mainly gone through two stages: single-machine serial processing and distributed parallel processing. Early single-machine serial processing methods relied on the hardware performance of a single computer, completing tasks such as land cover classification through sequential calculations pixel-by-pixel and region-by-region. However, this method is limited by the computing power and storage bottlenecks of a single machine, often requiring hours or even days to process image data larger than GB, making it difficult to meet the demands of real-time applications. To overcome the limitations of single-machine performance, distributed parallel processing technology emerged, dividing image data into multiple sub-tasks and distributing them to different computing nodes to achieve parallel execution of tasks. Existing distributed processing systems are mostly based on homogeneous cluster architectures, i.e., computing clusters composed of CPU nodes of the same type. While this improves processing efficiency to some extent, it still has significant shortcomings when facing the heterogeneity of multi-source image data and the differences in land cover complexity. In multi-source image data processing, a unified preprocessing workflow is usually adopted, performing the same radiometric correction, geometric correction, and other operations on images of all regions. However, the complexity of different land cover regions varies significantly. For example, urban built-up areas contain multiple land cover types such as buildings, roads, and vegetation, while desert areas have a single land cover type. A uniform preprocessing strategy not only leads to insufficient processing accuracy in complex areas but also wastes computational resources in simple areas. Meanwhile, the spatiotemporal registration accuracy of multi-source images directly affects subsequent processing results. Related technologies often employ traditional feature point matching algorithms such as SIFT and ORB, but the imaging mechanisms of SAR images differ significantly from optical images, easily generating a large number of mismatched feature points, increasing registration errors, and consequently affecting the accuracy of land cover classification. Regarding computational power allocation and task scheduling, while existing heterogeneous computing clusters have introduced various computing nodes such as CPUs, GPUs, and FPGAs, task allocation strategies are mostly based on empirical rules or simple load balancing algorithms. For example, some methods randomly assign all land cover classification tasks to GPU nodes without considering the matching degree between task complexity and node computational power characteristics; other methods only schedule tasks based on the current node load, ignoring the differences in task priority and processing accuracy requirements. This coarse-grained task allocation method easily leads to high-complexity tasks being assigned to low-computing-power nodes, or low-priority tasks occupying high-performance node resources, resulting in low overall cluster computational power utilization and difficulty in further improving processing efficiency.

[0004] Furthermore, during parallel processing, subtasks on adjacent computing nodes exhibit data dependencies in edge regions. Related technologies often address this issue through periodic global data synchronization; however, excessively high synchronization frequencies increase inter-node communication overhead, while insufficient synchronization frequencies lead to inconsistent classification results in edge regions, resulting in problems such as "seams." Simultaneously, the selection of land cover classification models lacks sufficient adaptability to the task. Existing methods often employ a single model to handle all subtasks without selecting a matching model based on node computing power and task complexity. For example, deploying a complex model on an FPGA-accelerated node may result in inference failure due to resource constraints, while deploying a lightweight model on a GPU node would waste computing resources.

[0005] In summary, remote sensing image processing technology still has many shortcomings in areas such as differentiated processing of multi-source data, heterogeneous cluster task scheduling, and collaborative computing between nodes, making it difficult to simultaneously meet the efficiency and accuracy requirements of processing massive amounts of multi-source remote sensing images. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method and system for parallel processing of remote sensing images in clusters. This addresses the technical problems that remote sensing image processing technology still has many deficiencies in areas such as differentiated processing of multi-source data, scheduling of heterogeneous cluster tasks, and collaborative computing between nodes, making it difficult to simultaneously meet the efficiency and accuracy requirements of processing massive multi-source remote sensing images.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution.

[0008] This invention first discloses a method for clustered parallel processing of remote sensing images, which includes the following steps: Simultaneously acquire optical remote sensing image data, SAR image data and digital elevation model data of the target area to construct the original multi-source dataset; The original multi-source data set is spatiotemporally registered using a feature point matching algorithm to obtain a multi-source image data set to be processed. Based on a deep learning semantic segmentation model, the set of multi-source image data to be processed is used to identify the complexity of ground features. Based on the results of the identification of ground feature complexity, the set of multi-source image data to be processed is divided into a first image region set and a second image region set. Preprocessing operations are performed on the first image region set and the second image region set respectively to obtain the first image region set to be processed and the second image region set to be processed. The first set of image regions to be processed and the second set of image regions to be processed are marked with the priority of land cover classification tasks to obtain the set of land cover classification tasks. The computing power characteristics of each node are collected using a heterogeneous computing cluster; Based on the computing power characteristics, the task node matching model is used to assign each land cover classification task in the land cover classification task set to the target node; The target nodes are used to perform parallel processing of the land cover classification task, output the land cover classification results of the sub-regions corresponding to the land cover classification task, and output the target accuracy land cover classification map based on the sub-region land cover classification results output by each target node.

[0009] The present invention further includes the following preferred embodiments: The simultaneous acquisition of optical remote sensing image data, SAR image data, and digital elevation model data of the target area to construct the original multi-source dataset further includes: Collect target area image data returned by optical remote sensing satellites, perform cloud-free screening on the target area image data, and determine the target time period and corresponding optical remote sensing image data. Collect SAR image data of the target area returned by SAR satellites within the target time period; Using InSAR technology to obtain digital elevation model data of the target area; The optical remote sensing image data, SAR image data, and digital elevation model data are aligned according to spatial coordinates to construct the original multi-source data set.

[0010] The step of performing spatiotemporal registration processing on the original multi-source data set using a feature point matching algorithm to obtain a multi-source image data set to be processed further includes: The SIFT algorithm was used to extract feature points from optical remote sensing images and SAR images in the original multi-source dataset, resulting in optical remote sensing image feature points and SAR image feature points. The K-nearest neighbor algorithm is used to perform preliminary matching of feature points in optical remote sensing images and SAR images to obtain the first set of feature point pairs. The RANSAC random sampling consensus algorithm is used to remove mismatched feature point pairs from the first feature point pair set to obtain the second feature point pair set. Spatiotemporal registration is performed on the set based on the second feature point to obtain the set of multi-source image data to be processed.

[0011] The deep learning semantic segmentation model adds a channel attention module to the model encoding end of the BiSeNet V2 model; the deep learning semantic segmentation model performs ground feature complexity identification on the multi-source image dataset to be processed, and divides the multi-source image dataset to be processed into a first image region set and a second image region set based on the ground feature complexity identification result, further including: The multi-source image data set to be processed is normalized. The normalized set of multi-source image data to be processed is input into the deep learning semantic segmentation model, and forward inference calculation is performed. The probability distribution of land cover categories for each pixel is obtained by using the Softmax function of the model output layer, and the category corresponding to the maximum probability is taken as the land cover category prediction result. The number of land cover categories in each image region is counted, and regions with 5 or more land cover categories are divided into the first image region set. Areas with fewer than 5 land cover categories are divided into a second image region set, and a region division mask map is generated.

[0012] The step of performing preprocessing operations on the first image region set and the second image region set respectively to obtain the first image region set to be processed and the second image region set to be processed further includes: Perform radiometric normalization and Gaussian filtering on the first image region set to obtain the enhanced preprocessed first image region set. A quadratic polynomial correction model is used to perform geometric correction on the first image region set after enhancement preprocessing to obtain the first image region set to be processed. A quadratic polynomial correction model is used to perform geometric correction on the second image region set to obtain the second image region set to be processed.

[0013] The step of prioritizing the first set of image regions to be processed and the second set of image regions to be processed to obtain a set of land cover classification tasks includes: Extract the number of land cover categories, image resolution data, and texture entropy data from the first and second image regions to be processed. The number of land cover categories, image resolution data, and texture entropy data of the first and second image regions to be processed are input into the classification task complexity evaluation model to obtain a set of task complexity quantification values. The set of task complexity quantification values ​​includes the task complexity quantification values ​​corresponding to each image region in the first and second image regions to be processed. Based on the task complexity threshold and the task complexity quantification value set, the first image region set and the second image region set to be processed are marked with the priority of the land cover classification task, so as to obtain the land cover classification task set. Then, a task attribute table containing task ID, complexity quantification value, priority level and accuracy time consumption constraint is generated for each land cover classification task in the land cover classification task set.

[0014] The heterogeneous computing cluster includes CPU nodes, GPU nodes, and FPGA acceleration nodes. The computing power characteristics include CPU floating-point operation characteristics, GPU tensor computation efficiency characteristics, and FPGA parallel throughput characteristics. The process of collecting the computing power characteristics of each node using the heterogeneous computing cluster includes: Utilize CPU nodes to collect floating-point arithmetic capabilities and memory usage. Utilize GPU nodes to acquire tensor computation efficiency and GPU memory utilization; Utilize FPGA nodes to collect parallel throughput and resource utilization; The floating-point computing power, memory usage, tensor computation efficiency, video memory usage, parallel throughput, and resource utilization are aggregated to the cluster scheduling center to construct a real-time computing power feature matrix.

[0015] The target node is one of a CPU node, a GPU node, and an FPGA acceleration node. The task node matching model is established using an improved Hungarian algorithm. The process of using the task node matching model to assign each land cover classification task in the land cover classification task set to the target node based on the computing power characteristics includes: The priority level and complexity quantification value of the land cover classification task are converted into the cost matrix penalty term parameters of the task node matching model; The real-time computing power feature matrix and the task parameters in the task attribute table are input into the task node matching model to generate the initial matching cost matrix of the task node. The minimum cost matching solution of the initial matching cost matrix is ​​obtained by improving the Hungarian algorithm; Based on the minimum cost matching solution, the task allocation scheme is output, and each land feature classification task is assigned to the corresponding target node, and a task allocation result table is generated.

[0016] The method of using target nodes to perform parallel processing of the land cover classification task, outputting the land cover classification results of the sub-regions corresponding to the land cover classification task, and outputting a land cover classification map with target accuracy based on the sub-region land cover classification results output by each target node, includes: The target node loads the corresponding level of land cover classification model. The FPGA acceleration node loads the first level land cover classification model, the GPU node loads the second level land cover classification model, and the CPU node loads the third level land cover classification model. The accuracy of the first level land cover classification model is higher than that of the second level land cover classification model, and the accuracy of the second level land cover classification model is higher than that of the third level land cover classification model. Utilize data sharing channels between adjacent target nodes to exchange geographic feature data in edge areas; The land cover classification calculation is performed in parallel at each target node, and the land cover feature parameters of the edge area are synchronized according to a preset period during the calculation process. Output the sub-region land cover classification results corresponding to the land cover classification task and save them in the standard remote sensing image format; A weighted fusion algorithm is used to stitch together the land cover classification results of each sub-region to obtain a preliminary classification map. The weight of the sub-result corresponding to the first image region set is higher than the weight of the sub-result corresponding to the second image region set. The accuracy of land cover classification and the Kappa coefficient of the preliminary classification map are calculated based on the confusion matrix, and the edge sharpness of the preliminary classification map is calculated using the gradient operator. A comprehensive evaluation score is obtained based on the accuracy of land cover classification, Kappa coefficient, and edge sharpness. If the comprehensive evaluation score is greater than or equal to the preset accuracy threshold, the preliminary classification map is output as the target accuracy land cover classification map.

[0017] This invention also discloses a remote sensing image clustering parallel processing system utilizing the aforementioned remote sensing image clustering parallel processing method, comprising: The acquisition module is used to simultaneously acquire optical remote sensing image data, SAR image data and digital elevation model data of the target area to construct the original multi-source data set; The matching module is used to perform spatiotemporal registration processing on the original multi-source data set using a feature point matching algorithm to obtain a multi-source image data set to be processed. The identification module is used to identify the ground feature complexity of the multi-source image data set to be processed based on a deep learning semantic segmentation model, and to divide the multi-source image data set to be processed into a first image region set and a second image region set based on the ground feature complexity identification result. The preprocessing module is used to perform preprocessing operations on the first image region set and the second image region set respectively to obtain the first image region set to be processed and the second image region set to be processed. The labeling module is used to label the first set of image regions to be processed and the second set of image regions to be processed with the priority of land cover classification tasks, so as to obtain a set of land cover classification tasks. The computing power module is used to collect the computing power characteristics of each node in a heterogeneous computing cluster. The allocation module is used to use the task node matching model to allocate each land cover classification task in the land cover classification task set to the target node based on the computing power characteristics. The processing module is used to perform parallel processing of the land cover classification task using the target nodes, output the land cover classification results of the sub-regions corresponding to the land cover classification task, and output the target accuracy land cover classification map based on the sub-region land cover classification results output by each target node.

[0018] Accordingly, this application also discloses a terminal, including a processor and a storage medium; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the aforementioned remote sensing image clustering parallel processing method.

[0019] Accordingly, this application also discloses a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the aforementioned remote sensing image clustering parallel processing method.

[0020] The beneficial effects of this invention are as follows: Compared with the prior art, this invention provides a method and system for clustered parallel processing of remote sensing images. Firstly, by simultaneously acquiring optical remote sensing image data, SAR image data, and digital elevation model data of the target area and constructing an original multi-source data set, it achieves complete acquisition and integration of multi-dimensional remote sensing information. Optical remote sensing images provide rich spectral information of ground features, SAR images have the advantage of all-day, all-weather observation, and digital elevation model data reflects terrain undulation characteristics. The simultaneous acquisition and aggregation of these three types of data avoids spatiotemporal deviations between different time periods. Secondly, by using a feature point matching algorithm to perform spatiotemporal registration processing on the original multi-source data set, spatial coordinate differences and temporal reference deviations between different types of remote sensing images are eliminated. Through feature point extraction, matching, and mismatch removal, it is ensured that multi-source data such as optical remote sensing images and SAR images are aligned within the same spatiotemporal framework. Then, by using a deep learning semantic segmentation model to identify the ground feature complexity of the processed multi-source image data set, a first image region set and a second image region set are divided accordingly, achieving refined differentiation of ground feature distribution characteristics. Deep learning semantic segmentation models accurately identify different land cover types and their distribution ranges. Complexity is quantified by statistically analyzing the number of land cover categories, enabling differentiated processing strategies for regions of varying complexity. Furthermore, targeted image optimization is achieved by performing preprocessing operations on the first and second image region sets. For the more complex first image region set, radiometric normalization and Gaussian filtering are performed sequentially to eliminate illumination differences and enhance texture details; for the less complex second image region set, only geometric correction is performed, simplifying the process while maintaining processing quality. This differentiated preprocessing approach improves image quality in complex areas while increasing processing efficiency in simpler areas. Subsequently, land cover classification task priorities are assigned to the first and second image region sets to be processed, generating a land cover classification task set and clarifying the priority order and constraints for processing each region. Multi-dimensional data such as the number of land cover categories and image resolution are extracted to calculate task complexity, and priority is assigned based on preset thresholds. A task attribute table containing task IDs and complexity quantification values ​​is generated, allowing subsequent task allocation to be carried out based on clear priorities and task characteristics. Furthermore, by leveraging the heterogeneous computing cluster to collect the computing power characteristics of each node, a comprehensive understanding of the cluster's real-time computing power distribution is achieved. Computing power acquisition agents are deployed on different nodes such as CPUs, GPUs, and FPGAs to collect specific computing power indicators, including floating-point arithmetic capabilities and tensor computation efficiency. These are then aggregated to construct a real-time computing power feature matrix, dynamically reflecting the load status and computing advantages of each node. Finally, a task node matching model is used to allocate the set of land cover classification tasks to target nodes based on computing power characteristics, achieving optimal task-node matching.Task priority and complexity are converted into cost matrix penalty terms. An initial matching cost matrix is ​​generated by combining this with a real-time computing power feature matrix. An improved Hungarian algorithm is used to find the minimum cost matching solution, allowing high-priority, high-complexity tasks to be assigned to nodes with stronger computing power, and low-priority, low-complexity tasks to suitable nodes, thus improving the overall computing power utilization and task processing efficiency of the cluster. Finally, by utilizing target nodes to perform parallel processing of land cover classification tasks and outputting sub-region land cover classification results, a target-precision land cover classification map is output based on the sub-results, achieving efficient processing and high-precision output of massive image data. Parallel computation of each target node significantly shortens the overall processing time, high-speed data sharing and feature synchronization between adjacent nodes ensure classification consistency in edge areas, and weighted fusion and comprehensive evaluation score assessment ensure that the accuracy of the stitched classification map meets the standards, ultimately achieving efficient and accurate remote sensing image land cover classification results. In summary, the remote sensing image clustering parallel processing method provided in this embodiment effectively solves many shortcomings of remote sensing image processing technology in terms of multi-source data differentiation processing, heterogeneous cluster task scheduling, and inter-node collaborative computing, making it difficult to simultaneously meet the efficiency and accuracy requirements of massive multi-source remote sensing image processing. Attached Figure Description

[0021] Figure 1 This is a flowchart of the remote sensing image clustering parallel processing method in this invention.

[0022] Figure 2 This is a schematic diagram of the structure of the remote sensing image clustered parallel processing system in this invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0024] The embodiments described in this application are merely some, not all, embodiments of the present invention. Based on the spirit of the present invention, other embodiments obtained by those skilled in the art without inventive effort are all within the protection scope of the present invention.

[0025] To address the shortcomings of existing technologies, this invention proposes a method and system for clustered parallel processing of remote sensing images. First, it simultaneously acquires optical remote sensing image data, SAR image data, and digital elevation model (DEM) data of the target area and constructs an original multi-source dataset, achieving complete acquisition and integration of multi-dimensional remote sensing information. Optical remote sensing images provide rich spectral information of ground features, SAR images offer all-weather, all-day observation advantages, and DEM data reflects terrain undulation characteristics. The simultaneous acquisition and aggregation of these three types of data avoids spatiotemporal deviations between different time periods. Second, a feature point matching algorithm is used to perform spatiotemporal registration processing on the original multi-source dataset, eliminating spatial coordinate differences and temporal reference deviations between different types of remote sensing images. Through feature point extraction, matching, and mismatch removal, it ensures that multi-source data such as optical remote sensing images and SAR images are aligned within the same spatiotemporal framework. Then, a deep learning-based semantic segmentation model is used to identify the ground feature complexity of the processed multi-source image dataset, and based on this, a first image region set and a second image region set are divided, achieving refined differentiation of ground feature distribution characteristics. Deep learning semantic segmentation models accurately identify different land cover types and their distribution ranges. Complexity is quantified by statistically analyzing the number of land cover categories, enabling differentiated processing strategies for regions of varying complexity. Furthermore, targeted image optimization is achieved by performing preprocessing operations on the first and second image region sets. For the more complex first image region set, radiometric normalization and Gaussian filtering are performed sequentially to eliminate illumination differences and enhance texture details; for the less complex second image region set, only geometric correction is performed, simplifying the process while maintaining processing quality. This differentiated preprocessing approach improves image quality in complex areas while increasing processing efficiency in simpler areas. Subsequently, land cover classification task priorities are assigned to the first and second image region sets to be processed, generating a land cover classification task set and clarifying the priority order and constraints for processing each region. Multi-dimensional data such as the number of land cover categories and image resolution are extracted to calculate task complexity, and priority is assigned based on preset thresholds. A task attribute table containing task IDs and complexity quantification values ​​is generated, allowing subsequent task allocation to be carried out based on clear priorities and task characteristics. Furthermore, by leveraging the heterogeneous computing cluster to collect the computing power characteristics of each node, a comprehensive understanding of the cluster's real-time computing power distribution is achieved. Computing power acquisition agents are deployed on different nodes such as CPUs, GPUs, and FPGAs to collect specific computing power indicators, including floating-point arithmetic capabilities and tensor computation efficiency. These are then aggregated to construct a real-time computing power feature matrix, dynamically reflecting the load status and computing advantages of each node. Finally, a task node matching model is used to allocate the set of land cover classification tasks to target nodes based on computing power characteristics, achieving optimal task-node matching.Task priority and complexity are converted into cost matrix penalty terms. An initial matching cost matrix is ​​generated by combining this with a real-time computing power feature matrix. An improved Hungarian algorithm is used to find the minimum cost matching solution, allowing high-priority, high-complexity tasks to be assigned to nodes with stronger computing power, and low-priority, low-complexity tasks to suitable nodes, thus improving the overall computing power utilization and task processing efficiency of the cluster. Finally, by utilizing target nodes to perform parallel processing of land cover classification tasks and outputting sub-region land cover classification results, a target-precision land cover classification map is output based on the sub-results, achieving efficient processing and high-precision output of massive image data. Parallel computation of each target node significantly shortens the overall processing time, high-speed data sharing and feature synchronization between adjacent nodes ensure classification consistency in edge areas, and weighted fusion and comprehensive evaluation score assessment ensure that the accuracy of the stitched classification map meets the standards, ultimately achieving efficient and accurate remote sensing image land cover classification results.

[0026] See Figure 1 As shown, the remote sensing image clustering parallel processing method disclosed in this invention includes the following steps: Step S101: Simultaneously collect optical remote sensing image data, SAR image data and digital elevation model data of the target area to construct the original multi-source data set.

[0027] Specifically, step S101 includes: Step 1011: Collect target area image data returned by optical remote sensing satellite, perform cloud-free screening on the target area image data, and determine the target time period and corresponding optical remote sensing image data.

[0028] Furthermore, during data acquisition, the target area image data transmitted from optical remote sensing satellites is received via a satellite data receiving terminal, containing spectral information of different bands of ground object reflection. After reception, an image quality assessment algorithm is used to screen the acquired image data for cloud-free conditions. By analyzing the grayscale distribution, texture features, and inter-band correlation of the images, image data with cloud cover or cloud shadow interference is identified and removed. Based on the capture time of the retained clear image data after screening, a continuous target time period with acceptable image quality is determined. All optical remote sensing image data that meet the requirements within this time period are summarized to form the optical remote sensing image dataset corresponding to the target time period.

[0029] Step 1012: Collect SAR image data of the target area returned by SAR satellite within the target time period.

[0030] Furthermore, SAR satellites observe target areas by actively transmitting microwave signals and receiving ground echo signals. During the acquisition process, the SAR satellite's data reception parameters are set according to the determined target time period to ensure a perfect match between the reception and target time periods. The SAR data receiving system receives SAR image data of the target area transmitted from the satellite. This data records ground feature information in the form of radar echo intensity and is unaffected by weather conditions. After reception, the SAR image data undergoes format conversion and preliminary verification to check for data integrity and transmission errors. Image files with missing data or transmission anomalies are discarded. The verified SAR image data is then organized in chronological order of capture time to form a SAR image dataset of the target area within the target time period.

[0031] Step 1013: Use InSAR technology to obtain digital elevation model data of the target area.

[0032] Furthermore, InSAR technology acquires topographic elevation information by interferometric processing of two or more SAR images of the same target area. The process begins by selecting multiple SAR images covering the target area within the target time period. These images must meet certain spatiotemporal baseline requirements to ensure effective interferometry. Preprocessing of the selected SAR images includes image registration, flattening, and filtering to eliminate systematic errors and noise interference between images. An interferometric phase map is obtained through interferometric phase calculation, and then phase unwrapping is performed to restore the wrapped phase values ​​to the true phase difference. Combining the orbital parameters of the SAR satellite and radar wavelength information, topographic elevation data of the target area is calculated through geometric relationships, ultimately generating a digital elevation model (DEM) of the target area.

[0033] Step 1014: Align the optical remote sensing image data, SAR image data, and digital elevation model data according to spatial coordinates to construct the original multi-source data set.

[0034] Furthermore, the spatial coordinate information of the three types of data is extracted. Optical remote sensing image data and SAR image data contain the geographic coordinate parameters of pixels, while digital elevation model data contains the latitude, longitude, and elevation information of grid points. Using the spatial coordinate system of the optical remote sensing image data as a benchmark, a coordinate transformation algorithm is employed to uniformly transform the spatial coordinates of the SAR image data and digital elevation model data, ensuring that all three types of data are within the same geographic coordinate framework. After the transformation, a spatial coordinate matching algorithm is used to align and associate information corresponding to the same geographic location in the three types of data, so that each geographic coordinate point corresponds to optical remote sensing spectral information, SAR echo information, and elevation information. The aligned three types of data are then integrated and stored to construct an original multi-source dataset containing multi-source remote sensing information.

[0035] Step S102: Use the feature point matching algorithm to perform spatiotemporal registration processing on the original multi-source data set to obtain the multi-source image data set to be processed.

[0036] Specifically, step S102 includes: Step 1021: Use the SIFT algorithm to extract feature points from the optical remote sensing image and SAR image in the original multi-source dataset, respectively, to obtain the optical remote sensing image feature points and SAR image feature points.

[0037] Furthermore, the SIFT algorithm, or Scale Invariant Feature Transform algorithm, can extract ground feature points with rotation and scaling invariance across different scale spaces. When processing optical remote sensing images, a scale space is first constructed, and extreme points are detected using the Difference of Gaussian operator. These extreme points are then precisely located and filtered. The principal orientation of the extreme points is then calculated to generate feature descriptors containing location, scale, and orientation information, forming the optical remote sensing image feature points. When processing SAR images, due to speckle noise, denoising preprocessing is performed first, followed by the same SIFT algorithm process to extract feature points, generating SAR image feature points. This ensures that the feature points of both types of images have a unified description format for subsequent matching.

[0038] Step 1022: Perform preliminary feature point matching between optical remote sensing image feature points and SAR image feature points using the K-nearest neighbor algorithm to obtain the first set of feature point pairs.

[0039] Furthermore, the K-nearest neighbor algorithm is a feature similarity-based matching algorithm. Its core is to find the K candidate feature points in the SAR image whose feature descriptors are most similar to each optical remote sensing image feature point. In operation, the descriptors of the optical remote sensing image feature points and the SAR image feature points are first converted into vectors of the same dimension. The Euclidean distance between the two sets of vectors is calculated; the smaller the distance, the higher the feature similarity. For each optical remote sensing image feature point, the K SAR image feature points with the smallest Euclidean distance are selected as candidate matching points. Then, a nearest neighbor to second nearest neighbor distance ratio method is used for further filtering. When the ratio of the nearest neighbor distance to the second nearest neighbor distance is less than a preset threshold, the nearest neighbor matching pair is retained. All matching pairs that meet the conditions are summarized to form the first feature point pair set.

[0040] Step 1023: Use the RANSAC random sampling consensus algorithm to remove mismatched feature point pairs from the first feature point pair set to obtain the second feature point pair set.

[0041] Furthermore, the RANSAC algorithm separates inliers and outliers through random sampling and model validation. Inliers are correctly matched feature point pairs, while outliers are mismatched feature point pairs. First, assuming a spatial transformation model, a small number of feature point pairs are randomly sampled from the first feature point pair set. The transformation model parameters are calculated based on these samples. Then, the number of feature point pairs in the first feature point pair set that satisfy the transformation model is counted, i.e., the number of inliers. This process of random sampling, model calculation, and inlier counting is repeated. When the number of inliers obtained in a given iteration is the highest and reaches a preset proportion, the transformation model parameters for that iteration are determined to be the optimal model parameters. All feature point pairs that satisfy the optimal model parameters are retained, while mismatched feature point pairs that do not satisfy the optimal parameters are removed, ultimately forming the second feature point pair set.

[0042] Step 1024: Perform spatiotemporal registration on the set based on the second feature points to obtain the set of multi-source image data to be processed.

[0043] Furthermore, based on the set of second feature point pairs, the spatial transformation matrix between the optical remote sensing image and the SAR image is calculated. Commonly used spatial transformation models include affine transformation models or projection transformation models. The coordinates of the second feature point pairs are fitted using the least squares method to obtain the values ​​of each element of the transformation matrix. Substituting the coordinates of each pixel in the SAR image into the transformation matrix, its corresponding coordinates in the optical remote sensing image coordinate system are calculated, achieving spatial alignment between the SAR and optical remote sensing images. Simultaneously, the acquisition timestamps of both types of images are used to uniformly label the time information of the image data, ensuring consistency in the time reference. The spatially aligned and time-referenced optical remote sensing images, SAR images, and digital elevation model data are integrated to construct the multi-source image dataset to be processed.

[0044] Step S103: Based on the deep learning semantic segmentation model, the ground feature complexity of the multi-source image data set to be processed is identified, and based on the ground feature complexity identification results, the multi-source image data set to be processed is divided into a first image region set and a second image region set.

[0045] Specifically, the deep learning semantic segmentation model adds a channel attention module to the model encoding end of the BiSeNet V2 model. Step S103 above includes: Step 1031: Normalize the multi-source image data set to be processed.

[0046] Furthermore, the multi-source image dataset to be processed includes optical remote sensing images, SAR images, and digital elevation model (DEM) data. The numerical ranges of different data types vary, and normalization is used to map these data to the same numerical range. During the process, the mean and standard deviation of each data type in the multi-source image dataset are first calculated. For optical remote sensing image data, a band mean normalization method is used, subtracting the band mean from the pixel value of each band and then dividing by the standard deviation. For SAR image data, a maximum-minimum normalization method is used to map the pixel values ​​to a preset range. The same maximum-minimum normalization method is used for DEM data to eliminate the influence of numerical differences in elevation values ​​across different terrain types. After normalization, it is ensured that all data types participate in subsequent model inference calculations at the same numerical scale.

[0047] Step 1032: Input the normalized multi-source image data set to be processed into the deep learning semantic segmentation model and perform forward inference calculation.

[0048] Furthermore, the improved BiSeNet V2 model incorporates a dual-channel feature extraction network: a spatial path and a semantic path. The spatial path extracts high-resolution spatial detail features, while the semantic path extracts deep semantic features through downsampling. A channel attention module added to the encoder is located in the feature output stage of the semantic path. First, global average pooling is performed on the feature map output by the semantic path to obtain feature descriptors for each channel. Then, attention weights for each channel are calculated using fully connected layers and activation functions. These attention weights are multiplied channel-by-channel with the original feature map to enhance key feature channels. Normalized multi-source data is concatenated in band order to form multi-channel input data. After being input into the improved model, the data sequentially passes through the spatial path, semantic path, and channel attention modules to complete feature extraction and fusion, outputting the feature vector of each pixel's land cover category.

[0049] Step 1033: Obtain the probability distribution of land cover categories for each pixel through the Softmax function of the model output layer, and take the category corresponding to the maximum probability as the land cover category prediction result.

[0050] Furthermore, the Softmax function is a normalized exponential function used to convert the feature vector output by the model into probability values ​​for various land cover categories, with the sum of all category probabilities equal to 1. In this operation, the feature vector output from the last layer of the improved BiSeNet V2 model is input into the Softmax function to calculate the probability of each pixel belonging to a specific land cover category, including predefined categories such as buildings, roads, vegetation, water bodies, and bare land. The probability distribution of each pixel is analyzed, and the land cover category with the highest probability value is selected as the final land cover category prediction result for that pixel, thus obtaining a pixel-level land cover category labeling map for the entire image. Each pixel in the labeling map corresponds to a unique land cover category identifier.

[0051] Step 1034: Count the number of land cover categories in each image region, and divide the regions with more than or equal to 5 land cover categories into the first image region set.

[0052] Furthermore, the multi-source image dataset to be processed is first divided into regions. A sliding window or regular grid method is used to segment the image into multiple image regions of uniform size, each containing a certain number of pixels. The pixel-level land cover category prediction results within each image region are statistically analyzed. The land cover category identifiers of all pixels within the region are traversed, and the number of different land cover categories appearing within that region is recorded. The statistically obtained number of land cover categories is compared with a preset complexity classification standard. When the number of land cover categories in a certain image region reaches or exceeds 5, the region is determined to be a region with high land cover complexity. All image regions that meet this condition are aggregated to form the first image region set.

[0053] Step 1035: Divide the areas with fewer than 5 land cover categories into the second image region set and generate a region division mask map.

[0054] Furthermore, the number of land cover categories obtained in step 1034 is further processed. When the number of land cover categories in a certain image area is less than 5, the area is determined to be an area with low land cover complexity. All such image areas are then aggregated to form a second image area set. The region division mask is a binary or indexed image used to visually identify the sets to which different image areas belong. Different pixel values ​​or colors are used in the mask to identify the areas corresponding to the first and second image area sets. For example, a first preset value is used to identify the areas of the first image area set, and a second preset value is used to identify the areas of the second image area set. After generating the mask, its spatial coordinates are kept consistent with the original image data, facilitating subsequent preprocessing and task allocation steps.

[0055] Step S104: Perform preprocessing operations on the first image region set and the second image region set respectively to obtain the first image region set to be processed and the second image region set to be processed.

[0056] Specifically, step S104 includes: Step 1041: Perform radiometric normalization and Gaussian filtering on the first image region set to obtain the enhanced preprocessed first image region set.

[0057] Furthermore, radiometric normalization is used to eliminate the problem of uneven radiometric brightness in images caused by factors such as lighting conditions and differences in sensor response. During the operation, a representative standard region is first selected from the first image region set, and the mean and variance of radiometric values ​​for each band within the standard region are calculated. Then, the pixel values ​​of each band in each image region are linearly transformed according to the radiometric statistics of the standard region to ensure that the radiometric characteristics of all image regions are consistent. Gaussian filtering is used to smooth the image, suppress noise, and preserve ground details. By constructing a Gaussian filter, with each pixel as the center, the weights of neighboring pixels are calculated according to a Gaussian function. The weighted sum of the neighboring pixel values ​​and their corresponding weights is used as the new value of the center pixel. This operation is performed sequentially on all pixels in the first image region set to complete the enhancement preprocessing.

[0058] Step 1042: Use a quadratic polynomial correction model to perform geometric correction on the first image region set after enhancement preprocessing to obtain the first image region set to be processed.

[0059] Furthermore, geometric correction is the process of eliminating geometric distortions in images caused by factors such as sensor pose changes and terrain undulations. A quadratic polynomial correction model achieves correction by establishing a quadratic function relationship between image pixel coordinates and geographic coordinates. In the operation, a certain number of ground control points are first selected from the first image region set after enhancement preprocessing. These ground control points must be clearly identifiable on the image and have accurate geographic coordinates. The image coordinates and geographic coordinates of the ground control points are substituted into the quadratic polynomial model, and the coefficient parameters of the model are solved using the least squares method to obtain the conversion formula between pixel coordinates and geographic coordinates. This conversion formula is then used to transform the coordinates of each pixel in the first image region set, resampling to generate image regions with eliminated geometric distortion, and summarizing these to form the first image region set to be processed.

[0060] Step 1043: Use a quadratic polynomial correction model to perform geometric correction on the second image region set to obtain the second image region set to be processed.

[0061] Furthermore, ground control points are first selected from the second image region set, with the selection criteria being uniform distribution, distinct features, and accurate geographic coordinates to ensure the accuracy of the calibration model. The image coordinates and geographic coordinates of the ground control points are input into a quadratic polynomial calibration model to solve for the model coefficient parameters and establish the transformation relationship between pixel coordinates and geographic coordinates. Based on this transformation relationship, pixel coordinate transformation and resampling are performed on each image region in the second image region set to eliminate geometric distortion, ensuring that the spatial positioning accuracy of the second image region set is consistent with that of the first image region set. All calibrated image regions are then combined to form the second image region set to be processed.

[0062] Step S105: Mark the land cover classification task priority of the first image region set to be processed and the second image region set to be processed to obtain the land cover classification task set.

[0063] Specifically, step S105 includes: Step 1051: Extract the number of land cover categories, image resolution data, and texture entropy data from the first image region set and the second image region set to be processed.

[0064] Furthermore, the number of land cover categories refers to the number of different land cover categories contained in each image region. This is achieved by directly counting the number of land cover categories corresponding to each image region by calling the land cover category prediction results and region division information generated in step S103. Image resolution data refers to the number of pixels per unit length in the image. Resolution parameters are read from the metadata information of the image region to be processed. Texture entropy data is an indicator reflecting the complexity of image texture. It is calculated by calculating the gray-level co-occurrence matrix of the image region, and then solving for the texture entropy value based on the gray-level co-occurrence matrix. The gray-level co-occurrence matrix describes the distribution pattern of pixels with different gray values ​​in the image at specific directions and distances; a larger texture entropy value indicates a more complex texture. The extracted three types of data are associated and stored according to image regions, forming multi-dimensional feature data groups corresponding to each region.

[0065] Step 1052: Input the land cover category quantity data, image resolution data, and texture entropy data of the first and second image regions to be processed into the classification task complexity evaluation model to obtain a task complexity quantification value set. The task complexity quantification value set contains the task complexity quantification value corresponding to each image region in the first and second image regions to be processed.

[0066] Furthermore, the task complexity quantification set includes the task complexity quantification values ​​corresponding to each image region in the first and second image region sets to be processed. The classification task complexity evaluation model is a regression model trained by a machine learning algorithm, used to map multi-dimensional feature data into a single complexity quantification value. During operation, the extracted land cover category quantity data, image resolution data, and texture entropy data are first standardized to eliminate the dimensional differences between different data dimensions. The standardized feature data is then input into the trained evaluation model, which calculates the weighted average of each feature using preset weight coefficients, outputting the task complexity quantification value corresponding to each image region. The complexity quantification values ​​of all image regions are then organized according to region identifiers to form the task complexity quantification value set.

[0067] Step 1053: Based on the task complexity threshold and the task complexity quantification value set, mark the land cover classification task priority of the first image region set and the second image region set to be processed to obtain the land cover classification task set, and generate a task attribute table for each land cover classification task in the land cover classification task set, which includes task ID, complexity quantification value, priority level and accuracy time constraint.

[0068] Furthermore, the task complexity threshold is a pre-defined critical value used to classify priority levels, determined based on the processing requirements of the actual application scenario and the cluster's computing power. Each quantized value in the task complexity quantification set is compared with the threshold; image areas with quantized values ​​above the threshold are marked as high-priority tasks, and those below are marked as low-priority tasks. If multiple thresholds exist, more priority levels can be defined. A unique task ID is assigned to each marked land cover classification task, and a task attribute table is generated by combining the corresponding complexity quantification value, priority level, and preset accuracy and time constraints. All marked tasks are integrated to form a land cover classification task set, and the task attribute table is used along with the set in the subsequent task allocation process.

[0069] Step S106: Collect the computing power characteristics of each node using a heterogeneous computing cluster.

[0070] Specifically, the heterogeneous computing cluster includes CPU nodes, GPU nodes, and FPGA acceleration nodes. Its computing power characteristics include CPU floating-point operation characteristics, GPU tensor computation efficiency characteristics, and FPGA parallel throughput characteristics. Step S106 includes: Step 1061: Use the CPU node to collect floating-point arithmetic capabilities and memory usage.

[0071] Furthermore, floating-point arithmetic capability is an indicator of the speed at which a CPU node performs floating-point operations. During data collection, this is determined by running a standard floating-point arithmetic test program on the CPU node and counting the number of floating-point operations completed per unit time. Memory utilization is an indicator reflecting the memory usage of the CPU node. It is obtained by reading the total physical memory capacity and the used capacity of the node in real time through the memory management interface provided by the operating system, and calculating the ratio of used capacity to total capacity. A fixed sampling period is set during the data collection process, and floating-point arithmetic capability and memory utilization data are continuously collected periodically to ensure data real-time performance and continuity. The collection results are temporarily stored in the local cache of the CPU node.

[0072] Step 1062: Utilize GPU nodes to collect tensor computation efficiency and video memory usage.

[0073] Furthermore, tensor computation efficiency is a core performance indicator for GPU nodes performing tensor operations. Tensor operations are widely used in deep learning and other scenarios. During data collection, a GPU tensor computation test suite is run, a standard tensor computation task is input, and the number of tensor operations completed per unit time is counted to obtain the tensor computation efficiency. Memory utilization is the percentage of GPU node memory resources used. This is obtained by using the memory query interface provided by the GPU driver to read the total memory capacity and the currently used memory capacity, calculating the ratio between the two to obtain the memory utilization rate. Similar to CPU node data collection, a sampling period is set to periodically collect data on these two indicators, and the collected data is stored in a temporary data file on the local GPU node.

[0074] Step 1063: Use FPGA nodes to collect parallel throughput and resource utilization.

[0075] Furthermore, parallel throughput refers to the amount of data or tasks that an FPGA node can process per unit of time. FPGA nodes implement parallel computing through hardware logic circuits. During data acquisition, a standard parallel processing task flow is input to the FPGA node, and the total number of tasks completed per unit of time is counted to determine the parallel throughput. Resource utilization refers to the usage of internal logic resources within the FPGA node. FPGA logic resources include lookup tables, flip-flops, DSP units, etc. Through the resource monitoring interface provided by the FPGA development tool, the total quantity and used quantity of various resources are read, and the proportion of used resources to total resources is calculated to obtain the resource utilization rate. These two indicators are collected at preset intervals, and the results are temporarily stored in a structured data format.

[0076] Step 1064: The floating-point computing power, memory usage, tensor computing efficiency, video memory usage, parallel throughput, and resource utilization are aggregated to the cluster scheduling center to construct a real-time computing power feature matrix.

[0077] Furthermore, the cluster scheduling center is the control core of the heterogeneous computing cluster, responsible for the overall management of nodes and task allocation. Each node transmits its collected computing power characteristic data to the cluster scheduling center via a high-speed network according to a sampling period. The scheduling center verifies and standardizes the received data, removing abnormal data and unifying the data format. Using node type as rows and computing power characteristics as columns, the standardized computing power characteristic data is filled into the corresponding positions in a matrix, forming a real-time computing power characteristic matrix. Each element in the matrix represents the corresponding computing power characteristic value of the corresponding node. The scheduling center updates this matrix periodically to ensure that it reflects the computing power status of each node in the cluster in real time.

[0078] Step S107: Using the task node matching model, each land cover classification task in the land cover classification task set is assigned to the target node based on the computing power characteristics.

[0079] Specifically, the target node is one of a CPU node, a GPU node, and an FPGA acceleration node. The task node matching model is established using an improved Hungarian algorithm. Step S107 includes: Step 1071: The priority level and complexity quantification value of the land cover classification task are converted into the cost matrix penalty term parameters of the task node matching model.

[0080] Furthermore, the penalty term parameter in the cost matrix is ​​a coefficient used to adjust the task allocation cost. Priority level reflects the urgency of task processing, while the complexity metric reflects the difficulty of task processing. The process first establishes a correspondence between priority levels and penalty coefficients; higher priority corresponds to a larger penalty coefficient, meaning that assigning a high-priority task to a node with lower computing power will result in a higher cost. Simultaneously, the complexity metric is normalized to transform it into a base cost coefficient in the cost matrix; higher complexity results in a larger base cost coefficient. The penalty coefficient corresponding to priority and the base cost coefficient corresponding to complexity are then weighted and fused to obtain the penalty term parameter for each land cover classification task, which is used in the subsequent construction of the cost matrix.

[0081] Step 1072: Input the real-time computing power feature matrix and the task parameters in the task attribute table into the task node matching model to generate the initial matching cost matrix of the task node.

[0082] Furthermore, the real-time computing power feature matrix includes computing power indicators such as floating-point operation capability and tensor computation efficiency of each node, while the task attribute table includes parameters such as the quantified value of task complexity and priority level. During operation, the indicators in the real-time computing power feature matrix are first standardized to eliminate dimensional differences between different indicators. For each land cover classification task and each target node, the fit between the task parameters and the node's computing power features is calculated. The fit is determined by the cosine similarity or Euclidean distance between the two; a lower fit corresponds to a higher allocation cost. Combining the penalty term parameters obtained in step 1071, the cost value corresponding to the fit is adjusted, and the adjusted cost value is filled into the matrix in the order of task row and node column to generate the initial matching cost matrix for task nodes.

[0083] Step 1073: Solve for the minimum cost matching solution of the initial matching cost matrix using the improved Hungarian algorithm.

[0084] Furthermore, the improved Hungarian algorithm is applicable to solving the minimum weight matching problem in a bipartite graph, specifically to find the allocation scheme with the minimum total cost between tasks and nodes. The operation first initializes algorithm parameters such as label vectors and matching vectors. The label vectors record the initial label values ​​of tasks and nodes, and the matching vectors record the matched task-node pairs. Then, the label vectors and matching vectors are updated by alternately searching for augmenting paths. An augmenting path is a path that starts from an unmatched task and passes through alternating unmatched and matched edges. This label update and augmenting path search process is repeated until all tasks or nodes are matched. The resulting matching solution is the minimum cost matching solution of the initial matching cost matrix, ensuring that the total cost of task allocation is minimized.

[0085] Step 1074: Output the task allocation scheme based on the minimum cost matching solution, assign each land feature classification task to the corresponding target node, and generate a task allocation result table.

[0086] Furthermore, the minimum cost matching solution is presented in the form of a task node correspondence, with each land cover classification task corresponding to a unique target node. During operation, the minimum cost matching solution is first parsed to determine the target node identifier for each task, forming a preliminary task allocation scheme. The allocation scheme is then validated for rationality, checking whether the task load of each target node exceeds its maximum computing power capacity. If any node exceeds its capacity, some tasks on that node are reassigned to other suitable nodes with lower loads. After successful validation, the final task allocation scheme is output, and information such as the corresponding target node and allocation cost are recorded for each task, generating a task allocation result table for subsequent task execution and monitoring.

[0087] Step S108: Use the target nodes to perform parallel processing of the land feature classification task, output the land feature classification results of the sub-regions corresponding to the land feature classification task, and output the target accuracy land feature classification map based on the sub-region land feature classification results output by each target node.

[0088] Step 1081: Load the corresponding level of land cover classification model using the target node. Specifically, the FPGA acceleration node loads the first-level land cover classification model, the GPU node loads the second-level land cover classification model, and the CPU node loads the third-level land cover classification model. The accuracy of the first-level land cover classification model is higher than that of the second-level land cover classification model, and the accuracy of the second-level land cover classification model is higher than that of the third-level land cover classification model.

[0089] Furthermore, the land cover classification model is a deep learning model trained to identify land cover categories. Different levels of models differ in network structure complexity and the number of parameters; higher-accuracy models have more complex structures and more parameters. During the loading process, model compatibility checks are first performed on each target node to ensure that the node hardware supports the operation of the corresponding model. The pre-trained land cover classification model files at each level are transferred to the corresponding target node, and the model parameters are read into the node's memory through the model loading interface to complete model initialization and prepare for subsequent classification calculations.

[0090] Step 1082: Use the data sharing channel between adjacent target nodes to exchange ground feature data in the edge area.

[0091] Furthermore, adjacent target nodes refer to nodes that process spatially adjacent sub-region land cover classification tasks, while edge regions refer to the overlapping or boundary areas of two adjacent sub-regions. The data sharing channel is a dedicated data transmission link established based on a high-speed network protocol, used to achieve low-latency data interaction between nodes. First, under the coordination of the cluster scheduling center, a data sharing channel is established for each adjacent target node, and communication parameters, including transmission rate and data format, are configured. Then, the edge range of the sub-region that each node is responsible for processing is determined, and the land cover feature data of the edge region is encapsulated according to a preset data interaction format. The feature data includes the spectral features and texture features of edge pixels, etc., and bidirectional transmission and reception of edge region feature data between adjacent nodes is achieved through the data sharing channel.

[0092] Step 1083: Perform land feature classification calculations in parallel using each target node, and synchronize the land feature feature parameters of the edge area according to a preset period during the calculation process.

[0093] Furthermore, during parallel execution, each target node invokes the loaded land cover classification model, inputs the assigned image area data to be processed into the model, and the model performs calculations such as feature extraction, feature fusion, and classification prediction on the input data to obtain preliminary land cover classification results. The preset period is a time interval determined based on the time consumption of classification calculation and the frequency of changes in edge area data. During the calculation process, each node periodically sends the land cover feature parameters of the edge area to neighboring nodes, while simultaneously receiving feature parameters sent by neighboring nodes. The received edge feature parameters from neighboring nodes are compared and fused with the local edge feature parameters to adjust the classification calculation results of the local edge area, ensuring the consistency of edge classification for adjacent sub-regions.

[0094] Step 1084: Output the sub-regional land cover classification results corresponding to the land cover classification task and save them in the standard remote sensing image format.

[0095] Furthermore, after the classification calculation is completed, each target node processes the calculated sub-region land cover classification results, converting the land cover category identifier of each pixel into the corresponding grayscale value or RGB color value to form a visualized sub-region classification image. The standard remote sensing image format is a common data format in the remote sensing field, supporting the associated storage of geographic coordinate information and image data. During the saving process, the sub-region classification image, along with corresponding geographic coordinate parameters, projection information, and other metadata, is written to a standard format file. The metadata also includes information such as task ID, processing time, and node identifier. The saved file undergoes integrity verification to ensure that there are no missing or incorrect data, completing the output and storage of the sub-region land cover classification results.

[0096] Step 1085: The weighted fusion algorithm is used to stitch together the land cover classification results of each sub-region to obtain a preliminary classification map. The weight of the sub-result corresponding to the first image region set is higher than the weight of the sub-result corresponding to the second image region set.

[0097] Furthermore, the weighted fusion algorithm is a method of merging and stitching together the classification results of different sub-regions by assigning different weights to them. The weights reflect the reliability of the sub-region classification results; the higher the reliability, the greater the weight. The first image region set has high feature complexity, making the classification of its corresponding sub-results more difficult; therefore, it is assigned higher weights to highlight the importance of its classification results. The operation first determines the weight values ​​of the classification results for each sub-region and establishes a weight allocation table. Then, according to the spatial relationship of the sub-regions, the classified images of each sub-region are placed in the overall coordinate frame. The pixel values ​​of overlapping areas are weighted and averaged according to their corresponding weights, while the original classification results are directly retained for non-overlapping areas. Finally, the images are stitched together to form a preliminary classification map covering the entire target area.

[0098] Step 1086: Calculate the land cover classification accuracy and Kappa coefficient of the preliminary classification map based on the confusion matrix, and use the gradient operator to calculate the edge sharpness of the preliminary classification map.

[0099] Furthermore, the confusion matrix is ​​used to evaluate classification accuracy. Rows represent actual land cover categories, columns represent predicted land cover categories, and matrix elements indicate the number of pixels that are actually in one category but are predicted to be in another. The confusion matrix is ​​constructed by statistically analyzing the pixel correspondence between the preliminary classification map and the actual land cover category map. Based on the confusion matrix, the land cover classification accuracy is calculated, which is the proportion of correctly classified pixels to the total number of pixels. The Kappa coefficient is calculated, which comprehensively considers the influence of random classification and can more objectively reflect classification accuracy. The gradient operator is used to detect the edge sharpness of the image. The edge gradient value is obtained by calculating the rate of change of pixel grayscale values ​​in the preliminary classification map, and the average gradient value of all edge pixels is used as the edge sharpness.

[0100] Step 1087: Obtain a comprehensive evaluation score based on the land cover classification accuracy, Kappa coefficient, and edge sharpness.

[0101] Furthermore, the comprehensive evaluation score is a quantitative assessment indicator of the overall quality of the preliminary classification map, and it needs to be calculated by weighting the importance of the three indicators. First, weight coefficients are assigned to the land cover classification accuracy, Kappa coefficient, and edge sharpness, with these coefficients determined based on the actual application's requirements for classification accuracy and edge quality. The three indicators are then normalized to ensure a uniform value range. Next, each normalized indicator is multiplied by its corresponding weight coefficient, and the products are summed to obtain the comprehensive evaluation score. A higher score indicates a better quality preliminary classification map.

[0102] Step 1088: If the comprehensive evaluation score is greater than or equal to the preset accuracy threshold, output the preliminary classification map as the target accuracy land cover classification map.

[0103] Furthermore, if the comprehensive evaluation score is greater than or equal to the preset accuracy threshold, a preliminary classification map is output as the target accuracy land cover classification map. If the comprehensive evaluation score is less than the preset accuracy threshold, a guided filtering algorithm is used to optimize the edge regions of the preliminary classification map. The guided filtering algorithm can smooth edge region noise while preserving image structure information. Using the preliminary classification map as a guide map, filtering calculations are performed on edge region pixels, adjusting the category labels of edge pixels, and improving the accuracy of edge classification. After optimization, the land cover classification accuracy, Kappa coefficient, and edge sharpness are recalculated to obtain a new comprehensive evaluation score. The optimization and evaluation process is repeated until the comprehensive evaluation score is greater than or equal to the preset accuracy threshold, at which point the classification map is output as the target accuracy land cover classification map. During output, the target accuracy land cover classification map is converted into a standard remote sensing image format, along with a classification accuracy evaluation report. The report includes the comprehensive evaluation score, values ​​of various indicators, and confusion matrix information, so that users can understand the accuracy of the classification map.

[0104] The beneficial effects of this invention are as follows: First, by simultaneously acquiring optical remote sensing image data, SAR image data, and digital elevation model data of the target area and constructing an original multi-source data set, the complete acquisition and integration of multi-dimensional remote sensing information is achieved. Optical remote sensing images provide rich spectral information of ground features, SAR images have the advantage of all-day, all-weather observation, and digital elevation model data reflects the topographic relief characteristics. The simultaneous acquisition and collection of these three types of data avoids spatiotemporal deviations between different time periods. Second, by using a feature point matching algorithm to perform spatiotemporal registration processing on the original multi-source data set, spatial coordinate differences and time reference deviations between different types of remote sensing images are eliminated. Through feature point extraction, matching, and mismatch removal, it is ensured that multi-source data such as optical remote sensing images and SAR images are aligned within the same spatiotemporal framework. Then, by using a deep learning semantic segmentation model to identify the ground feature complexity of the processed multi-source image data set, and accordingly dividing it into a first image region set and a second image region set, a refined distinction of ground feature distribution characteristics is achieved. Deep learning semantic segmentation models accurately identify different land cover types and their distribution ranges. Complexity is quantified by statistically analyzing the number of land cover categories, enabling differentiated processing strategies for regions of varying complexity. Furthermore, targeted image optimization is achieved by performing preprocessing operations on the first and second image region sets. For the more complex first image region set, radiometric normalization and Gaussian filtering are performed sequentially to eliminate illumination differences and enhance texture details; for the less complex second image region set, only geometric correction is performed, simplifying the process while maintaining processing quality. This differentiated preprocessing approach improves image quality in complex areas while increasing processing efficiency in simpler areas. Subsequently, land cover classification task priorities are assigned to the first and second image region sets to be processed, generating a land cover classification task set and clarifying the priority order and constraints for processing each region. Multi-dimensional data such as the number of land cover categories and image resolution are extracted to calculate task complexity, and priority is assigned based on preset thresholds. A task attribute table containing task IDs and complexity quantification values ​​is generated, allowing subsequent task allocation to be carried out based on clear priorities and task characteristics. Furthermore, by leveraging the heterogeneous computing cluster to collect the computing power characteristics of each node, a comprehensive understanding of the cluster's real-time computing power distribution is achieved. Computing power acquisition agents are deployed on different nodes such as CPUs, GPUs, and FPGAs to collect specific computing power indicators, including floating-point arithmetic capabilities and tensor computation efficiency. These are then aggregated to construct a real-time computing power feature matrix, dynamically reflecting the load status and computing advantages of each node. Finally, a task node matching model is used to allocate the set of land cover classification tasks to target nodes based on computing power characteristics, achieving optimal task-node matching.Task priority and complexity are converted into cost matrix penalty terms. An initial matching cost matrix is ​​generated by combining this with a real-time computing power feature matrix. An improved Hungarian algorithm is used to find the minimum cost matching solution, allowing high-priority, high-complexity tasks to be assigned to nodes with stronger computing power, and low-priority, low-complexity tasks to suitable nodes, thus improving the overall computing power utilization and task processing efficiency of the cluster. Finally, by utilizing target nodes to perform parallel processing of land cover classification tasks and outputting sub-region land cover classification results, a target-precision land cover classification map is output based on the sub-results, achieving efficient processing and high-precision output of massive image data. Parallel computation of each target node significantly shortens the overall processing time, high-speed data sharing and feature synchronization between adjacent nodes ensure classification consistency in edge areas, and weighted fusion and comprehensive evaluation score assessment ensure that the accuracy of the stitched classification map meets the standards, ultimately achieving efficient and accurate remote sensing image land cover classification results. In summary, the remote sensing image clustering parallel processing method provided in this embodiment effectively solves many shortcomings of remote sensing image processing technology in terms of multi-source data differentiation processing, heterogeneous cluster task scheduling, and inter-node collaborative computing, making it difficult to simultaneously meet the efficiency and accuracy requirements of massive multi-source remote sensing image processing.

[0105] This embodiment also provides a remote sensing image clustered parallel processing system, which is used to implement the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the system described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0106] This embodiment provides a remote sensing image clustering parallel processing system, such as Figure 2 As shown, it includes: The acquisition module 201 is used to simultaneously acquire optical remote sensing image data, SAR image data and digital elevation model data of the target area to construct the original multi-source data set.

[0107] The matching module 202 is used to perform spatiotemporal registration processing on the original multi-source data set using a feature point matching algorithm to obtain a multi-source image data set to be processed.

[0108] The identification module 203 is used to identify the ground feature complexity of the multi-source image data set to be processed based on a deep learning semantic segmentation model, and to divide the multi-source image data set to be processed into a first image region set and a second image region set based on the ground feature complexity identification result.

[0109] The preprocessing module 204 is used to perform preprocessing operations on the first image region set and the second image region set respectively to obtain the first image region set to be processed and the second image region set to be processed.

[0110] The labeling module 205 is used to label the first set of image regions to be processed and the second set of image regions to be processed with the priority of the land cover classification task to obtain the land cover classification task set.

[0111] The computing power module 206 is used to collect the computing power characteristics of each node using a heterogeneous computing cluster.

[0112] The allocation module 207 is used to allocate each land cover classification task in the land cover classification task set to the target node based on the computing power characteristics using the task node matching model.

[0113] The processing module 208 is used to perform parallel processing of the land feature classification task using the target nodes, output the land feature classification results of the sub-regions corresponding to the land feature classification task, and output the target accuracy land feature classification map based on the sub-region land feature classification results output by each target node.

[0114] Based on the spirit of this invention, those skilled in the art will readily conceive of a computer program product derived from the aforementioned remote sensing image clustering parallel processing method. The computer program product may include a computer-readable storage medium on which computer-readable program instructions are loaded to enable a processor to implement various aspects of this disclosure. That is, this application also includes a terminal comprising a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps according to the aforementioned remote sensing image clustering parallel processing method.

[0115] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example, but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0116] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0117] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for clustered parallel processing of remote sensing images, characterized in that, Includes the following steps: Simultaneously acquire optical remote sensing image data, SAR image data and digital elevation model data of the target area to construct the original multi-source dataset; The original multi-source data set is spatiotemporally registered using a feature point matching algorithm to obtain a multi-source image data set to be processed. Based on a deep learning semantic segmentation model, the set of multi-source image data to be processed is used to identify the complexity of ground features. Based on the results of the identification of ground feature complexity, the set of multi-source image data to be processed is divided into a first image region set and a second image region set. Preprocessing operations are performed on the first image region set and the second image region set respectively to obtain the first image region set to be processed and the second image region set to be processed. The first set of image regions to be processed and the second set of image regions to be processed are marked with the priority of land cover classification tasks to obtain the set of land cover classification tasks. The computing power characteristics of each node are collected using a heterogeneous computing cluster; Based on the computing power characteristics, the task node matching model is used to assign each land cover classification task in the land cover classification task set to the target node; The target nodes are used to perform parallel processing of the land cover classification task, output the land cover classification results of the sub-regions corresponding to the land cover classification task, and output the target accuracy land cover classification map based on the sub-region land cover classification results output by each target node.

2. The method according to claim 1, characterized in that, The simultaneous acquisition of optical remote sensing image data, SAR image data, and digital elevation model data of the target area to construct the original multi-source dataset further includes: Collect target area image data returned by optical remote sensing satellites, perform cloud-free screening on the target area image data, and determine the target time period and corresponding optical remote sensing image data. Collect SAR image data of the target area returned by SAR satellites within the target time period; Using InSAR technology to obtain digital elevation model data of the target area; The optical remote sensing image data, SAR image data, and digital elevation model data are aligned according to spatial coordinates to construct the original multi-source data set.

3. The method according to claim 1, characterized in that, The step of performing spatiotemporal registration processing on the original multi-source data set using a feature point matching algorithm to obtain a multi-source image data set to be processed further includes: The SIFT algorithm was used to extract feature points from optical remote sensing images and SAR images in the original multi-source dataset, resulting in optical remote sensing image feature points and SAR image feature points. The K-nearest neighbor algorithm is used to perform preliminary matching of feature points in optical remote sensing images and SAR images to obtain the first set of feature point pairs. The RANSAC random sampling consensus algorithm is used to remove mismatched feature point pairs from the first feature point pair set to obtain the second feature point pair set. Spatiotemporal registration is performed on the set based on the second feature point to obtain the set of multi-source image data to be processed.

4. The method according to claim 3, characterized in that, The deep learning semantic segmentation model adds a channel attention module to the model encoding end of the BiSeNet V2 model; the deep learning semantic segmentation model performs ground feature complexity identification on the multi-source image dataset to be processed, and divides the multi-source image dataset to be processed into a first image region set and a second image region set based on the ground feature complexity identification result, further including: The multi-source image data set to be processed is normalized. The normalized set of multi-source image data to be processed is input into the deep learning semantic segmentation model, and forward inference calculation is performed. The probability distribution of land cover categories for each pixel is obtained by using the Softmax function of the model output layer, and the category corresponding to the maximum probability is taken as the land cover category prediction result. The number of land cover categories in each image region is counted, and regions with 5 or more land cover categories are divided into the first image region set. Areas with fewer than 5 land cover categories are divided into a second image region set, and a region division mask map is generated.

5. The method according to claim 4, characterized in that, The step of performing preprocessing operations on the first image region set and the second image region set respectively to obtain the first image region set to be processed and the second image region set to be processed further includes: Perform radiometric normalization and Gaussian filtering on the first image region set to obtain the enhanced preprocessed first image region set. A quadratic polynomial correction model is used to perform geometric correction on the first image region set after enhancement preprocessing to obtain the first image region set to be processed. A quadratic polynomial correction model is used to perform geometric correction on the second image region set to obtain the second image region set to be processed.

6. The method according to claim 5, characterized in that, The step of prioritizing the first set of image regions to be processed and the second set of image regions to be processed to obtain a set of land cover classification tasks includes: Extract the number of land cover categories, image resolution data, and texture entropy data from the first and second image regions to be processed. The number of land cover categories, image resolution data, and texture entropy data of the first and second image regions to be processed are input into the classification task complexity evaluation model to obtain a set of task complexity quantification values. The set of task complexity quantification values ​​includes the task complexity quantification values ​​corresponding to each image region in the first and second image regions to be processed. Based on the task complexity threshold and the task complexity quantification value set, the first image region set and the second image region set to be processed are marked with the priority of the land cover classification task, so as to obtain the land cover classification task set. Then, a task attribute table containing task ID, complexity quantification value, priority level and accuracy time consumption constraint is generated for each land cover classification task in the land cover classification task set.

7. The method according to claim 6, characterized in that, The heterogeneous computing cluster includes CPU nodes, GPU nodes, and FPGA acceleration nodes. The computing power characteristics include CPU floating-point operation characteristics, GPU tensor computation efficiency characteristics, and FPGA parallel throughput characteristics. The process of collecting the computing power characteristics of each node using the heterogeneous computing cluster includes: Utilize CPU nodes to collect floating-point arithmetic capabilities and memory usage. Utilize GPU nodes to acquire tensor computation efficiency and GPU memory utilization; Utilize FPGA nodes to collect parallel throughput and resource utilization; The floating-point computing power, memory usage, tensor computation efficiency, video memory usage, parallel throughput, and resource utilization are aggregated to the cluster scheduling center to construct a real-time computing power feature matrix.

8. The method according to claim 7, characterized in that, The target node is one of a CPU node, a GPU node, and an FPGA acceleration node. The task node matching model is established using an improved Hungarian algorithm. The process of using the task node matching model to assign each land cover classification task in the land cover classification task set to the target node based on the computing power characteristics includes: The priority level and complexity quantification value of the land cover classification task are converted into the cost matrix penalty term parameters of the task node matching model; The real-time computing power feature matrix and the task parameters in the task attribute table are input into the task node matching model to generate the initial matching cost matrix of the task node. The minimum cost matching solution of the initial matching cost matrix is ​​obtained by improving the Hungarian algorithm; Based on the minimum cost matching solution, the task allocation scheme is output, and each land feature classification task is assigned to the corresponding target node, and a task allocation result table is generated.

9. The method according to claim 8, characterized in that, The method of using target nodes to perform parallel processing of the land cover classification task, outputting the land cover classification results of the sub-regions corresponding to the land cover classification task, and outputting a land cover classification map with target accuracy based on the sub-region land cover classification results output by each target node, includes: The target node loads the corresponding level of land cover classification model. The FPGA acceleration node loads the first level land cover classification model, the GPU node loads the second level land cover classification model, and the CPU node loads the third level land cover classification model. The accuracy of the first level land cover classification model is higher than that of the second level land cover classification model, and the accuracy of the second level land cover classification model is higher than that of the third level land cover classification model. Utilize data sharing channels between adjacent target nodes to exchange geographic feature data in edge areas; The land cover classification calculation is performed in parallel at each target node, and the land cover feature parameters of the edge area are synchronized according to a preset period during the calculation process. Output the sub-region land cover classification results corresponding to the land cover classification task and save them in the standard remote sensing image format; A weighted fusion algorithm is used to stitch together the land cover classification results of each sub-region to obtain a preliminary classification map. The weight of the sub-result corresponding to the first image region set is higher than the weight of the sub-result corresponding to the second image region set. The accuracy of land cover classification and the Kappa coefficient of the preliminary classification map are calculated based on the confusion matrix, and the edge sharpness of the preliminary classification map is calculated using the gradient operator. A comprehensive evaluation score is obtained based on the accuracy of land cover classification, Kappa coefficient, and edge sharpness. If the comprehensive evaluation score is greater than or equal to the preset accuracy threshold, the preliminary classification map is output as the target accuracy land cover classification map.

10. A remote sensing image clustered parallel processing system, characterized in that, include: The acquisition module is used to simultaneously acquire optical remote sensing image data, SAR image data and digital elevation model data of the target area to construct the original multi-source data set; The matching module is used to perform spatiotemporal registration processing on the original multi-source data set using a feature point matching algorithm to obtain a multi-source image data set to be processed. The identification module is used to identify the ground feature complexity of the multi-source image data set to be processed based on a deep learning semantic segmentation model, and to divide the multi-source image data set to be processed into a first image region set and a second image region set based on the ground feature complexity identification result. The preprocessing module is used to perform preprocessing operations on the first image region set and the second image region set respectively to obtain the first image region set to be processed and the second image region set to be processed. The labeling module is used to label the first set of image regions to be processed and the second set of image regions to be processed with the priority of land cover classification tasks, so as to obtain a set of land cover classification tasks. The computing power module is used to collect the computing power characteristics of each node in a heterogeneous computing cluster. The allocation module is used to use the task node matching model to allocate each land cover classification task in the land cover classification task set to the target node based on the computing power characteristics. The processing module is used to perform parallel processing of the land cover classification task using the target nodes, output the land cover classification results of the sub-regions corresponding to the land cover classification task, and output the target accuracy land cover classification map based on the sub-region land cover classification results output by each target node.

11. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the remote sensing image clustering parallel processing method according to any one of claims 1-9.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the remote sensing image clustering parallel processing method according to any one of claims 1-9.

Images