Remote sensing image processing method based on deep learning

Abstract

The present invention discloses a remote sensing image processing method based on deep learning, which relates to the technical field of remote sensing image processing. First, all buildings containing building features are identified from remote sensing images, and the buildings are marked as complete buildings and buildings with occlusion. For buildings with occlusion, they are matched to complete buildings based on similarity matching, and then the complete buildings are used as templates to assist in reconstructing the outline of the occluded buildings. Through an accurate outline matching algorithm, the best correspondence between the occluded buildings and the complete buildings can be ensured, thereby improving the accuracy of the reconstructed outline. Using the outline of the complete building as a template can reduce the reconstruction error caused by the lack of outline information caused by occlusion. The method of directly using outline matching and splicing is simpler in calculation, can speed up the processing speed, and improve the overall efficiency. It can effectively handle the problem of building occlusion in remote sensing images, and improve the application scope and practicality of remote sensing image analysis.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and in particular to a remote sensing image processing method based on deep learning. Background Art

[0002] Deep learning technology has developed rapidly in the field of optical remote sensing target detection, forming a new paradigm for data-driven optical remote sensing image target detection. High-quality remote sensing data has become an important foundation for the research of such algorithms.

[0003] Domestic and foreign research institutions have released a number of optical remote sensing image target detection datasets, laying the foundation for algorithm development. These datasets are developing in the direction of high quality, large scale, and multiple categories.

[0004] Related algorithms have been applied and developed in horizontal box target detection, rotated box target detection, and subdivided detection directions (such as small target detection and fine-grained detection).

[0005] The Chinese invention patent application with publication number CN113033385A discloses a method and system for remote sensing identification of illegal buildings based on deep learning, and discloses in detail the following steps: obtaining remote sensing images of buildings and roads in a ground environment as the original data set, convolving, pooling, and deconvolving the vector data of the remote sensing image, extracting house features to obtain a vector data feature map, vectorizing the house outline pixels, and converting the vector data feature map into a raster image for model training; inputting the raster image into a deep learning network to obtain a building outline training model; inputting the remote sensing image to be tested as a test set into the deep learning network, performing remote sensing image classification on the test set and outputting the results to achieve target recognition of the remote sensing image.

[0006] The above technical solution discloses a technical idea of ​​obtaining the outline of a house by analyzing remote sensing images based on deep learning. However, the problem currently existing in practice is that buildings in some remote sensing images are often obstructed, and the distortion of the building outline output using the above existing technology is more obvious. Summary of the invention

[0007] In order to solve the above technical problems, the present invention provides a remote sensing image processing method based on deep learning. The following technical solutions are adopted:

[0008] The remote sensing image processing method based on deep learning includes the following steps:

[0009] Step 1: Obtain remote sensing images, use convolutional neural networks to extract features from remote sensing images, and classify labels based on feature recognition results;

[0010] Step 2: Mark all image areas containing building features according to the label classification in step 1;

[0011] Step 3, using an edge detection algorithm to perform contour integrity detection on the contours of all image areas containing building features, and classifying and marking the building areas as complete building areas and areas with blocked buildings according to the contour integrity detection results;

[0012] Step 4, extract feature vectors from the area with blocked buildings, and use cosine similarity to calculate the similarity between the blocked building area and all complete buildings;

[0013] Step 5, using the complete building with the highest similarity as the reconstruction template, based on the contour matching algorithm, extracting the contours corresponding to the complete building and the occluded part of the occluded building area as the splicing item;

[0014] Step 6: Use an image stitching algorithm to stitch the stitching item with the outline of the unobstructed area in the area with the obstructed building, and complete the reconstruction of the area with the obstructed building to eliminate the occlusion.

[0015] By adopting the above technical solution, in order to solve the problem of building occlusion in remote sensing images, all buildings containing building features are first identified from the remote sensing images, and the buildings are marked as complete buildings and buildings with occlusions. For the buildings with occlusions, they are matched to the complete buildings based on similarity matching, and then the complete buildings are used as templates to assist in reconstructing the outlines of the occluded buildings.

[0016] Through the precise contour matching algorithm, the best correspondence between the occluded buildings and the complete buildings can be ensured, thereby improving the accuracy of the reconstructed contours.

[0017] Using the outline of a complete building as a template can reduce reconstruction errors caused by missing outline information due to occlusion. Compared with complex deep learning models, directly using outline matching and stitching methods is computationally simpler, which can speed up processing and improve overall efficiency.

[0018] This method can effectively deal with the problem of building occlusion in remote sensing images and improve the application scope and practicality of remote sensing image analysis.

[0019] Accurate building outline information is crucial for urban planning, land use, and disaster assessment. By reconstructing the outlines of obstructed buildings, more complete data support can be provided.

[0020] The method is easy to realize automated processing flow, reduce manual intervention, and reduce operation difficulty and maintenance cost.

[0021] By directly matching and stitching contours, the reconstruction process is more intuitive and the results are more interpretable, making it easier for professionals to analyze and make decisions.

[0022] This method does not depend on a specific remote sensing image type or resolution, has strong adaptability, and can be applied to a variety of remote sensing data sources.

[0023] Accurate contour information provides good basic data for subsequent advanced analysis such as building attribute extraction, change detection, and 3D modeling.

[0024] Optionally, in step 1, the convolutional neural network uses the following formula to extract features and classify remote sensing images:

[0025] ;

[0026] in is the output classification result, is the activation function, and is the weight matrix, and is the bias term, is the input image data.

[0027] By adopting the above technical methods, the features of the sensed images can be extracted and classified efficiently, and the building areas can be quickly identified, which facilitates the subsequent labeling of the building areas.

[0028] Optionally, in step 3, the contour integrity detection method includes the following steps:

[0029] Step 31, using an edge detection algorithm to identify the contour line of the building, and tracing the contour line to form the building outline according to the edge detection result;

[0030] Step 32, evaluating the integrity of the building outline to obtain a building outline integrity evaluation result, wherein the building outline integrity evaluation result includes a complete building outline and an incomplete building outline;

[0031] Step 33: Mark the corresponding building area with a complete building outline as a complete building area, and mark the corresponding building area with an incomplete building outline as an obstructed building area.

[0032] Optionally, in step 32, the integrity evaluation of the building outline includes the following steps:

[0033] Step 321, check the closure of the building outline, check whether the outline forms a closed shape, if the distance between the start point and the end point of the outline is less than a set threshold, or the shape of the outline conforms to the geometric characteristics of a complete building, then the building outline is judged to be closed;

[0034] Step 322, building outline continuity check, by calculating the distances between adjacent points on the outline, if all distances are less than a set distance threshold, the outline is judged to be continuous;

[0035] Step 323: If it is determined that the building outline is closed and continuous, an evaluation result indicating that the building outline is complete is output; otherwise, an evaluation result indicating that the building outline is incomplete is output.

[0036] By adopting the above technical solutions, the following steps are usually required to identify whether the outline of a building in a remote sensing image is complete: Preprocessing, image enhancement: Improve the contrast and clarity of the image to make the outline of the building more obvious. Noise removal: Use filters to remove noise from the image to avoid affecting the recognition of the outline. Building extraction, use deep learning models (such as CNN) to extract features in the image, which help distinguish buildings from other landforms. Apply semantic segmentation techniques, such as Mask R-CNN or U-Net, to identify building pixels in the image. Edge detection, use edge detection algorithms (such as Canny edge detector) to identify the outline of the building, and based on the edge detection results, trace the outline to form a closed building outline.

[0037] Contour integrity assessment mainly includes two aspects: a. Contour closure check, which checks whether the contour line forms a closed shape. If the distance between the start and end points of the contour line is less than a certain threshold, or the shape of the contour line conforms to the geometric characteristics of the building, the contour is considered closed. b. Contour continuity check, which checks whether the contour line is continuous and has no breaks. The distances between adjacent points on the contour line can be calculated. If all distances are less than a certain threshold, the contour is considered continuous.

[0038] Optionally, the formula for calculating the similarity between the blocked building area and all complete buildings using cosine similarity in step 4 is:

[0039] ;

[0040] in and are the feature vectors of occluded buildings and complete buildings, respectively.

[0041] By adopting the above technical solution, feature vectors are extracted for buildings with occlusion. The similarity between the occluded building and the complete building is calculated using cosine similarity. The complete building with the greatest similarity to the occluded building area can be quickly obtained. Since the styles and types of buildings in the remote sensing image area are generally similar, for example, the architectural styles in the same community or village are usually similar, the corresponding unoccluded parts of the complete building obtained by this similarity calculation are used to replace the occluded parts of the occluded building, which can ensure the best correspondence between the occluded building and the complete building, thereby improving the accuracy of the reconstructed contour.

[0042] Optionally, in step 5, the contour matching algorithm Hausdorff distance is used to match the best matching part of the contour of the complete building area and the contour of the obstructed building area, and the corresponding part of the successfully matched complete building contour is cut out.

[0043] Optionally, the contour matching algorithm is Hausdorff distance. The formula for calculating Hausdorff distance is:

[0044] ;

[0045] in is a set of contour points and the contour point set The Hausdorff distance between It is a contour point and contour points The distance between them, sup is the supremum, inf is the infimum, and the smaller the Hausdorff distance, the more similar they are.

[0046] By adopting the above technical solution, it is possible to achieve the best matching part of matching the outline of the complete building area and the outline of the area with occluded buildings, thereby providing a more accurate splicing item material for reconstructing the outline.

[0047] Optionally, step 6 includes the following sub-steps:

[0048] Step 61, extracting the unobstructed part of the outline of the area where the building is obstructed as the main outline splicing item, and intercepting the corresponding obstructed part of the outline of the complete building that is successfully matched as the auxiliary outline splicing item;

[0049] Step 62: Splice the main contour splicing item and the auxiliary contour splicing item to obtain a reconstructed contour of the area where the occluded building exists.

[0050] Optionally, in step 62, the splicing calculation formula is:

[0051]

[0052] in is the reconstructed contour, is the main outline concatenation item, It is a secondary contour splicing item.

[0053] Optionally, step 6 further includes sub-step 63, performing smoothing processing on the reconstructed contour by using Gaussian filtering.

[0054] By adopting the above technical solution, the spliced ​​outline is smoothed to eliminate the mutation at the splicing point. An image filtering algorithm, such as Gaussian filtering, can be used. The information of the complete building can be directly used to reconstruct the outline of the occluded building, which reduces the complexity of the algorithm and may improve the processing speed and accuracy.

[0055] In summary, the present invention includes at least one of the following beneficial technical effects:

[0056] The present invention can provide a remote sensing image processing method based on deep learning, which first identifies all buildings containing building features from remote sensing images, marks the buildings as complete buildings and buildings with occlusions, and for buildings with occlusions, matches them to complete buildings based on similarity matching, and then uses the complete buildings as templates to assist in reconstructing the outlines of the occluded buildings;

[0057] Through the precise contour matching algorithm, the best correspondence between the occluded buildings and the complete buildings can be ensured, thereby improving the accuracy of the reconstructed contours.

[0058] Using the outline of a complete building as a template can reduce reconstruction errors caused by missing outline information due to occlusion. Compared with complex deep learning models, directly using outline matching and stitching methods is computationally simpler, which can speed up processing and improve overall efficiency.

[0059] It can effectively deal with the problem of building occlusion in remote sensing images and improve the application scope and practicality of remote sensing image analysis.

[0060] Accurate building outline information is crucial for urban planning, land use, and disaster assessment. By reconstructing the outlines of obstructed buildings, more complete data support can be provided.

[0061] By directly matching and stitching contours, the reconstruction process is more intuitive and the results are more interpretable, making it easier for professionals to analyze and make decisions.

[0062] Accurate contour information provides good basic data for subsequent advanced analysis such as building attribute extraction, change detection, and 3D modeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1It is a flowchart of a remote sensing image processing method based on deep learning of the present invention; DETAILED DESCRIPTION

[0064] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0065] The embodiment of the present invention discloses a remote sensing image processing method based on deep learning.

[0066] Reference Figure 1 Embodiment 1, a remote sensing image processing method based on deep learning, comprising the following steps:

[0067] Step 1: Obtain remote sensing images, use convolutional neural networks to extract features from remote sensing images, and classify labels based on feature recognition results;

[0068] Step 2: Mark all image areas containing building features according to the label classification in step 1;

[0069] Step 3, using an edge detection algorithm to perform contour integrity detection on the contours of all image areas containing building features, and classifying and marking the building areas as complete building areas and areas with blocked buildings according to the contour integrity detection results;

[0070] Step 4, extract feature vectors from the area with blocked buildings, and use cosine similarity to calculate the similarity between the blocked building area and all complete buildings;

[0071] Step 5, using the complete building with the highest similarity as the reconstruction template, based on the contour matching algorithm, extracting the contours corresponding to the complete building and the occluded part of the occluded building area as the splicing item;

[0072] Step 6: Use an image stitching algorithm to stitch the stitching item with the outline of the unobstructed area in the area with the obstructed building, and complete the reconstruction of the area with the obstructed building to eliminate the occlusion.

[0073] In order to solve the problem of building occlusion in remote sensing images, all buildings containing building features are first identified from remote sensing images, and the buildings are marked as complete buildings and buildings with occlusions. For buildings with occlusions, they are matched to complete buildings based on similarity matching, and then the complete buildings are used as templates to assist in reconstructing the outlines of the occluded buildings.

[0074] Through the precise contour matching algorithm, the best correspondence between the occluded buildings and the complete buildings can be ensured, thereby improving the accuracy of the reconstructed contours.

[0075] Using the outline of a complete building as a template can reduce reconstruction errors caused by missing outline information due to occlusion. Compared with complex deep learning models, directly using outline matching and stitching methods is computationally simpler, which can speed up processing and improve overall efficiency.

[0076] This method can effectively deal with the problem of building occlusion in remote sensing images and improve the application scope and practicality of remote sensing image analysis.

[0077] Accurate building outline information is crucial for urban planning, land use, and disaster assessment. By reconstructing the outlines of obstructed buildings, more complete data support can be provided.

[0078] The method is easy to realize automated processing flow, reduce manual intervention, and reduce operation difficulty and maintenance cost.

[0079] By directly matching and stitching contours, the reconstruction process is more intuitive and the results are more interpretable, making it easier for professionals to analyze and make decisions.

[0080] This method does not depend on a specific remote sensing image type or resolution, has strong adaptability, and can be applied to a variety of remote sensing data sources.

[0081] Accurate contour information provides good basic data for subsequent advanced analysis such as building attribute extraction, change detection, and 3D modeling.

[0082] In Example 2, in step 1, the convolutional neural network uses the following formula to extract features and classify remote sensing images:

[0083] ;

[0084] in is the output classification result, is the activation function, and is the weight matrix, and is the bias term, is the input image data.

[0085] It can efficiently extract and classify features of sensory images, quickly identify building areas, and facilitate subsequent labeling of building areas.

[0086] In embodiment 3, in step 3, the contour integrity detection method comprises the following steps:

[0087] Step 31, using an edge detection algorithm to identify the contour line of the building, and tracing the contour line to form the building outline according to the edge detection result;

[0088] Step 32, evaluating the integrity of the building outline to obtain a building outline integrity evaluation result, wherein the building outline integrity evaluation result includes a complete building outline and an incomplete building outline;

[0089] Step 33: Mark the corresponding building area with a complete building outline as a complete building area, and mark the corresponding building area with an incomplete building outline as an obstructed building area.

[0090] In Example 4, in step 32, the integrity evaluation of the building outline includes the following steps:

[0091] Step 321, check the closure of the building outline, check whether the outline forms a closed shape, if the distance between the start point and the end point of the outline is less than a set threshold, or the shape of the outline conforms to the geometric characteristics of a complete building, then the building outline is judged to be closed;

[0092] Step 322, building outline continuity check, by calculating the distances between adjacent points on the outline, if all distances are less than a set distance threshold, the outline is judged to be continuous;

[0093] Step 323: If it is determined that the building outline is closed and continuous, an evaluation result indicating that the building outline is complete is output; otherwise, an evaluation result indicating that the building outline is incomplete is output.

[0094] Identifying whether the outline of a building in a remote sensing image is complete usually requires the following steps: Preprocessing, Image enhancement: Improve the contrast and clarity of the image to make the outline of the building more obvious. Noise removal: Use filters to remove noise from the image to avoid affecting the recognition of the outline. Building extraction, use deep learning models (such as CNN) to extract features in the image, which help distinguish buildings from other landforms. Apply semantic segmentation techniques, such as Mask R-CNN or U-Net, to identify building pixels in the image. Edge detection, use edge detection algorithms (such as Canny edge detector) to identify the outline of the building, and based on the edge detection results, trace the outline to form a closed building outline.

[0095] Contour integrity assessment mainly includes two aspects: a. Contour closure check, which checks whether the contour line forms a closed shape. If the distance between the start and end points of the contour line is less than a certain threshold, or the shape of the contour line conforms to the geometric characteristics of the building, the contour is considered closed. b. Contour continuity check, which checks whether the contour line is continuous and has no breaks. The distances between adjacent points on the contour line can be calculated. If all distances are less than a certain threshold, the contour is considered continuous.

[0096] In Example 5, the formula for calculating the similarity between the blocked building area and all complete buildings using cosine similarity in step 4 is:

[0097] ;

[0098] in and are the feature vectors of occluded buildings and complete buildings, respectively.

[0099] For buildings with occlusion, feature vectors are extracted. The cosine similarity is used to calculate the similarity between the occluded building and the complete building. The complete building with the greatest similarity to the occluded building area can be quickly obtained. Since the styles and types of buildings in the remote sensing image area are generally similar, for example, the architectural styles of the same community or village are usually similar, the corresponding unoccluded parts of the complete building obtained by this similarity calculation are used to replace the occluded parts of the occluded building, which can ensure the best correspondence between the occluded building and the complete building, thereby improving the accuracy of the reconstructed contour.

[0100] In Example 6, in step 5, the contour matching algorithm Hausdorff distance is used to match the best matching part of the contour of the complete building area and the contour of the area with occluded buildings, and the corresponding part of the successfully matched complete building contour is cut out.

[0101] In Example 7, the contour matching algorithm is the Hausdorff distance, and the formula for calculating the Hausdorff distance is:

[0102] ;

[0103] in is a set of contour points and the contour point set The Hausdorff distance between It is a contour point and contour points The distance between them, sup is the supremum, inf is the infimum, and the smaller the Hausdorff distance, the more similar they are.

[0104] It is possible to achieve the best matching part of matching the outline of the complete building area and the outline of the area with occluded buildings, thereby providing stitching item materials with more accurate reconstructed outlines.

[0105] In Example 8, step 6 includes the following sub-steps:

[0106] Step 61, extracting the unobstructed part of the outline of the area where the building is obstructed as the main outline splicing item, and intercepting the corresponding obstructed part of the outline of the complete building that is successfully matched as the auxiliary outline splicing item;

[0107] Step 62: Splice the main contour splicing item and the auxiliary contour splicing item to obtain a reconstructed contour of the area where the occluded building exists.

[0108] In step 62, the splicing calculation formula is:

[0109]

[0110] in is the reconstructed contour, is the main outline concatenation item, It is a secondary contour splicing item.

[0111] In Example 9, step 6 further includes a sub-step 63 of smoothing the reconstructed contour using a Gaussian filter.

[0112] Smooth the spliced ​​outline to eliminate the mutation at the splicing point. Image filtering algorithms such as Gaussian filtering can be used. The information of the complete building can be directly used to reconstruct the outline of the occluded building, which reduces the complexity of the algorithm and may improve the processing speed and accuracy.

[0113] The following specific embodiments are used to illustrate the implementation principle of the remote sensing image processing method based on deep learning of the present invention:

[0114] A set of high-resolution remote sensing images covering a large area of ​​a city. Our goal is to identify and classify buildings in these images and reconstruct the outlines of buildings that are obscured.

[0115] Acquire remote sensing images, use convolutional neural networks to extract features from remote sensing images, and classify labels based on feature recognition results;

[0116] Through identification, it was found that there were 82 buildings in the area covered by the remote sensing image.

[0117] According to the label classification in step 1, all 82 building image areas containing building features are marked;

[0118] The edge detection algorithm is used to detect the contour integrity of all image areas containing building features. According to the contour integrity detection results, the building areas are classified and marked as complete building areas and areas with blocked buildings. Among them, there are 80 complete building areas, and 2 buildings are blocked. The 82 building areas are numbered from 1 to 82, and the numbers of the blocked buildings are 2 and 10.

[0119] Extract feature vectors from the area with blocked buildings, and use cosine similarity to calculate the similarity between the blocked building area and all complete buildings;

[0120] The complete building with the highest similarity is used as the reconstruction template. Based on the contour matching algorithm, the contours corresponding to the complete building and the occluded part of the area where the occluded building exists are extracted as the splicing item;

[0121] After the calculation is completed, the complete building numbered 3 has the greatest similarity with building numbered 2, and the occluded area of ​​building numbered 2 is located in the southeast corner. The contour of the area corresponding to the southeast corner of the complete building numbered 3 is intercepted as the splicing item; the complete building numbered 10 has the greatest similarity with building numbered 9, and the occluded area of ​​building numbered 10 is located in the southeast corner. The contour of the area corresponding to the southeast corner of the complete building numbered 9 is intercepted as the splicing item;

[0122] An image stitching algorithm is used to stitch the stitching items with the outline of the unobstructed area in the area with occluded buildings, so as to complete the reconstruction of the area with occluded buildings and eliminate the occlusion.

[0123] After splicing, the complete outlines of buildings No. 2 and No. 10 are obtained, completing the reconstruction of the outlines of the obstructed buildings.

[0124] The above are all preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A remote sensing image processing method based on deep learning, characterized in that: The following steps are involved: Step 1: Obtain remote sensing images, use convolutional neural networks to extract features from remote sensing images, and classify labels based on feature recognition results; Step 2: Mark all image areas containing building features according to the label classification in step 1; Step 3, using an edge detection algorithm to perform contour integrity detection on the contours of all image areas containing building features, and classifying and marking the building areas as complete building areas and areas with blocked buildings according to the contour integrity detection results; Step 4, extract feature vectors from the area with blocked buildings, and use cosine similarity to calculate the similarity between the blocked building area and all complete buildings; Step 5, using the complete building with the highest similarity as the reconstruction template, based on the contour matching algorithm, extracting the contours corresponding to the complete building and the occluded part of the occluded building area as the splicing item; Step 6: Use an image stitching algorithm to stitch the stitching item with the outline of the unobstructed area in the area with the obstructed building, and complete the reconstruction of the area with the obstructed building to eliminate the occlusion.

2. The remote sensing image processing method based on deep learning according to claim 1, characterized in that: In step 1, the convolutional neural network uses the following formula to extract features and classify remote sensing images: ； in is the output classification result, is the activation function, and is the weight matrix, and is the bias term, is the input image data.

3. The remote sensing image processing method based on deep learning according to claim 2, characterized in that: In step 3, the contour integrity detection method includes the following steps: Step 31, using an edge detection algorithm to identify the contour line of the building, and tracing the contour line to form the building outline according to the edge detection result; Step 32, evaluating the integrity of the building outline to obtain a building outline integrity evaluation result, wherein the building outline integrity evaluation result includes a complete building outline and an incomplete building outline; Step 33: Mark the corresponding building area with a complete building outline as a complete building area, and mark the corresponding building area with an incomplete building outline as an obstructed building area.

4. The remote sensing image processing method based on deep learning according to claim 3, characterized in that: In step 32, the integrity assessment of the building outline includes the following steps: Step 321, check the closure of the building outline, check whether the outline forms a closed shape, if the distance between the start point and the end point of the outline is less than a set threshold, or the shape of the outline conforms to the geometric characteristics of a complete building, then the building outline is judged to be closed; Step 322, building outline continuity check, by calculating the distances between adjacent points on the outline, if all distances are less than a set distance threshold, the outline is judged to be continuous; Step 323: If it is determined that the building outline is closed and continuous, an evaluation result indicating that the building outline is complete is output; otherwise, an evaluation result indicating that the building outline is incomplete is output.

5. The remote sensing image processing method based on deep learning according to claim 4, characterized in that: The formula for calculating the similarity between the blocked building area and all complete buildings using cosine similarity in step 4 is: ； in and are the feature vectors of occluded buildings and complete buildings, respectively.

6. The remote sensing image processing method based on deep learning according to claim 5, characterized in that: In step 5, the contour matching algorithm Hausdorff distance is used to match the best matching part of the contour of the complete building area and the contour of the obstructed building area, and the corresponding part of the successfully matched complete building contour is cut out.

7. The remote sensing image processing method based on deep learning according to claim 6, characterized in that: The contour matching algorithm is Hausdorff distance, and the formula for calculating Hausdorff distance is: ； in is a set of contour points and the contour point set The Hausdorff distance between It is a contour point and contour points The distance between them, sup is the supremum, inf is the infimum, and the smaller the Hausdorff distance, the more similar they are.

8. The remote sensing image processing method based on deep learning according to claim 7, characterized in that: Step 6 includes the following sub-steps: Step 61, extracting the unobstructed part of the outline of the area where the building is obstructed as the main outline splicing item, and intercepting the corresponding obstructed part of the outline of the complete building that is successfully matched as the auxiliary outline splicing item; Step 62: The main contour stitching item and the auxiliary contour stitching item are stitched together to obtain a reconstructed contour of the area where the occluded building exists.

9. The remote sensing image processing method based on deep learning according to claim 8, characterized in that: In step 62, the splicing calculation formula is: in is the reconstructed contour, is the main outline concatenation item, It is a secondary contour splicing item.

10. The remote sensing image processing method based on deep learning according to claim 9, characterized in that: Step 6 also includes a sub-step 63 of smoothing the reconstructed contour using a Gaussian filter.

Images