Land class information surveying and mapping system based on remote sensing interpretation

Abstract

The invention provides a land type information surveying and mapping system based on remote sensing interpretation, and the system comprises an image processing module which is used for obtaining a high-resolution remote sensing image and carrying out the superpixel segmentation preprocessing; the land class extraction module is used for capturing texture features of land classes and identifying initial edge features, and associating the texture features and the initial edge features to obtain an initial land class region block; the land class classification module is used for dividing the initial land class region blocks into two classes of land class edge boundary integrity and land class edge boundary incompleteness, performing edge boundary reinforcement and edge boundary repair on the initial land class region blocks, then performing reinforcement, outputting land class repair region blocks and performing scene classification; and the land class integration module is used for fusing the classified land class restoration region blocks into a land class map capable of accurately reflecting land cover land class types. According to the method, different land types are processed respectively and fused into the land map capable of accurately reflecting the land cover land type, and an important practical basis is provided for subsequent land utilization.

Description

Technical Field

[0001] The present invention relates to the field of image recognition technology, and in particular to a land type information surveying and mapping system based on remote sensing interpretation. Background Art

[0002] High-resolution remote sensing images are one of the important data sources in land survey work. They can provide rich and effective surface information, which is conducive to the accurate and detailed interpretation of natural resource land types by indoor workers. High-resolution remote sensing images have high spatial resolution and can obtain the details and texture information of objects, which provides favorable conditions for automatic interpretation. However, traditional high-resolution remote sensing image automatic interpretation methods have weak ability to extract and interpret visual features. Generally, they need to manually design features based on a large amount of prior knowledge, and the completeness of the features cannot be guaranteed. Therefore, the existing high-resolution remote sensing image automatic interpretation methods have weak generalization ability when carrying out interpretation tasks for images covering a large range of study areas, and it is difficult to systematically meet the business application needs of land type information extraction. Summary of the invention

[0003] In view of this, the purpose of the present invention is to provide a land type information mapping system based on remote sensing interpretation to solve the technical problem that it is difficult to accurately identify and extract land types in a large area based on remote sensing images. The system includes:

[0004] The image processing module is used to obtain high-resolution remote sensing images and use super-pixel segmentation algorithm to perform segmentation preprocessing on the high-resolution remote sensing images to obtain super-pixel blocks composed of multiple pixel points with spectral and spatial similarities;

[0005] The land type extraction module is used to capture the texture features of the land type in the multispectral image through the semantic segmentation algorithm, and then identify the initial edge features of the land type through the edge detection algorithm, and associate the texture features of the land type with the edge features to obtain the initial land type area block;

[0006] The land classification module is used to classify the initial land class area blocks into two categories: those with complete land class edge boundaries and those with incomplete land class edge boundaries. The land class area blocks with complete land class edge boundaries are reinforced, and the land class area blocks with incomplete land class edge boundaries are reinforced after repairing the edge boundaries. The land class repaired area blocks are output and the scene classification is performed.

[0007] The land class integration module merges the classified land restoration area blocks into a land class map that can accurately reflect the land cover types on the surface.

[0008] Furthermore, the superpixel segmentation algorithm is used to perform segmentation preprocessing on the high-resolution remote sensing image, including:

[0009] S11, performing superpixel segmentation on the high-resolution remote sensing image, fusing the over-segmented small superpixel blocks with the adjacent superpixel blocks generated by normal segmentation, to obtain multiple superpixel blocks;

[0010] S12, selecting a seed node, and finding all other nodes in the superpixel block where the selected seed node is located as the neighboring nodes of the seed node;

[0011] S13, connecting the seed node and the neighboring node, constructing the seed adjacent edge, obtaining the adjacent weight of the seed adjacent edge and setting a threshold value, which is used to filter the edge with an adjacent weight value less than the threshold;

[0012] S14, calculating the adjacency weight of each node in the seed adjacent edge according to step S12 and step S13, concatenating all nodes of the seed adjacent edge to obtain the adjacency matrix of the seed adjacent edge, that is, obtaining the reconstructed high-resolution remote sensing image.

[0013] Furthermore, the texture features of the land are extracted through the semantic segmentation algorithm, including:

[0014] S21, labeling each superpixel block and performing normalization processing;

[0015] S22, extracting color features by calculating the average value of the LAB color space of all pixels in each superpixel block, and extracting texture features in each superpixel block by using the gray level co-occurrence matrix;

[0016] S23, using a deep learning model to learn high-level semantic features of superpixel blocks, counting the number of pixels of each category in each superpixel block, and assigning the category with the largest number of pixels to the entire superpixel block;

[0017] S24. After fusing the high-level semantic features extracted by the deep learning model with the color features and texture features of the superpixels, each superpixel block is classified through the softmax function to generate a segmentation result map with the same size as the input high-resolution remote sensing image.

[0018] Furthermore, the initial edge features of the land type are identified through edge detection algorithms, including:

[0019] S31, randomly selecting a pixel point in the unknown category within the superpixel block area as the central initial pixel point, and finding all pixels that fall within a preset bandwidth of the central initial pixel point;

[0020] S32, superimposing the vectors from all the pixels to the central initial point in step S31 to obtain an offset vector, calculating the modulus of the offset vector and taking it as the moving distance, and moving the central initial pixel point according to the offset vector and the moving distance;

[0021] S33, iterating step S31 and step S32 until all pixels are divided, counting the access frequency of each category to each pixel, and the category with the largest access frequency is the category described in the current point set;

[0022] S34, calculating the grayscale gradient in the neighborhood around the class with the highest access frequency in the superpixel block to detect the edge, connecting the detected edge points into a complete edge contour, and combining it with the high-level semantic features extracted in step S23.

[0023] Furthermore, the boundary of the land area blocks with complete boundary is strengthened, including:

[0024] S41, using an edge extraction operator based on Gaussian filtering optimization to preliminarily identify the edge contour of the target land area block, and then performing deep convolution feature fusion with the high-level semantic features extracted in step S23 to suppress image noise;

[0025] S42, using the upsampling of bilinear interpolation method and the attention mechanism to adaptively increase the weight of edge information, strengthen the detailed features of the edge contour of the land area block, and obtain the refined edge boundary;

[0026] S43, replacing the initial edge boundary with the refined edge boundary, re-associating it with the texture feature, and obtaining a land restoration area block.

[0027] Furthermore, the edge boundaries of the land area blocks with incomplete land boundary are repaired and strengthened, including:

[0028] S51, filling edge information into the land area blocks with incomplete land boundary to construct an edge guidance map, so as to provide structural guidance for the restoration of edge boundary restoration;

[0029] S52, decomposing the land area block with incomplete land edge boundaries into a high-level detail map and a low-level incomplete map, and constructing an edge repair module, using the edge guide map as the input of the content repair module, and repairing the low-level incomplete map to obtain an initial edge repair map;

[0030] S53, upsampling the initial edge repair map, combining it with the high-level detail map to obtain an intermediate edge repair map, and using it together with the edge guide map as input to the content repair module to complete the repair and optimization of the land area blocks with incomplete land boundary boundaries, and obtain an edge repair map;

[0031] S54. Repeat steps S41 to S43 for the edge restoration until a land restoration area block is obtained.

[0032] Furthermore, the classified land restoration area blocks are merged into a land type map that can accurately reflect the land type of the surface, including:

[0033] S61. Design an edge perception module to emphasize the edge information in the land restoration area block to maintain the clarity and accuracy of the land boundary during the fusion process;

[0034] S62, integrating the high-level semantic features extracted by the deep learning model with the texture and edge features of the land type;

[0035] S63. Import the integrated results into GIS software for mapping to generate the final land type map.

[0036] Furthermore, the scene classification in the land classification module includes building areas, water areas, cultivated land areas, forest areas and grassland areas.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] The present invention provides a land class information mapping system based on remote sensing interpretation. By performing superpixel segmentation on high-resolution remote sensing images, the discrimination accuracy for subsequent image processing of the system is improved, thereby improving the recognition efficiency of semantic segmentation and edge detection algorithms in a land class extraction module; in a land class classification module, the initial land class area blocks are divided into two categories: land class edge boundary complete and land class edge boundary incomplete. By respectively performing edge boundary reinforcement and edge boundary repair and then reinforcement processing, the refined extraction of different land class forms is achieved, thereby ensuring the accuracy of land class scene classification; in a land class integration module, the land classes with good scene classification are merged into a land class map that can accurately reflect the land class types covered on the surface, thereby providing an important practical basis for subsequent further land utilization. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only preferred embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0040] Figure 1 It is a schematic diagram of the overall structure of a land information surveying and mapping system based on remote sensing interpretation provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0041] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The enumerated embodiments are only used to explain the present invention and are not used to limit the scope of the present invention.

[0042] Reference Figure 1 This embodiment provides a land information surveying and mapping system based on remote sensing interpretation, the system comprising:

[0043] The image processing module is used to obtain high-resolution remote sensing images, and uses superpixel segmentation algorithms to perform segmentation preprocessing on high-resolution remote sensing images to obtain superpixel blocks composed of multiple pixel points with spectral and spatial similarities. In order to ensure the accuracy of land type map extraction in high-resolution remote sensing images, superpixel reconstruction is performed on high-resolution remote sensing images to obtain superpixel blocks composed of multiple pixel points with spectral and spatial similarities. The accuracy of extracting features from superpixel blocks can be improved, such as shape, spectrum, texture, etc., and these features help provide high-quality segmentation preprocessing results for subsequent image task processing.

[0044] The superpixel segmentation algorithm is used to perform segmentation preprocessing on high-resolution remote sensing images, including:

[0045] S11. Perform superpixel segmentation on the high-resolution remote sensing image, and fuse the over-segmented small superpixel blocks with the adjacent superpixel blocks generated by normal segmentation to obtain multiple superpixel blocks.

[0046] S12: Select a seed node, and find all other nodes in the superpixel block where the selected seed node is located as the neighboring nodes of the seed node.

[0047] S13, connecting the seed node and the neighboring nodes, constructing the seed adjacent edge, obtaining the adjacent weight of the seed adjacent edge and setting a threshold, which is used to filter the edges with adjacent weights less than the threshold.

[0048] S14, calculating the adjacency weight of each node in the seed adjacent edge according to step S12 and step S13, concatenating all nodes of the seed adjacent edge to obtain the adjacency matrix of the seed adjacent edge, that is, obtaining the reconstructed high-resolution remote sensing image.

[0049] In this embodiment, superpixel segmentation is performed, and the purpose of selecting seed nodes is to initialize the clustering process and ensure that the segmentation algorithm can be carried out effectively. Seed nodes are usually evenly distributed in the image and serve as the initial center of clustering, which helps to ensure that the final superpixel is evenly covered in the image and no area is over-segmented or ignored. The superpixel segmentation algorithm starts the iterative clustering process through these initial centers, that is, since each pixel point will be searched by multiple seed nodes, each pixel point will have a distance from the surrounding seed nodes, and the seed node corresponding to the minimum value is taken as the clustering center of the pixel point. By reselecting or optimizing the positions of these nodes in the local neighborhood of the seed nodes, the segmentation effect in the edge and texture areas of the image can be improved, providing high-quality data input for the semantic segmentation algorithm and edge detection algorithm in the next land type extraction module.

[0050] The land type extraction module is used to capture the texture features of land types in multispectral images through semantic segmentation algorithms, and then identify the initial edge features of land types through edge detection algorithms, and associate the texture features of land types with edge features to obtain the initial land type area blocks. Texture features describe the local arrangement pattern of pixels in an image, while edge features describe the location of brightness changes in an image. Combining texture features and edge features can improve the accuracy of boundary area segmentation of land type restoration area blocks in multispectral images.

[0051] The texture features of the land are extracted through the semantic segmentation algorithm, including:

[0052] S21. Label each superpixel block and perform normalization.

[0053] S22. The color features are extracted by calculating the average value of the LAB color space of all pixels in each superpixel block, and the texture features in each superpixel block are extracted by using the gray-level co-occurrence matrix.

[0054] S23. Use a deep learning model to learn the high-level semantic features of superpixel blocks, count the number of pixels of each category in each superpixel block, and assign the category with the largest number of pixels to the entire superpixel block.

[0055] S24. After fusing the high-level semantic features extracted by the deep learning model with the color features and texture features of the superpixels, each superpixel block is classified through the softmax function to generate a segmentation result map with the same size as the input high-resolution remote sensing image.

[0056] In this embodiment, firstly, the super pixel block is labeled and normalized to reduce the complexity of subsequent data processing. By respectively obtaining the color features, texture features and high-level semantic features in the high-resolution remote sensing image, different land classes can be distinguished more accurately in the land class boundary area, so that the segmentation result is more consistent with the boundary of the actual object. Different land classes usually have different color features. The LAB color space can better reflect the human eye's perception of color. By calculating the LAB average value of all pixels in the super pixel block, the color information of a certain area can be obtained, which helps to distinguish different land classes. In the analysis of the land class repair area block, the grayscale co-occurrence matrix is ​​a method for describing the texture features of the image. By analyzing the grayscale relationship between pixels to extract texture information, it helps to distinguish different land classes and identify surface cover changes. In the boundary area of ​​the land class repair area block, the deep learning model can capture subtle semantic differences, which is used to further improve the accuracy of image segmentation. The number of pixels of each category in each super pixel is counted, and the category with the largest number of pixels is assigned to the entire super pixel to retain the texture details of the image.

[0057] The initial edge features of land types are identified through edge detection algorithms, including:

[0058] S31. Randomly select a pixel point from the unknown category in the superpixel block area as the central initial pixel point, and find out all the pixels that fall within the preset bandwidth of the central initial pixel point.

[0059] S32, superimposing the vectors from all the pixels to the center initial point in step S31 to obtain an offset vector, calculating the modulus of the offset vector and taking it as the moving distance, and moving the center initial pixel point according to the offset vector and the moving distance.

[0060] S33, iterating step S31 and step S32 until all pixels are divided, counting the access frequency of each category to each pixel, and the category with the highest access frequency is the category described in the current point set.

[0061] S34, calculating the grayscale gradient in the neighborhood around the class with the highest access frequency in the superpixel block to detect the edge, connecting the detected edge points into a complete edge contour, and combining it with the high-level semantic features extracted in step S23.

[0062] In this embodiment, based on high-resolution optical remote sensing images, human vision is often more sensitive to boundary information perception, and has a certain degree of recognition even if there are fuzzy boundaries between adjacent land types. The attribution of pixels in the boundary area between adjacent objects in remote sensing images is the key to the classification of fuzzy boundary objects. When performing edge contour division, the present invention does not focus on category information to perform edge contour division, but determines the number of categories through iteration, and completes the edge contour division based on the number of categories, which can avoid the subsequent classification accuracy reduction due to inconsistent category division standards.

[0063] The land classification module is used to divide the initial land class area blocks into two categories: complete land class edge boundaries and incomplete land class edge boundaries. The land class area blocks with complete land class edge boundaries are strengthened, and the land class area blocks with incomplete land class edge boundaries are strengthened after repairing the edge boundaries. The land class repair area blocks are output and the scene classification is performed. The strength of the land class edge boundaries in the initial land class area blocks roughly extracted by the land class extraction module is not consistent. Some land class edge boundaries are complete, and some land class edge boundaries are incomplete. If the edge boundaries of both are directly strengthened at the same time, the accuracy of the final output land class map will decrease, and the edge boundaries will still be unclear. Therefore, the two are first distinguished, and then processed separately, and the land class repair area blocks are output. Scene classification is performed based on the repaired land class repair area blocks to improve the classification accuracy.

[0064] Strengthen the boundary of land area blocks with complete boundary, including:

[0065] S41, using the edge extraction operator based on Gaussian filter optimization to preliminarily identify the edge contour of the target land area block, and then perform deep convolution feature fusion with the high-level semantic features extracted in step S23 to suppress noise. The noise in the initial land area block is generally approximately Gaussian distributed, so a Gaussian filter is used to suppress the noise, and the performance of the edge extraction operator is improved by adjusting the parameters of the Gaussian filter to retain more target edge feature information.

[0066] S42. Adopt upsampling of bilinear interpolation method and attention mechanism to adaptively increase the weight of edge information, strengthen the detailed features of the edge contour of land-type area blocks, and obtain refined edge boundaries. When upsampling is enlarged back to the original image size, it is easy to cause edge blurring. The upsampling of bilinear interpolation method takes into account the correlation influence of the four neighboring points around the sampling point to the sampling point, improves the continuity of the image edge, and thus reduces edge blurring. After sampling, more detailed features in the image edge are retained. Combined with the attention mechanism, the feature channels related to image information are retained or enhanced, and the feature channels related to noise information are removed or weakened, so as to enhance the aggregation ability of edge boundary speckle and the ability to resist background interference, thereby enhancing the strengthening of the edge boundary of land-type area blocks under complex backgrounds.

[0067] S43, replacing the initial edge boundary with the refined edge boundary, re-associating it with the texture feature, and obtaining a land restoration area block.

[0068] The edge boundaries of land area blocks with incomplete land boundaries are repaired and strengthened, including:

[0069] S51. Fill in edge information for land area blocks with incomplete land boundary boundaries to construct an edge guidance map, so as to provide structural guidance for edge boundary repair.

[0070] S52, decompose the land area block with incomplete land edge boundaries into a high-level detail map and a low-level incomplete map, and construct an edge repair module, use the edge guide map as the input of the content repair module, and repair the low-level incomplete map to obtain an initial edge repair map. The low-level incomplete map has a larger receptive field and can obtain more global information, while the high-level detail map contains more detailed information.

[0071] S53, upsampling the initial edge repair map, combining it with the high-level detail map to obtain an intermediate edge repair map, and using it together with the edge guide map as the input of the content repair module to complete the repair and optimization of the land area blocks with incomplete land boundary boundaries, and obtain an edge repair map. Through progressive repair, the edge repair map can obtain rich detail information and global information, thereby generating a high-quality repair result.

[0072] S54. Repeat steps S41 to S43 for the edge restoration until a land restoration area block is obtained.

[0073] The scene classification in the land classification module includes building area, water area, cultivated land area, forest area and grassland area. In this embodiment, the scene is classified into building area, water area, cultivated land area, forest area and grassland area. The accuracy of the land restoration area block obtained by the above image processing can adapt to different classification standards.

[0074] The land type integration module integrates the classified land type restoration area blocks into a land type map that can accurately reflect the land type of the surface. It can reflect the land type of the surface and provide important information support for land management and planning.

[0075] The classified land restoration area blocks are merged into a land type map that can accurately reflect the land type of the surface, including:

[0076] S61. Design an edge perception module to emphasize the edge information in the land restoration area block, so as to maintain the clarity and accuracy of the land boundary during the fusion process. It can provide richer context information and enhance the ability to express different land features.

[0077] S62, integrating the high-level semantic features extracted by the deep learning model with the texture and edge features of the land class. By integrating the high-level semantic features of the deep learning model with the texture and edge features of the land class, different land classes in the land class boundary area can be distinguished more accurately.

[0078] S63, importing the integration results into GIS software for mapping to generate a final land class map. Using GIS software, these area blocks are merged according to geographical location, and different symbols, colors and patterns are assigned to different land classes, so as to clearly distinguish each land class on the final land class map.

[0079] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A land information surveying and mapping system based on remote sensing interpretation, characterized in that: The system comprises: The image processing module is used to obtain high-resolution remote sensing images and use super-pixel segmentation algorithm to perform segmentation preprocessing on the high-resolution remote sensing images to obtain super-pixel blocks composed of multiple pixel points with spectral and spatial similarities; The land type extraction module is used to capture the texture features of the land type in the multispectral image through the semantic segmentation algorithm, and then identify the initial edge features of the land type through the edge detection algorithm, and associate the texture features of the land type with the edge features to obtain the initial land type area block; The land classification module is used to classify the initial land class area blocks into two categories: those with complete land class edge boundaries and those with incomplete land class edge boundaries. The land class area blocks with complete land class edge boundaries are reinforced, and the land class area blocks with incomplete land class edge boundaries are reinforced after repairing the edge boundaries. The land class repaired area blocks are output and the scene classification is performed. The land class integration module merges the classified land restoration area blocks into a land class map that can accurately reflect the land cover types on the surface.

2. A land information surveying and mapping system based on remote sensing interpretation according to claim 1, characterized in that: The superpixel segmentation algorithm is used to perform segmentation preprocessing on high-resolution remote sensing images, including: S11, performing superpixel segmentation on the high-resolution remote sensing image, fusing the over-segmented small superpixel blocks with the adjacent superpixel blocks generated by normal segmentation, to obtain multiple superpixel blocks; S12, selecting a seed node, and finding all other nodes in the superpixel block where the selected seed node is located as the neighboring nodes of the seed node; S13, connecting the seed node and the neighboring node, constructing the seed adjacent edge, obtaining the adjacent weight of the seed adjacent edge and setting a threshold value, which is used to filter the edge with an adjacent weight value less than the threshold; S14, calculating the adjacency weight of each node in the seed adjacent edge according to step S12 and step S13, concatenating all nodes of the seed adjacent edge to obtain the adjacency matrix of the seed adjacent edge, that is, obtaining the reconstructed high-resolution remote sensing image.

3. A land information surveying and mapping system based on remote sensing interpretation according to claim 2, characterized in that: The texture features of the land are extracted through the semantic segmentation algorithm, including: S21, labeling each superpixel block and performing normalization processing; S22, extracting color features by calculating the average value of the LAB color space of all pixels in each superpixel block, and extracting texture features in each superpixel block by using the gray level co-occurrence matrix; S23, using a deep learning model to learn high-level semantic features of superpixel blocks, counting the number of pixels of each category in each superpixel block, and assigning the category with the largest number of pixels to the entire superpixel block; S24. After fusing the high-level semantic features extracted by the deep learning model with the color features and texture features of the superpixels, each superpixel block is classified through the softmax function to generate a segmentation result map with the same size as the input high-resolution remote sensing image.

4. A land information surveying and mapping system based on remote sensing interpretation according to claim 3, characterized in that: The initial edge features of land types are identified through edge detection algorithms, including: S31, randomly selecting a pixel point in the unknown category within the superpixel block area as the central initial pixel point, and finding all pixels that fall within a preset bandwidth of the central initial pixel point; S32, superimposing the vectors from all the pixels to the central initial point in step S31 to obtain an offset vector, calculating the modulus of the offset vector and taking it as the moving distance, and moving the central initial pixel point according to the offset vector and the moving distance; S33, iterating step S31 and step S32 until all pixels are divided, counting the access frequency of each category to each pixel, and the category with the largest access frequency is the category described in the current point set; S34, calculating the grayscale gradient in the neighborhood around the class with the highest access frequency in the superpixel block to detect the edge, connecting the detected edge points into a complete edge contour, and combining it with the high-level semantic features extracted in step S23.

5. The land information surveying and mapping system based on remote sensing interpretation according to claim 3 is characterized in that: Strengthen the boundary of land area blocks with complete boundary, including: S41, using an edge extraction operator based on Gaussian filtering optimization to preliminarily identify the edge contour of the target land area block, and then performing deep convolution feature fusion with the high-level semantic features extracted in step S23 to suppress image noise; S42, using the upsampling of bilinear interpolation method and the attention mechanism to adaptively increase the weight of edge information, strengthen the detailed features of the edge contour of the land area block, and obtain the refined edge boundary; S43, replacing the initial edge boundary with the refined edge boundary, re-associating it with the texture feature, and obtaining a land restoration area block.

6. A land information surveying and mapping system based on remote sensing interpretation according to claim 5, characterized in that: The edge boundaries of land area blocks with incomplete land boundaries are repaired and strengthened, including: S51, filling edge information into the land area blocks with incomplete land boundary to construct an edge guidance map, so as to provide structural guidance for the restoration of edge boundary restoration; S52, decomposing the land area block with incomplete land edge boundaries into a high-level detail map and a low-level incomplete map, and constructing an edge repair module, using the edge guide map as the input of the content repair module, and repairing the low-level incomplete map to obtain an initial edge repair map; S53, upsampling the initial edge repair map, combining it with the high-level detail map to obtain an intermediate edge repair map, and using it together with the edge guide map as input to the content repair module to complete the repair and optimization of the land area blocks with incomplete land boundary boundaries, and obtain an edge repair map; S54. Repeat steps S41 to S43 for the edge restoration until a land restoration area block is obtained.

7. The land information surveying and mapping system based on remote sensing interpretation according to claim 3 is characterized in that: The classified land restoration area blocks are merged into a land type map that can accurately reflect the land type of the surface, including: S61. Design an edge perception module to emphasize the edge information in the land restoration area block to maintain the clarity and accuracy of the land boundary during the fusion process; S62, integrating the high-level semantic features extracted by the deep learning model with the texture and edge features of the land type; S63. Import the integrated results into GIS software for mapping to generate the final land type map.

8. The land information surveying and mapping system based on remote sensing interpretation according to claim 1 is characterized in that: The scene classification in the land classification module includes building areas, water areas, cultivated land areas, forest areas and grassland areas.

Images