A method and device for generating a large-area remote sensing image land cover map

Abstract

The application provides a large-area remote sensing image ground cover map generation method and device, which aims to ensure the integrity of the remote sensing image ground cover map. The large-area remote sensing image ground cover map generation method comprises the following steps: obtaining the prediction results of ground objects corresponding to first, second and third block modes respectively; obtaining large-scale ground object information fusion results from the union area of the first and second prediction results based on the processing priority of the large-scale ground objects; on this basis, the union area of the first and second prediction results is obtained based on the processing priority of the medium-scale ground objects, and the medium-scale ground object data is fused with the third prediction result to obtain the medium-scale ground object information fusion result; and based on the processing priority of the small-scale ground objects, the small-scale ground object data in the third prediction result is fused with the medium-scale ground object information fusion result to generate a target area remote sensing image ground cover map.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method and apparatus for generating land cover maps from large-area remote sensing images. Background Technology

[0002] With the rapid development of remote sensing technology and the continuous improvement of image spatial resolution, semantic segmentation of land cover technology, as a core component of remote sensing image processing, can accurately identify different land cover types such as crops, forests, water bodies, and buildings, providing data support for generating land cover maps.

[0003] Currently, land cover maps are typically obtained using a "block prediction-information aggregation" approach. A common practice is to use fixed-size, overlapping blocks for prediction, then perform pixel-level voting on multiple predictions for the same location within the overlapping area (e.g., selecting the category with the highest frequency) to complete the full map stitching. However, this method not only ignores the inherent scale characteristics and spatial context of land features, easily generating spatially discontinuous and isolated pixels that contradict surrounding features, making it easy for land features to lose details due to fixed block segmentation, but also amplifies the discontinuity in stitching adjacent sub-remote sensing images due to the block effect caused by fixed blocks. These shortcomings result in fragmented patches and incomplete land feature morphology in the generated land cover map.

[0004] Therefore, there is an urgent need for a method to generate land cover maps from large-area remote sensing imagery. Summary of the Invention

[0005] This application provides a method and apparatus for generating land cover maps from large-area remote sensing imagery. The purpose is to design a multi-modal segmentation method based on the size characteristics of land features, and to sequentially fuse land feature data at different scales by combining processing priorities. This preserves the integrity of large-scale land features, restores the details of medium-scale land features, and accurately captures information about small-scale land features, reducing block effects and classification bias. This ensures the integrity, continuity, and precision of the land cover map from large-area remote sensing imagery, and solves the technical problem of incomplete land feature morphology in the generation of land cover maps from large-area remote sensing imagery in the prior art.

[0006] To address the aforementioned technical problems, this application provides the following technical solutions: The first aspect of this application provides a method for generating land cover maps from large-area remote sensing imagery, comprising: When the target area is larger than a preset area, the remote sensing image of the target area is sliced ​​into blocks using the first block mode, the second block mode, and the third block mode, respectively. The block results are then processed into prediction results for land features using a land feature recognition model. Specifically, the first block mode uses a preset size for segmentation, the second block mode uses a preset size and a preset offset for segmentation, and the third block mode uses a preset ratio size smaller than the preset size for segmentation. The first prediction result has processing priorities for large-scale, medium-scale, and small-scale land features. The first block mode is used to preserve the macroscopic features of large-scale land features with a preset size, the second block mode is used to cover the segmentation lines of the first block mode with a preset size and a preset offset, and the third block mode is used to capture the detailed features of small-scale land features with a preset ratio size smaller than the preset size. Based on the processing priority of large-scale land features, the fusion result of large-scale land feature information is obtained from the union region of the first prediction result and the second prediction result. Based on the large-scale land cover information fusion results, and based on the processing priority of mesoscale land cover, the mesoscale land cover data obtained from the union region of the first and second prediction results is fused with the mesoscale land cover data in the third prediction result to obtain the mesoscale land cover information fusion result. Based on the processing priority of small-scale features, the small-scale feature data in the third prediction result is fused with the fusion result of the medium-scale feature information to generate a remote sensing image land cover map of the target area.

[0007] A second aspect of this application provides an apparatus for generating land cover maps from large-area remote sensing imagery, comprising: The acquisition unit is used to slice the remote sensing image of the target area into blocks using a first block mode, a second block mode, and a third block mode when the area of ​​the target area is larger than a preset area. The block results are then processed into prediction results for land features using a land feature recognition model. The first block mode uses a preset size for segmentation, the second block mode uses a preset size and a preset offset for segmentation, and the third block mode uses a preset ratio size smaller than the preset size for segmentation. The first prediction result has processing priorities for large-scale, medium-scale, and small-scale land features. The first block mode is used to preserve the macroscopic features of large-scale land features with a preset size. The second block mode is used to cover the segmentation lines of the first block mode with a preset size and a preset offset. The third block mode is used to capture the detailed features of small-scale land features with a preset ratio size smaller than the preset size. The unit is used to obtain the fusion result of large-scale land cover information from the union region of the first prediction result and the second prediction result based on the processing priority of the large-scale land cover of the unit. The obtaining unit is used to, based on the large-scale ground feature information fusion result and the processing priority of the medium-scale ground features, fuse the medium-scale ground feature data obtained in the union region of the first prediction result and the second prediction result with the medium-scale ground feature data in the third prediction result to obtain the medium-scale ground feature information fusion result. The obtained unit is used to fuse the small-scale ground feature data in the third prediction result with the fusion result of the medium-scale ground feature information based on the processing priority of small-scale ground features, and generate a remote sensing image land cover map of the target area.

[0008] A third aspect of this application provides a storage medium comprising a stored program, wherein, when the program is executed, the device containing the storage medium is controlled to perform the aforementioned method for generating a large-area remote sensing image land cover map.

[0009] A fourth aspect of this application provides an electronic device, the device including at least one processor, at least one memory and a bus connected to the processor; wherein the processor and the memory communicate with each other through the bus; the processor is used to call program instructions in the memory to execute the above-described method for generating a large-area remote sensing image land cover map.

[0010] Compared to existing technologies, this application provides a method for generating land cover maps from large-area remote sensing imagery. By introducing a block-segmentation strategy and performing progressive, layered fusion based on the processing priority of land features, it effectively solves the technical problems caused by fixed block segmentation and pixel voting in existing technologies. Specifically, this method first ensures that land features of each scale can be identified in the most suitable slice of remote sensing imagery based on different block segmentation patterns, thus avoiding the fragmentation and loss of detail of large-scale land features caused by fixed block segmentation at the source. Second, this method does not simply superimpose all prediction results at once, but fuses them according to a priority order from large scale to small scale: first, large-scale land features are fused to establish a macro framework, ensuring its continuity and integrity; then, medium-scale land features are gradually incorporated to enrich details, and finally, small-scale land features are superimposed to capture fine information. This layered and progressive fusion mechanism not only significantly reduces the "block effect" caused by inconsistent block boundaries, but also greatly maintains the spatial continuity of land feature classification results by prioritizing the fixing of macro-scale land feature outlines and then filling in details. Through the above mechanism, it is ensured that the generated land cover map can accurately reflect the true distribution of land features within a large area. Attached Figure Description

[0011] The above and other objects, features, and advantages of exemplary embodiments of this application will become readily understood by reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of this application are illustrated by way of example and not limitation, with the same or corresponding reference numerals denoteing the same or corresponding parts, wherein: Figure 1 A flowchart illustrating a method for generating land cover maps from large-area remote sensing imagery is shown schematically. Figure 2 The flowchart schematically illustrates another method for generating land cover maps from large-area remote sensing imagery. Figure 3 A schematic diagram of a device for generating land cover maps from large-area remote sensing imagery is shown. Figure 4 The diagram schematically illustrates the structure of another device for generating land cover maps from large-area remote sensing imagery. Detailed Implementation

[0012] Exemplary embodiments of this application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of this application are shown in the drawings, it should be understood that this application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of this application and to fully convey the scope of this application to those skilled in the art.

[0013] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains.

[0014] Based on this, the applicant proposes a novel method to optimize map accuracy through a hierarchical processing strategy. The core of this method lies in two aspects: First, designing a multi-modal segmentation mechanism based on differences in feature size characteristics to provide the optimal recognition environment for features at different scales; second, abandoning the traditional simple voting strategy and adopting a progressive fusion process based on feature processing priorities. This process follows a logical order from macro to micro, prioritizing the integrity of large-scale features before gradually integrating detailed information, thereby improving classification accuracy while maintaining spatial continuity. Specifically, as follows... Figure 1 As shown, Step 101: When the target area is larger than the preset area, the remote sensing image of the target area is sliced ​​into blocks using the first block mode, the second block mode and the third block mode respectively, and the block results are processed into prediction results of ground features using the ground feature recognition model.

[0015] In this step, the first segmentation mode is used to preserve the macroscopic features of large-scale features at a preset size. The second segmentation mode is used to cover the segmentation lines of the first segmentation mode with a preset size and a preset offset. The third segmentation mode is used to capture the detailed features of small-scale features at a preset scale smaller than the preset size. The first segmentation mode uses a preset size for segmentation, the second segmentation mode uses a preset size and a preset offset for segmentation, and the third segmentation mode uses a preset scale smaller than the preset size for segmentation. The preset offset is set according to the actual situation. For example, taking the first segmentation mode as a reference, the row and column offsets are S / 2. The preset scale smaller than the preset size can be 1 / 2. The processing priority of large-scale, medium-scale, and small-scale features in the first prediction result is determined. The prediction results are the stitched land cover data obtained by using a land cover recognition model to obtain sliced ​​remote sensing images under different segmentation strategies (first segmentation mode, second segmentation mode, and third segmentation mode). The different segmentation strategies are based on the size characteristics of land cover and specifically include the first segmentation mode, the second segmentation mode, and the third segmentation mode. The first segmentation mode corresponds to the first prediction result, the second segmentation mode corresponds to the second prediction result, and the third segmentation mode corresponds to the third prediction result. The first prediction result contains the types of land cover in the target area and the processing priority of the land cover types. That is, the first prediction result has the processing priority of large-scale, medium-scale, and small-scale land cover. The types of land cover include large-scale land cover, medium-scale land cover, and small-scale land cover. The first segmentation mode preserves the macroscopic features of land features with a preset size. The second segmentation mode covers the segmentation line of the first mode with a preset size plus a preset offset to reduce the block effect. The third segmentation mode captures the detailed features of small-scale land features with a preset ratio size smaller than the preset size. The three are based on the same preset size and form a complementary segmentation system of "macroscopic preservation - boundary optimization - detail supplementation" to jointly adapt to the recognition needs of land features at different scales.

[0016] Furthermore, for example, the first partitioning pattern: starting point (0,0), block size 4096×4096. The second partitioning pattern: starting point (d,d), block size 4096×4096. If the two grid lines overlap, then there exist integers m,m ， Make 4096m = d + 4096m ， Therefore, as long as d (d is the offset) is not an integer multiple of 4096, the two grid lines will not overlap, so d∈{1,2,…,4095}.

[0017] In this step, the preset size is the basic block size. Different block strategies specifically include: acquiring the basic block size and the original remote sensing image of the target area; the first block mode is to crop the original remote sensing image of the target area into multiple slice remote sensing images with the same basic block size; the second block mode is to crop the original remote sensing image of the target area with the basic block size and offset the row and column directions by half the basic block size to obtain multiple slice remote sensing images; the third block mode is to crop the original remote sensing image into multiple slice remote sensing images with a size half that of the basic block size. Specifically, firstly, using the training sample size of the land cover recognition model as the basic unit, and combining it with the hardware computing power limit, multiple sets of candidate block sizes are generated. The optimal basic block size is then selected in a representative verification area using a pixel-level consistency index. Simultaneously, the original remote sensing image of the target area is acquired to ensure that the image contains complete land cover information. The first mode neatly cuts the image according to the basic block size, preserving the macroscopic shape of ground features. The second mode cuts the image according to the basic block size, offsetting it by half in each row and column direction to cover the block lines of the first mode. The third mode cuts the original image into slices half the size of the basic block, accurately capturing small-scale ground feature details and forming a multi-dimensional block data system. Candidate block sizes are multiple sets of alternative clipping sizes generated based on the model sample size and hardware computing power, providing a basis for determining the basic block size. The pixel-level consistency index is used to measure the consistency between the predicted results of different block sizes and the actual annotations, and is the core standard for selecting the optimal basic block size. The block line is the boundary line formed after clipping in the first block mode, which is prone to ground feature breaks; the second mode can cover this line through offset clipping.

[0018] It is worth noting that obtaining the basic block size includes obtaining the training sample size of the land cover recognition model and the GPU computing power limit; based on the training sample size and the GPU computing power limit, a candidate block size set containing 1U, 2U, 4U, and 8U specifications is generated; in the verification area, the candidate block sizes are used to perform block prediction and stitching processing on remote sensing images respectively, and the manually labeled ground truth of the verification area is used as a reference to calculate the pixel-level consistency index between the predicted image corresponding to each candidate block size and the ground truth; the change law of the pixel-level consistency index with the increase of block size is analyzed, and when the block size increases to a certain specification and the pixel-level consistency index no longer significantly improves, the block size of that specification is determined as the optimal basic block size. This method accurately solves the problems of subjective selection of basic block size and poor adaptability to model and hardware in the prior art. By combining the model sample size and GPU computing power to generate a candidate set, and then using the ground truth of the verification area and the consistency index to select the optimal size, it ensures that the block size is both compatible with the model and does not exceed the hardware capacity, laying an accurate foundation for subsequent block prediction and reducing the block effect from the source.

[0019] One method for obtaining prediction results corresponding to different segmentation strategies is as follows: All sliced ​​remote sensing images obtained under different segmentation modes are input into the land cover recognition model to obtain the first prediction result, the second prediction result, and the third prediction result. The sample data of the land cover recognition model consists of remote sensing images of basic segmentation size and corresponding labeled raster maps, where pixel values ​​in the labeled raster map represent the types of land cover. Specifically, according to the training requirements of the land cover recognition model, remote sensing images of segmentation size determined by the inherent characteristics of land cover are selected as sample base maps. For various land cover types such as crops, forests, and buildings in the sample base map, different pixel values ​​are used to accurately label the categories in the labeled raster map, constructing an "image-label" corresponding training sample set. The model parameters are optimized through iterative training to ensure that the model can accurately output land cover category prediction results. Subsequently, all sliced ​​remote sensing images under the three segmentation modes are input into the trained land cover recognition model according to a unified data format. The model outputs the land cover semantic segmentation results for each slice. Then, based on the spatial coordinates of each slice in the original remote sensing image, the prediction results of slices under the same segmentation mode are stitched together to finally obtain the first, second, and third prediction results corresponding to the three segmentation modes, respectively. The "image-annotation" corresponding training sample set consists of remote sensing images and annotation rasters with segmentation sizes determined by the inherent characteristics of the land cover. Figure 1 A dataset consisting of one-to-one correspondences provides input and label references for training the land cover recognition model. The semantic segmentation result is the pixel-level land cover category prediction result output by the land cover recognition model for the sliced ​​remote sensing image, clearly reflecting the distribution of various land cover categories within the slice. Spatial coordinate stitching is the process of integrating the scattered slice prediction results into a complete regional prediction result based on the position coordinates of the slice in the original image, ensuring the continuity of the spatial distribution of land cover categories. The above steps realize a complete closed loop from land cover recognition model training to prediction stitching, accurately solving the core problems of "mismatch between sliced ​​prediction and land cover scale" and "easy loss of details during prediction result stitching." During the training phase, precise "image-label" matching ensures the model's recognition accuracy for various land cover categories; during prediction, multi-mode slices are input in a unified format to ensure consistent model output; the stitching phase integrates based on spatial coordinates to avoid spatial misalignment of land cover categories. The final three types of prediction results are adapted to the processing needs of large, medium, and small-scale land cover categories, providing high-quality data support for subsequent hierarchical fusion, reducing block effects from the source, and improving the accuracy and continuity of large-area remote sensing image classification maps.

[0020] In this step, acquiring remote sensing images of the target area using different segmentation strategies and the corresponding prediction results includes: acquiring the number of patches in the connected regions of each land cover type in the first prediction result, and calculating the area of ​​a single patch for each land cover type; for each land cover type, acquiring the number of patches with a single patch area greater than an area threshold and the total number of all patches, and calculating the proportion of large-area patches; when the proportion of large-area patches is greater than or equal to a first threshold, the land cover type is determined to be a large-scale land cover; when the proportion of large-area patches is less than the first threshold but greater than a second threshold, the land cover type is determined to be a medium-scale land cover, wherein the first threshold is greater than the second threshold; when the proportion of large-area patches is less than or equal to the second threshold, the land cover type is determined to be a small-scale land cover. This method constructs a quantitative classification standard for land cover scale through patch quantity statistics, area calculation, and proportion threshold determination, accurately solving the core problem that existing technologies cannot adaptively classify types based on the inherent scale of land covers. It avoids the fragmentation of large-scale land covers due to improper segmentation and prevents small-scale land covers from being missed, providing a clear basis for subsequent layered fusion. Meanwhile, by using thresholds to distinguish mesoscale features, we can ensure that such features can balance macroscopic continuity with microscopic detail processing, optimize the block segmentation strategy and the adaptability of feature scale from the source, and lay the foundation for improving the accuracy of the final map.

[0021] Step 102: Based on the processing priority of large-scale land features, obtain the fusion result of large-scale land feature information from the union region of the first prediction result and the second prediction result.

[0022] In this step, the processing priority can be set according to the actual situation, or it can be sorted in ascending order of the total area of ​​each land type in the first prediction result.

[0023] From the large-scale land features determined in step 101, a processing priority list is generated (e.g., forest → crops → grassland → horticultural fruit trees → buildings). For each land feature category in the list, spatial overlay analysis is performed sequentially to extract its union region in the first and second prediction results. A union mask is constructed using binary masking technology to accurately mark the spatial distribution range of this type of land feature in the two prediction results. Furthermore, a mask is constructed for the obtained union region to prevent modification of the processed union region when processing other land features later. In the fusion result to be generated, modification marker records are retrieved to accurately identify regions that have not been modified by any previous processing steps (e.g., prediction results). Operations are performed only on unmodified regions, avoiding efficiency losses caused by repeated processing and maximizing the preservation of effective information in the original image, preventing feature distortion caused by overprocessing. Pixel writing and modification marker updating: The current pixel value is written to all unmodified regions in the fusion result (i.e., the intersection of the union region and the unmodified region). Immediately after writing, the modification marker is updated to record the processing status of these regions. For areas marked as "modified," no further write operations are performed, regardless of whether they fall within the union of the current pixels, to ensure that the retained key information is not overwritten. Subsequently, the Canny edge detection algorithm is used to extract the complete outline of the current feature patch. The length of the intersecting pixels between each outline and predefined block lines is calculated using a coordinate calculation tool. The predefined block lines are determined based on the first and second block modes, as follows: When the starting point of the first block mode is (0,0), and the block size is S*S, the block line position is: x=0,S,2S,…; the starting point of the second block mode is (d,d), and the block size is S*S; therefore, the block line is x=d,d+S,d+2S,… (d is a preset offset). The preset block line positions are mS and d+m. ， S, where m and m ，It is a non-negative integer. If the intersection length exceeds a preset threshold, the "preserved value cache" for that region is searched: if a historical preserved value exists, the region's pixel value is restored to the preserved value; if not, the current pixel value is stored in the cache and marked as a new preserved value, while the region in the "modify label matrix" is reset to "0" to reserve space for subsequent optimization. Finally, all large-scale land cover categories and their corresponding contours in the priority list are processed iteratively until all categories have been processed, resulting in the large-scale land cover information fusion result. This step aims to eliminate large-scale feature fragmentation and ensure spatial continuity. It achieves this goal through three key design aspects: First, prioritizing features by total land area to avoid covering large-scale features in small areas, thus addressing the imbalance of "large areas being prioritized over small areas" in traditional fusion methods. Second, using binary masks and modified marker records, unprocessed areas are precisely located, eliminating efficiency losses from repeated processing and preventing feature distortion caused by overprocessing. Third, combining Canny edge detection and a value retention cache library, feature outline breaks at block lines are specifically repaired, significantly reducing the block effect. The resulting large-scale feature fusion lays a complete and continuous macroscopic feature skeleton for subsequent layered fusion, effectively solving the problem of insufficient integrity caused by block splicing in existing technologies.

[0024] Step 103: Based on the large-scale ground feature information fusion result, and based on the processing priority of mesoscale ground features, the mesoscale ground feature data obtained in the union region of the first prediction result and the second prediction result are fused with the mesoscale ground feature data in the third prediction result to obtain the mesoscale ground feature information fusion result.

[0025] In this step, using the large-scale land cover information fusion result (F_large) output in step 102 as the base layer, the processing priority of the mesoscale land cover determined in step 101 is first extracted, and various land cover types are processed in order of priority. For the current mesoscale land cover category to be processed, the first step is to perform large-scale basic optimization: following the core logic of "union fusion + boundary repair" in step 102, the union region of this type of land cover in the first prediction result and the second prediction result is extracted through spatial overlay analysis tools. The union mask is constructed using binary masking technology to accurately mark the spatial distribution of land cover in the two prediction results. The second step involves enhancing the scale details: From the third prediction result, spatial coordinate matching is performed based on the geographic coordinate system of the original remote sensing image to extract all patches of the mesoscale feature (the slices can fully preserve details such as narrow waterways and side roads); using a logical union algorithm, these patches are overlaid with the corresponding areas of the large-scale optimized F_large layer—if the corresponding area in F_large is unmodified or the feature category is inconsistent, the pixel values ​​of the patches in the third prediction result are replaced to supplement the detailed information; if the area already contains large-scale features, the original pixel values ​​are retained to avoid conflicts. This process is repeated for all mesoscale feature categories, simultaneously updating the "modification marker matrix" and the "retained value cache," ultimately generating the mesoscale feature information fusion result. The mesoscale feature information fusion result is a fusion of large-scale and mesoscale features.

[0026] Step 104: Based on the processing priority of small-scale features, the small-scale feature data in the third prediction result is fused with the fusion result of the medium-scale feature information to generate a remote sensing image land cover map of the target area.

[0027] For small-scale features, they are processed sequentially according to the priority of ascending total pixel area by land type (processing priority of small-scale features). First, from the third prediction result C, all small patches belonging to the current processing type are extracted based on the land feature category attribute, ensuring complete extraction of small-scale features without omission. Then, based on the geographic coordinate system of the original remote sensing image, precise spatial coordinate matching is performed, and these small patches are overlaid and fused with the fusion result of the medium-scale land feature information obtained in step 103 (F_medium), prioritizing the preservation of small patch information to fill the detail gaps in F_medium. After completing the fusion of all small-scale features, the complete remote sensing image semantic segmentation result is output in GeoTIFF format, and then the raster data is converted into vector data using block processing and polygonization algorithms through the GDAL library, finally generating a large-area remote sensing image land cover map.

[0028] Furthermore, this application provides a more detailed discussion on the generation of land cover maps from large-area remote sensing imagery, such as... Figure 2As shown, the details are as follows: Step 201: When the target area is larger than the preset area, acquire remote sensing images of the target area with different segmentation strategies and the corresponding prediction results.

[0029] The discussion in step 201 is the same as that in step 101, and will not be repeated here.

[0030] Step 202: Based on the processing priority of large-scale features, obtain the large-scale feature information fusion result from the union region of the first prediction result and the second prediction result; based on the large-scale feature information fusion result, and based on the processing priority of mesoscale features, fuse the mesoscale feature data obtained from the union region of the first prediction result and the second prediction result with the mesoscale feature data in the third prediction result to obtain the mesoscale feature information fusion result.

[0031] In this step, the process of obtaining the large-scale feature information fusion result from the union region of the first and second prediction results based on the processing priority of large-scale features includes: acquiring features based on the processing priority of large-scale features and processing the pixel values ​​of each feature individually, wherein the pixel value processing is used to extract the union region of the first and second prediction results; and creating a union mask based on the union region, wherein the union mask is used to mark the union region as having been corrected. This processing method accurately solves the problem in the prior art that "large-scale feature block fusion is prone to repeated processing or omission of key areas". Processing the pixel values ​​one by one according to the processing priority can ensure that small-area large-scale features are not covered by large-area features, avoiding the imbalance of "emphasizing area and neglecting details" in traditional fusion. By extracting the union region and creating a union mask, the corrected areas can be accurately marked, which not only eliminates the efficiency loss and feature distortion caused by repeated processing, but also clarifies the effective range of subsequent processing, ensuring that each region is accurately optimized during large-scale feature fusion, significantly improving the coherence and integrity of the fusion result, and laying a solid macro foundation for subsequent layered fusion.

[0032] After obtaining the large-scale land cover information fusion result, the method further includes: using a contour extraction function to identify the contour boundary of each land cover in the large-scale land cover information fusion result; calculating the intersection length between the contour boundary of the land cover and a predefined block line; when the intersection length is greater than a preset threshold, updating the union mask of the land cover corresponding to the contour boundary of the land cover in the large-scale land cover information fusion result; when the intersection length is less than or equal to the preset threshold, determining that the land cover corresponding to the contour boundary of the land cover is complete in the large-scale land cover information fusion result. Specifically, determining that the land cover corresponding to the contour boundary of the land cover is complete in the large-scale land cover information fusion result when the intersection length is less than or equal to the preset threshold is because the contour boundary of the land cover is only slightly affected by the block effect when the intersection length is less than or equal to the preset threshold. When the intersection length exceeds a preset threshold, the system checks whether there are previously recorded retained values ​​in the affected area where the feature corresponding to the outline boundary of the patch is located. These retained values ​​are pixel values ​​of features marked as priority for retention during historical processing. If a retained value exists, the pixel values ​​of the affected area are restored to the retained value to ensure the continuity of key feature boundaries. If no retained value exists, the current pixel value of the affected area is determined as the new retained value, and the modified marking of the affected area is revoked, allowing the affected area to re-enter an unmodified state. This processing flow precisely solves the core problems of broken feature outlines and discontinuous boundaries at block lines in existing technologies. By extracting the outline and determining the intersection length, areas affected by block effects can be quickly located. Updating the union mask and dynamically correcting pixel values ​​using the retained value mechanism not only restores historical key feature boundaries but also reserves space for subsequent fine processing, avoiding boundary distortion. Simultaneously, it directly confirms areas with complete outlines, reducing invalid operations. Ultimately, it significantly weakens block splicing traces, ensuring the smoothness and spatial continuity of large-scale feature boundaries, laying a high-quality foundation for subsequent medium- and small-scale feature fusion.

[0033] Step 203: Based on the processing priority of small-scale features, the small-scale feature data in the third prediction result is fused into the medium-scale feature information fusion result to generate a remote sensing image land cover map of the target area.

[0034] For small-scale features, they are processed sequentially according to the ascending priority of the total pixel area of ​​the land cover type (the processing priority of small-scale features). First, from the third prediction result, all small patches belonging to the current processing type are extracted based on the land cover category attribute, ensuring complete extraction of small-scale features without omission. Then, based on the geographic coordinate system of the original remote sensing image, precise spatial coordinate matching is performed, and these small patches are overlaid and fused with the fusion result of the medium-scale land cover information obtained in step 202 (F_medium), prioritizing the preservation of small patch information to fill the detail gaps in F_medium. After completing the fusion of all small-scale features, the complete remote sensing image semantic segmentation result is output in GeoTIFF format, and then the raster data is converted into vector data using block processing and polygonization algorithms through the GDAL library, finally generating a large-area remote sensing image land cover map.

[0035] Furthermore, this application provides a more specific example of the above method, as follows: This embodiment trains a land cover recognition model based on samples covering multiple categories. The samples consist of 512×512 pixel remote sensing images and corresponding labeled raster maps. Pixel values ​​1 to 11 in the labeled maps represent 11 land cover categories: 1-crops, 2-horticultural fruit trees, 3-forest, 4-grassland, 5-water bodies, 6-buildings, 7-roads, 8-planting facility land, 9-aquaculture facility land, 10-other facility land, 11-other land use. After training, the model weights with the best performance on the validation set are selected, and the obtained optimal model is used to extract and predict land covers from large-area remote sensing images.

[0036] The method proposed in this invention is used to process the prediction results obtained from the land cover identification model. The specific steps performed in this embodiment are as follows: Step 1: Design the block size for large-area remote sensing image classification prediction based on the land feature scale characteristics, and perform land feature classification prediction. According to the proportion of large-area patches in each land feature category to the total number of patches in that land feature category, classify the land features into large-scale, medium-scale, and small-scale land features, and adapt the corresponding processing flow.

[0037] Specifically, a base block size S is first determined as the clipping benchmark. S is determined by using the 512×512 pixel training sample size of the feature recognition model as the basic unit, and considering the GPU computing power limit, generating a set of candidate block sizes (e.g., 512×512 pixels, 1024×1024 pixels, 2048×2048 pixels, 4096×4096 pixels). In a representative validation region, predictions are made and stitched using all candidate sizes. The pixel-level consistency index between the predicted and ground truth regions is calculated, using the manually labeled ground truth region as the criterion. This consistency index uses the Intersection over Union (IOU) value to measure the degree of overlap between the predicted and ground truth regions. Subsequently, analysis is performed. When increasing the block size no longer significantly improves the consistency index (i.e., the change in the IOU is less than a preset change), the corresponding block size is determined as the optimal value S.

[0038] Through experimental analysis, the baseline block size S was determined to be 4096×4096 pixels. Based on this, the original remote sensing image was cropped in multiple sizes to generate three different sizes of sliced ​​remote sensing image sets.

[0039] 1) The original raster image is cropped into 4096×4096 pixel slices. This slice size can better preserve the overall macroscopic features of the ground features and helps to obtain information on the large-scale distribution of ground features.

[0040] 2) The original raster image is cropped to a size of 4096×4096, with rows and columns offset by 2048. This offset cropping method is used to cover the block lines formed by the previous 4096×4096 blocks during the subsequent fusion process, thereby reducing the potential breakage of ground feature information caused by block division and ensuring good continuity of ground feature extraction results at block edges.

[0041] 3) Crop the original raster image into 2048×2048 pixel slices. Small slices help capture local details of ground features and make up for the lack of detail in large slices.

[0042] For the slices obtained from the above three modes, the trained land cover recognition model is used to predict the slices, and the prediction results are stitched together. Finally, the land cover semantic segmentation prediction results of each slice are output, which are denoted as prediction result A (first prediction result, corresponding to the 4096×4096 slice stitching result), prediction result B (second prediction result, corresponding to the 4096×4096 size slice stitching result with a 2048 offset in rows and columns), and prediction result C (third prediction result, corresponding to the 2048×2048 slice stitching result). The offset d=S / 2=2048. This is because only d=2048 has symmetry and complementarity. The boundary of the first segmentation mode is at the center of the block in the second segmentation mode, and the boundary of the second segmentation mode is at the center of the block in the first segmentation mode. Symmetry and complementarity mean that the boundaries of the two segments are in the exact center of each other's blocks. This maximizes the degree of boundary offset and avoids the two predictions being affected by the boundary at the same location.

[0043] Selecting the first 4096×4096 block, the preset results can be displayed using the result images of prediction result A, prediction result B, and prediction result C. In prediction result A, the patches are complete, but the detailed features are fewer than in prediction result C. Prediction result B is similar to prediction result A in overall outline, and this offset cropping method can cover the block lines formed in prediction result A (where different colors represent the types of land features: 1-white for crops, 2-blue-green for horticultural fruit trees, 3-green for forests, 4-light green for grasslands, 5-blue for water bodies, 6-red for buildings, 7-yellow for roads, 8-purple for agricultural facilities, 9-light blue-purple for aquaculture facilities, 10-light purple-orange for other agricultural facilities, and 11-orange for other land uses). Subsequently, the correspondence between the block scale and the feature scale is established. For each category of connected regions in the prediction result A, the area of ​​each individual patch is calculated, and an area threshold T is set. l For 6000 pixels and the percentage threshold P l 30%, percentage threshold P h It is 80%. Area threshold T l To determine whether a single land cover is a large-area land cover, the proportion R of large-area land cover in each land cover category to the total number of land cover in that category is calculated, and the proportion threshold P is used. l and the percentage threshold P h This is used to determine the interval to which the percentage belongs. Then, the corresponding processing flow is matched according to the interval to which the percentage belongs.

[0044] Type X (large-scale ground object): R ≥ 80%, which belongs to the large-scale ground object processing flow, so this ground object category only needs to undergo the processing flow for large-scale ground objects. In prediction result A, the ground types belonging to this category include crops, horticultural fruit trees, forests, grasslands, and buildings.

[0045] Type Y (medium-scale ground object): 30%<R<80%, which belongs to the medium-scale ground object processing flow, so this ground object category not only needs to undergo the processing flow for large-scale ground objects, but also needs to proceed to the next processing flow for small-scale ground objects. In prediction result A, the ground types belonging to this category include water areas and roads.

[0046] Type Z (small-scale ground object): R ≤30%, which belongs to the small-scale ground object processing flow, so it is considered that only the processing flow for small-scale ground objects needs to be performed. In prediction result A, the ground types belonging to this category include planting facility land, aquaculture facility land, other facility land, and other land.

[0047] Step 2: Count the total area of each ground object pixel in the prediction stitching result of the reference block, and sort the corresponding ground object types in the processing flows for ground objects of different scales in ascending order according to the total area of the ground type pixels.

[0048] Specifically, a pixel-level analysis is performed on prediction result A. First, the total area of each ground type in A is counted. For the ground types belonging to Type X (large-scale ground object), Type Y, and Type Z in Step 1, sorting is performed in ascending order according to the calculated total area of the ground types, to obtain sorting result List_X (large-scale processing priority), sorting result List_Y (medium-scale ground object processing priority), and sorting result List_Z (small-scale ground object processing priority), which are respectively: 1) Sorting result List_X: Pixel 3 (forest) < Pixel 1 (crop) < Pixel 4 (grassland) < Pixel 2 (horticultural fruit tree) < Pixel 6 (building). 2) Sorting result List_Y: Pixel 5 (water area) < Pixel 7 (road) 3) Sorting result List_Z: Pixel 10 (other facility land) < Pixel 9 (aquaculture facility land) < Pixel 8 (planting facility land) < Pixel 11 (other land). In this order, the priority of ground type processing is determined, to ensure that small-area ground types are processed first and avoid being covered by large-area ground types.

[0049] Step 3: Input the ground objects classified as belonging to the large-scale ground object processing flow, and according to the ascending sorting result of the ground objects, perform operations such as union region extraction and mask construction, accurate identification of unmodified regions, pixel writing and modification mark update, contour detection and block boundary interactive processing, and global loop control for each ground object category, to obtain the large-scale ground object information fusion result F_large. The details are expanded in the following sub-steps: (1) First, start the pixel loop. Based on the sorting result List_X determined in step 2, process pixel value 3 (forest) first, and then process pixel value 6 (building) in turn.

[0050] (2) Union Region Extraction and Mask Construction. For the pixel value being processed, the union region of it in image A and image B is extracted, and a union mask is created based on the range of this region. The mask will accurately mark the spatial distribution of the pixel value in the union of A and B, providing a spatial positioning basis for subsequent region state judgment and processing operations.

[0051] (3) Accurate identification of unmodified regions. In the fusion result to be generated, the regions that have not been modified by any previous processing steps are accurately identified by searching the modification marker records. Only operating on unmodified regions can avoid the efficiency loss caused by repeated processing, and can retain the effective information in the original image to the maximum extent, preventing feature distortion caused by overprocessing.

[0052] (4) Pixel writing and modification mark update. Write the current pixel value to all unmodified regions in the fusion result (i.e., the intersection of the union region and the unmodified region). Immediately after writing, update the modification mark and record the processing status of these regions. For regions that have been marked as "modified", no further writing operation will be performed, regardless of whether they belong to the union range of the current pixel, to ensure that the key information that has been preserved is not overwritten.

[0053] (5) Contour detection and block boundary interaction processing. The contour extraction function is called to identify the complete contour of the land cover corresponding to the current pixel category, and then the following loop processing is performed on each contour: 1) Block boundary intersection detection: Calculate the intersection length between the contour boundary and the predefined block lines (one every 2048 pixels). The block lines are determined by the boundary positions when the initial image is segmented and cropped.

[0054] 2) Threshold Judgment and Special Processing Trigger: If the intersection length exceeds a set threshold, it indicates that the contour boundary is significantly affected by the block effect and requires processing. In this embodiment, the threshold is set to 30 pixel values. First, check whether there are previously recorded reserved values ​​(i.e., pixel values ​​marked as priority to be retained in historical processing) in the affected area. If there are reserved values, restore the pixel values ​​of the area to the reserved values ​​to ensure the continuity of key feature boundaries; if there are no reserved values, maintain the current pixel values ​​and record them as new reserved values. Second, remove the "modified" mark of the area, allowing it to re-enter the unmodified state, reserving adjustment space for subsequent higher priority or more refined processing steps.

[0055] In prediction result A, due to the use of block-based prediction, the boundary of the grassland patch located at the edge of the block was misidentified as a building patch, resulting in an incomplete patch boundary. In prediction result B, however, since the patch was not located at the block boundary, complete grassland category information was preserved. By performing a union operation on prediction results A and B, and then performing contour detection and pixel value restoration / preservation on the union result, the complementary information in their spatial distribution was effectively integrated, correcting the boundary misclassification problem caused by block prediction, and ultimately obtaining a more complete and continuous patch.

[0056] 3) Contour loop control: After processing a single contour, determine whether there are any unprocessed contours. If so, continue processing the next contour. If all contours have been processed, proceed to the next stage of pixel loop.

[0057] (6) Global loop control. After completing the processing of all current pixel values, record the processing progress and determine if there are any unprocessed pixel values. If so, process the next pixel value in priority order until all pixel values ​​have been processed.

[0058] This step significantly reduces the block effect through meticulous pixel-level looping and boundary interaction processing.

[0059] Step 4: Based on the large-scale information fusion result F_large, perform large-scale basic optimization and small-scale detail enhancement operations on each mesoscale land cover category according to the ascending order of land cover results, update the integrated mesoscale land cover information in F_large, and obtain the mesoscale land cover information fusion result F_medium.

[0060] Based on the large-scale information fusion results, for the land features in the sorted result List_Y from step 2, namely roads and water bodies, a fusion logic of "boundary coordination + detail supplementation" is adopted: (1) Large-scale basic optimization: Following the processing procedure for large-scale features in step 3, the boundary of large-scale patches in this type of feature is corrected to weaken the block effect caused by the dividing line and ensure the continuity of the contiguous area.

[0061] For example, taking pixel value 7 (road) as an example, following the processing flow for large-scale features in step 3, all patches with pixel value 7 in prediction results A and B are extracted. A logical "union" operation is performed on these two sets of patches, and contour detection and pixel value restoration / preservation are performed on the union result. This can alleviate the problems of feature fragmentation and inconsistent boundaries caused by block segmentation, and generate a preliminary fused patch. (2) Small-scale detail enhancement: Extract the land cover patches of this type in the prediction result C, and perform a logical "union" operation with the result after large-scale optimization by matching spatial coordinates.

[0062] In large-scale optimization results, some detailed information may be missed. For example, after large-scale optimization, there may be missing information in some parts of the road. The small-scale prediction result C contains rich details. The result after combining the two successfully fills in the detailed information captured in C while ensuring the continuity of the main body, thus achieving the complementary advantages of the whole and the details.

[0063] Step 5: Based on the mesoscale information fusion result F_medium, perform patch fusion operation for each small-scale land cover category according to the ascending order of land cover results. Update the small-scale land cover information in F_medium to obtain the small-scale land cover information fusion result F_small. Then output the final land cover information fusion result.

[0064] Specifically, for the land feature categories in the sorting result List_Z of step 2, namely, agricultural facility land, aquaculture facility land, other facility land, and other land, the typical characteristic of this type of land feature is that most of its patches are small in area, making them prone to being missed in large-scale patch predictions, while maintaining a relatively complete shape in small-scale patch predictions. First, all land feature patches belonging to this type are extracted from the prediction result C. Then, these small patches from the prediction result C are directly fused with the result optimized in step 4, ultimately forming a land feature fusion result with richer detailed features.

[0065] Finally, after completing the small-scale feature fusion, the semantic segmentation results obtained from the above processing are integrated into a complete raster image and output in GeoTIFF format. A raster-to-vector algorithm is used to convert raster feature patches into vector files. The GDAL library is used for raster-to-vector conversion, and a block-based processing method is adopted to address the need for large file processing. The selected block size is 4096×4096 pixels. This size is chosen to ensure stable program operation, strictly limiting the amount of data processed in a single operation to the limits of the current computer's memory. A polygonization algorithm is then used to convert the pixel information of each block into vector polygons. Finally, the generated vector results are merged and geometrically broken down according to attribute values ​​to form complete polygonal feature vector data.

[0066] Furthermore, as a response to the above Figure 1-2The present invention also provides an apparatus for generating large-area remote sensing image land cover maps, which is used to obtain accurate large-area remote sensing image land cover maps. The embodiments of this apparatus correspond to the foregoing method embodiments. For ease of reading, this embodiment will not repeat the details of the foregoing method embodiments one by one, but it should be clear that the apparatus in this embodiment can correspondingly implement all the contents of the foregoing method embodiments. For example... Figure 3 As shown: The acquisition unit 31 is used to slice the remote sensing image of the target area into blocks using a first block mode, a second block mode, and a third block mode when the area of ​​the target area is larger than a preset area. The block results are then processed into prediction results for land features using a land feature recognition model. The first block mode uses a preset size for segmentation, the second block mode uses a preset size and half of a preset offset for segmentation, and the third block mode uses half of the preset size for segmentation. The first prediction result has processing priorities for large-scale, medium-scale, and small-scale land features. The first block mode is used to preserve the macroscopic features of large-scale land features with a preset size. The second block mode is used to cover the segmentation lines of the first block mode with a preset size and a preset offset. The third block mode is used to capture the detailed features of small-scale land features with a preset ratio size smaller than the preset size. Unit 32 is used to obtain the fusion result of large-scale land cover information from the union region of the first prediction result and the second prediction result based on the processing priority of large-scale land cover in the acquisition unit; The obtaining unit 32 is used to, based on the large-scale ground feature information fusion result and the processing priority of the medium-scale ground features, fuse the medium-scale ground feature data obtained in the union region of the first prediction result and the second prediction result with the medium-scale ground feature data in the third prediction result to obtain the medium-scale ground feature information fusion result. The obtaining unit 32 is used to fuse the small-scale ground feature data in the third prediction result with the fusion result of the medium-scale ground feature information based on the processing priority of small-scale ground features, and generate a remote sensing image land cover map of the target area.

[0067] Furthermore, such as Figure 4 As shown, the preset size is the basic block size. The acquisition unit 31 includes an acquisition module 311, which includes: Acquire the basic block size and the original remote sensing image of the target area; The first segmentation mode is to cut the original remote sensing image of the target area into multiple slice remote sensing images with the same size as the basic segmentation; The second segmentation mode is to cut the original remote sensing image of the target area into multiple slice remote sensing images by cropping it with a basic segmentation size and offsetting it in both row and column directions by half the basic segmentation size. The third segmentation mode is to cut the original remote sensing image into multiple slices of remote sensing image with a size of half that of the basic segmentation size.

[0068] Furthermore, such as Figure 4 As shown, the acquisition unit 31 includes an acquisition module 312, which includes: All sliced ​​remote sensing images obtained under different segmentation modes are input into the land feature recognition model to obtain the first prediction result, the second prediction result and the third prediction result. The sample data of the land feature recognition model consists of remote sensing images with segmentation size determined by the inherent characteristics of land features and corresponding labeled raster maps, wherein the pixel values ​​in the labeled raster maps represent the types of land features.

[0069] Furthermore, such as Figure 4 As shown, the acquisition unit 31 includes a calculation module 313, which includes: Obtain the number of patches in the connected regions for each land cover type in the first prediction result, and calculate the area of ​​a single patch for each land cover type; For each land cover type, obtain the number of individual patches with an area greater than the area threshold and the total number of all patches, and calculate the proportion of large-area patches; When the proportion of large-area patches is greater than or equal to the first threshold, it is determined that the type of land cover belongs to large-scale land cover; When the proportion of large-area patches is less than a first threshold and greater than a second threshold, the type of land cover is determined to be a mesoscale land cover, wherein the first threshold is greater than the second threshold; When the proportion of large-area patches is less than or equal to the second threshold, the type of land cover is determined to be small-scale land cover.

[0070] Furthermore, such as Figure 4 As shown, the unit includes a marking module 321, which includes: Based on the processing priority of large-scale features, features are acquired and pixel values ​​are processed for each feature. The pixel value processing is used to extract the union region of the first prediction result and the second prediction result. A union mask is created based on the union region, and the union mask is used to mark that the union region has been corrected.

[0071] Furthermore, such as Figure 4 As shown, the obtaining unit 32 includes: The contour extraction function is used to identify the contour boundary of each land feature in the large-scale land feature information fusion result; the intersection length of the contour boundary of the land feature with the predefined block line is calculated; when the intersection length is greater than a preset threshold, the union mask of the land feature corresponding to the contour boundary of the land feature in the large-scale land feature information fusion result is updated; when the intersection length is less than or equal to the preset threshold, it is determined that the land feature corresponding to the contour boundary of the land feature has a complete contour in the large-scale land feature information fusion result.

[0072] Furthermore, such as Figure 4 As shown, the obtaining unit 32 includes: When the intersection length is greater than a preset threshold, check whether there is a previously recorded reserved value in the affected area where the feature corresponding to the outline boundary of the patch is located. The reserved value is the pixel value of the feature that was marked as needing to be reserved in the historical processing. If the reserved value exists, the pixel value of the affected area is restored to the reserved value to ensure the continuity of the key feature boundary. If the reserved value does not exist, the current pixel value of the affected area is determined as the new reserved value, and the modified mark of the affected area is revoked, so that the affected area returns to the unmodified state.

[0073] Furthermore, embodiments of this application also provide a computing device, the computing device comprising: at least one processor, and a memory, wherein the memory stores instructions executable by the processor, the instructions being executed by the processor, thereby enabling the processor to perform the above-described operations. Figure 1-2 The method for generating land cover maps from large-area remote sensing imagery described herein.

[0074] Furthermore, embodiments of this application also provide a readable storage medium for storing a computer program, wherein the computer program, when running, controls the device where the storage medium is located to perform the above-described actions. Figure 1-2 The method for generating land cover maps from large-area remote sensing imagery described herein.

[0075] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0076] It is understood that the relevant features in the above methods and apparatus can be referenced interchangeably. Furthermore, the terms "first," "second," etc., in the above embodiments are used to distinguish between embodiments and do not represent the superiority or inferiority of any particular embodiment.

[0077] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0078] The algorithms and displays provided herein are not inherently related to any particular computer, virtual system, or other device. Various general-purpose systems can also be used in conjunction with the teachings herein. The required structure for constructing such systems is apparent from the above description. Furthermore, this invention is not directed to any particular programming language. It should be understood that the contents of the invention described herein can be implemented using various programming languages, and the above description of specific languages ​​is for the purpose of disclosing the best mode of implementation of the invention.

[0079] In addition, the memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0080] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0081] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0082] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0083] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0084] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0085] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0086] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0087] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0088] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The above descriptions are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0089] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for generating land cover maps from large-area remote sensing imagery, characterized in that, include: When the target area is larger than a preset area, the remote sensing image of the target area is sliced ​​into blocks using a first block mode, a second block mode, and a third block mode, respectively. Each block result is then processed into a prediction result for ground features using a ground feature recognition model. Specifically, the first block mode uses a preset size for segmentation, the second block mode uses a preset size and a preset offset for segmentation, and the third block mode uses a preset ratio smaller than the preset size for segmentation. The first prediction result prioritizes the processing of large-scale, medium-scale, and small-scale ground features. The first block mode preserves the macroscopic features of large-scale ground features using the preset size, the second block mode covers the segmentation lines of the first block mode using the preset size and preset offset, and the third block mode captures the detailed features of small-scale ground features using a preset ratio smaller than the preset size. The first block mode corresponds to the first prediction result, the second block mode corresponds to the second prediction result, and the third block mode corresponds to the third prediction result. Wherein, the preset size is the basic block size; the step of processing each block result into a prediction result of the land feature using a land feature recognition model includes: The process involves acquiring the basic block size and the original remote sensing image of the target area; the first block mode is to crop the original remote sensing image of the target area into multiple slice remote sensing images with the same basic block size; the second block mode is to crop the original remote sensing image of the target area with the basic block size and offset the row and column directions by half the basic block size to obtain multiple slice remote sensing images; the third block mode is to crop the original remote sensing image into multiple slice remote sensing images with a size of half the basic block size. All sliced ​​remote sensing images obtained under different segmentation modes are input into the land cover recognition model to obtain the first prediction result, the second prediction result and the third prediction result. The sample data of the land cover recognition model consists of remote sensing images with segmentation size determined by the inherent characteristics of land cover and corresponding labeled raster maps, wherein the pixel values ​​in the labeled raster maps represent the types of land cover. The number of patches in the connected regions for each land cover type in the first prediction result is obtained, and the area of ​​a single patch in each land cover type is calculated. For each land cover type, the number of patches with a single patch area greater than an area threshold and the total number of all patches are obtained, and the proportion of large-area patches is calculated. When the proportion of large-area patches is greater than or equal to a first threshold, the land cover type is determined to be a large-scale land cover. When the proportion of large-area patches is less than the first threshold but greater than a second threshold, the land cover type is determined to be a medium-scale land cover, wherein the first threshold is greater than the second threshold. When the proportion of large-area patches is less than or equal to the second threshold, the land cover type is determined to be a small-scale land cover. Based on the processing priority of large-scale features, the fusion result of large-scale feature information is obtained from the union region of the first prediction result and the second prediction result, including: based on the processing priority of large-scale features, features are obtained and pixel values ​​are processed for each feature, the pixel value processing is used to extract the union region of the first prediction result and the second prediction result; a union mask is created based on the union region, the union mask is used to mark that the union region has been corrected. After obtaining the large-scale land cover information fusion result, the method further includes: using a contour extraction function to identify the contour boundary of each land cover in the large-scale land cover information fusion result; calculating the intersection length between the contour boundary of the land cover and a predefined block line; and updating the union mask of the land cover corresponding to the contour boundary of the land cover in the large-scale land cover information fusion result when the intersection length is greater than a preset threshold. When the intersection length is less than or equal to a preset threshold, it is determined that the land feature corresponding to the outline boundary of the patch is complete in the large-scale land feature information fusion result; When the intersection length is greater than a preset threshold, check whether there is a previously recorded reserved value in the affected area where the feature corresponding to the outline boundary of the patch is located. The reserved value is the pixel value of the feature that was marked as needing to be reserved in the historical processing. If the reserved value exists, the pixel value of the affected area is restored to the reserved value to ensure the continuity of the key feature boundary. If the reserved value does not exist, the current pixel value of the affected area is determined as the new reserved value, and the corrected mark of the affected area is revoked, so that the affected area re-enters the uncorrected state. Based on the large-scale ground feature information fusion results, and based on the processing priority of mesoscale ground features, the mesoscale ground feature data obtained from the union region of the first and second prediction results are fused with the mesoscale ground feature data in the third prediction result to obtain the mesoscale ground feature information fusion results. Based on the processing priority of small-scale features, the small-scale feature data in the third prediction result is fused with the fusion result of the medium-scale feature information to generate a remote sensing image land cover map of the target area.

2. A device for generating land cover maps from large-area remote sensing imagery, characterized in that, include: The acquisition unit is used to slice the remote sensing image of the target area into blocks using a first block mode, a second block mode, and a third block mode when the area of ​​the target area is larger than a preset area. The segmentation results are then processed into prediction results for land features using a land feature recognition model. Specifically, the first block mode uses a preset size for segmentation, the second block mode uses a preset size and a preset offset for segmentation, and the third block mode uses a preset ratio size smaller than the preset size for segmentation. The first prediction result prioritizes the processing of large-scale, medium-scale, and small-scale land features. The first block mode preserves the macroscopic features of large-scale land features using the preset size, the second block mode covers the segmentation lines of the first block mode using the preset size and a preset offset, and the third mode captures the detailed features of small-scale land features using a preset ratio size smaller than the preset size. The first block mode corresponds to the first prediction result, the second block mode corresponds to the second prediction result, and the third block mode corresponds to the third prediction result. The preset size is the basic block size. The acquisition unit includes an acquisition module, which includes: acquiring a basic block size and the original remote sensing image of the target area; a first block mode is to crop the original remote sensing image of the target area into multiple slice remote sensing images with the same size as the basic block; a second block mode is to crop the original remote sensing image of the target area with the basic block size and to divide it into multiple slice remote sensing images with an offset of half the basic block size in both row and column directions; a third block mode is to crop the original remote sensing image into multiple slice remote sensing images with a size of half the basic block size. The acquisition unit includes an acquisition module, which includes: inputting all sliced ​​remote sensing images obtained under different block modes into the land cover recognition model to obtain the first prediction result, the second prediction result and the third prediction result. The sample data of the land cover recognition model consists of remote sensing images with block sizes determined by the inherent characteristics of the land cover and corresponding labeled raster maps, wherein the pixel values ​​in the labeled raster maps represent the types of land cover. The acquisition unit includes a calculation module, which includes: acquiring the number of patches in the connected regions of each land feature type in the first prediction result, and calculating the area of ​​a single patch in each land feature type; for each land feature type, acquiring the number of patches with a single patch area greater than an area threshold and the total number of all patches, and calculating the proportion of large-area patches; when the proportion of large-area patches is greater than or equal to a first threshold, determining that the land feature type belongs to a large-scale land feature; when the proportion of large-area patches is less than the first threshold but greater than a second threshold, determining that the land feature type belongs to a medium-scale land feature, wherein the first threshold is greater than the second threshold; when the proportion of large-area patches is less than or equal to the second threshold, determining that the land feature type belongs to a small-scale land feature; The unit is used to obtain the fusion result of large-scale land cover information from the union region of the first prediction result and the second prediction result based on the processing priority of the large-scale land cover of the unit. The obtaining unit includes a labeling module, which includes: acquiring land features based on the processing priority of large-scale land features and processing the pixel values ​​of each land feature, wherein the pixel value processing is used to extract the union region of the first prediction result and the second prediction result; and creating a union mask based on the union region, wherein the union mask is used to mark that the union region has been corrected. The obtaining unit includes: using a contour extraction function to identify the contour boundary of each land feature in the large-scale land feature information fusion result; calculating the intersection length between the contour boundary of the land feature and a predefined block line; when the intersection length is greater than a preset threshold, updating the union mask of the land feature corresponding to the contour boundary of the land feature in the large-scale land feature information fusion result; when the intersection length is less than or equal to the preset threshold, determining that the land feature corresponding to the contour boundary of the land feature has a complete contour in the large-scale land feature information fusion result; The obtaining unit includes: when the intersection length is greater than a preset threshold, checking whether there is a previously recorded retained value in the affected area where the feature corresponding to the outline boundary of the patch is located, wherein the retained value is the pixel value of the feature marked as needing to be retained in the historical processing; if the retained value exists, restoring the pixel value of the affected area to the retained value to ensure the continuity of the key feature boundary; if the retained value does not exist, determining the current pixel value of the affected area as the new retained value, and canceling the corrected mark of the affected area, so that the affected area re-enters the uncorrected state; The obtaining unit is used to, based on the large-scale ground feature information fusion result and the processing priority of the medium-scale ground features, fuse the medium-scale ground feature data obtained in the union region of the first prediction result and the second prediction result with the medium-scale ground feature data in the third prediction result to obtain the medium-scale ground feature information fusion result. The obtained unit is used to fuse the small-scale ground feature data in the third prediction result with the fusion result of the medium-scale ground feature information based on the processing priority of small-scale ground features, and generate a remote sensing image land cover map of the target area.

3. A storage medium, characterized in that, The storage medium includes a stored program, wherein, when the program is running, it controls the device where the storage medium is located to execute the method for generating a land cover map from a large area of ​​remote sensing imagery as described in claim 1.

4. An electronic device, characterized in that, The device includes at least one processor, at least one memory connected to the processor, and a bus; wherein the processor and the memory communicate with each other through the bus; the processor is used to call program instructions in the memory to execute the method for generating a land cover map of a large area remote sensing image as described in claim 1.

Images