A tracking-based high-resolution remote sensing image edge repairing method

Abstract

A high-resolution remote sensing image edge repairing method based on tracking, first, a deep convolutional neural network is designed according to the remote sensing edge detection task, samples are made and edge detection model and edge tracking model are trained; the remote sensing image is input into the edge detection model to extract the ground feature edge, so as to obtain the ground feature edge intensity map. The remote sensing image and the random starting point file are input into the edge tracking model to iteratively generate the edge tracking graph, and for the discontinuous ground feature edge, the edge detection result and the edge tracking result are fused to automatically determine the edge width and complete the edge. The method completes the broken line repairing work through the topological information provided by the edge tracking algorithm for the discontinuous edge part in the prediction result of the edge detection algorithm, and simultaneously uses the edge binary graph to assist in connecting the broken edge tracking graph, or creates an edge in the position where the edge does not exist to complete the edge tracking graph. The method can improve the edge extraction accuracy and efficiency, and improve the accuracy of the edge tracking model.

Description

Technical Field

[0001] This invention relates to remote sensing image processing technology and remote sensing image information extraction methods. Specifically, it is a method that combines remote sensing image edge detection with edge tracking and line break repair methods. Background Technology

[0002] In the field of remote sensing, extracting ground feature boundaries using high-resolution aerial imagery is a research hotspot. This is because manually drawing ground feature boundaries is both laborious and expensive, and these boundaries constantly change over time. Deep learning networks, on the other hand, can quickly and accurately identify ground feature boundaries, effectively meeting real-world needs.

[0003] In the past, many methods for extracting ground object boundaries have been proposed from different perspectives, generally falling into two categories: line-based partitioning and surface-based aggregation. Traditional methods such as belief propagation and graph cutting assume that the spatial relationships of objects depend only on their relative positions in a static state, thus defining billions of edges. Conditional Random Fields (CRFs), on the other hand, use fully connected layers to define edge potentials to capture spatial information while preserving object boundaries. This inference method converges in just a few seconds on standard pixel images. With the widespread application of deep learning in computer vision, new concepts and methods such as semantic segmentation, instance segmentation, and panoptic segmentation have brought entirely new perspectives to remote sensing image segmentation schemes. Convolutional Neural Networks (CNNs) have produced significant results in visual semantic segmentation. Fully Convolutional Neural Networks (FCNs) achieve segmentation of objects of different sizes by replacing fully connected layers with convolutional layers in CNNs, using an encoder to extract features at multiple levels, and finally a decoder to combine the encoded features. Furthermore, other FCN structures have been extensively studied; UNet and SegNet also use the well-known encoder-decoder structure to segment images. DeepLab introduces a spatial pyramid module, using dilated convolutional layers with different dilation rates connected in parallel to capture contextual information of varying ranges. They perform better in near-field and medical image segmentation. Inspired by the successful application of deep learning in semantic segmentation, CNNs and FCNs have been introduced into ground feature boundary segmentation. ResNet, LinkNet, and D-Link-Net are among the best networks for road extraction. D-Link-Net, in particular, uses LinkNet with a pre-trained encoder (ResNet34) as its backbone and has additional dilated convolutional layers in the central part. It also adopts a cascaded and parallel mode to increase the receptive field of the convolutional kernel while keeping the number of parameters constant, so that the size of the output feature map remains unchanged. It also improves the segmentation accuracy by combining features at different scales.

[0004] Because previous methods ignored the structural and topological information of feature boundaries, resulting in poor connectivity, researchers opted to use graph neural networks and extensive post-processing to ensure feature boundary connectivity. Unlike segmentation methods, graph neural network-based boundary detection methods can directly output the graph structure. Later road centerline tracking algorithms used an iterative search strategy convolutional neural network (RoadTracer) to track continuous and complete road networks. To address the issue of RoadTracer's inability to detect high-quality road intersections due to stride limitations, heatmap prediction was used instead of direction prediction to display the probability of adjacent vertex distributions; however, this approach suffers from low performance and inability to identify points that are too close together. Sat2Graph predicts a 19-dimensional tensor from high-quality aerial imagery, containing all the necessary information for the road network, which can be decoded and used to obtain the road network graph. In real-time instance segmentation, unlike directly regressing object boundary point coordinates in images, DeepSnake uses recurrent convolution to better utilize the cyclic graph structure of the contour, handling errors in target localization through initial contour and contour deformation. Among the most advanced boundary extraction algorithms currently available, PolyWord uses graph neural networks to predict the connection strength between each pair of vertices and correctly connects them to create accurate polygons, which is helpful for extracting vector polygons in geographic and mapping applications. Summary of the Invention

[0005] The purpose of this invention is to overcome the problem of imprecise and incomplete edge extraction of ground objects in high-resolution remote sensing images due to factors such as blurred boundaries and mutual occlusion. This invention proposes an edge extraction method for high-resolution remote sensing images based on edge tracking. It obtains relatively complete ground object edges through edge detection algorithms and reconstructs relatively complete edge tracking maps through edge tracking algorithms. It detects and extracts indistinct and broken edge positions and predicts the tracking route of broken edges by using tracking. By combining the advantages of both methods, complete edge information can be extracted.

[0006] To achieve the above objectives, the technical solution proposed by this invention is as follows:

[0007] 1. A high-resolution remote sensing edge repair method based on edge tracking, characterized by comprising the following steps:

[0008] Step 1: Design or modify a deep convolutional neural network based on the remote sensing target extraction task, create samples, and train to obtain an edge detection model and an edge tracking model.

[0009] The detailed steps for step 1 are as follows:

[0010] (1-1) Acquiring high-resolution remote sensing images. Optical satellite remote sensing data with visible-near-infrared sensors or aerial remote sensing data with general optical cameras can be used directly or fused with panchromatic images, depending on the resolution requirements.

[0011] (1-2) Cropping remote sensing images. Select remote sensing images of typical targets within the production area for cropping (the cropping range should have a certain degree of randomness and broad coverage according to the size of the production area).

[0012] (1-3) Cropping edge-tracking remote sensing images. Select the area where typical targets are located in the production area to fill the edges of the edge-detected remote sensing images.

[0013] (1-4) Create edge detection samples. Use ArcGIS or other GIS software to draw edge detection training samples. In this embodiment, it is necessary to mark the edges of cultivated land and generate corresponding cultivated land samples according to the requirements of the edge detection model.

[0014] (1-5) Create edge tracking samples. Convert the farmland samples corresponding to edge detection into .graph files, input the farmland sample .graph files into the random starting point algorithm, and generate a random starting point .json file.

[0015] (1-6) Collect images and corresponding annotations as training samples according to the task requirements. If detection accuracy is required, test samples can be prepared separately.

[0016] (1-7) Design of deep convolutional neural networks. In order to train the edge detection model and the edge tracking model, DLinkNet34_3plus and RoadTracer neural networks are selected in this invention. They can be replaced or modified accordingly depending on the production target network.

[0017] (1-8) Set training hyperparameters.

[0018] (1-9) Input samples and train the model. Input the training samples into the DLinkNet34_3plus model and train it according to the hyperparameters to obtain an edge detection model that can extract the edge of farmland; input the training samples into RoadTracer and train it according to the hyperparameters to obtain an edge tracking model that can extract the edge of farmland.

[0019] Step 2: Input the remote sensing image into the edge detection model, that is, use the model to extract the boundaries of ground features, thereby obtaining the edge intensity map of ground features; at the same time, input the remote sensing image into the edge tracking model, that is, use the model to extract the edge tracking contour of ground features, thereby obtaining the edge tracking map of ground features.

[0020] The detailed steps for step 2 are as follows:

[0021] (2-1) Input high-resolution remote sensing images into the edge detection model to obtain the edge intensity map of cultivated land.

[0022] (2-2) Input high-resolution remote sensing imagery into the edge tracking model to obtain farmland tracking maps, including the following two steps:

[0023] (2-2-1) Input the edge detection intensity map into the corner detection algorithm to detect the vertices in the cultivated land sample that produce angular offsets, output the corner pixel coordinates as a .txt file, use it as the starting point of the multi-start point tracking algorithm, and output the corner detection schematic diagram.

[0024] set up Let w(x, y) represent the pixel grayscale value at point (x, y) in the image, and w(x, y) represent the sliding window for Harris corner detection centered at (x, y), i.e., the weighting function. When the window moves by (Δx, Δy), the mathematical changes in the pixel grayscale values ​​in the window before and after the sliding are shown in formula (1):

[0025]

[0026] (2-2-2) Input the high-resolution remote sensing image prediction samples, random starting point .json file, corner detection .txt file into the edge tracking model, and use the post-processing algorithm to obtain the disconnected farmland edge tracking .graph file.

[0027] Step 3: Based on the edge intensity map of ground features obtained from edge detection and the discontinuous edge tracking map of ground features obtained from edge tracing, the two results are overlaid and analyzed. The location of the broken line is determined by the fusion algorithm, and the edge is extended according to the tracking results to complete the broken line repair work. Finally, the edge detection result map and the complete edge tracking result map are generated.

[0028] The detailed steps for step 3 are as follows:

[0029] (3-1): The edge intensity map obtained from step 2 is a grayscale image with pixel values ​​ranging from 0 to 255. The edge intensity map is then converted into a binary image using a thresholding algorithm. The inter-class variance calculation for the OTSU thresholding algorithm is shown in formula (2):

[0030]

[0031] , This represents the probability of grayscale values ​​being less than and greater than this threshold. , The average gray level of each part. Given the overall grayscale mean of the image, extract the threshold Thresh corresponding to the maximum inter-class variance; this value is the optimal threshold.

[0032] (3-2): Detect discontinuous parts based on the edge binary map and the edge tracing map. Specifically, divide the vertices and edges of the edge tracing map into multiple segments. If the edge detection binary map and the edge tracing map only partially cover each other, then the part is considered discontinuous, and the next operation is performed on the part.

[0033] (3-3): Automatically infer edge width in discontinuous parts. Specifically, based on experience, a buffer width is created and denoted as Buffer, the edge binary map is denoted as Seg, the intersection of Buffer and the edge binary map is denoted as area, the intersection of the single pixel width on the edge tracing map and the edge binary map is denoted as length, and num(·) counts the number of pixels in the intersection area. Therefore, the edge width can be calculated using formula (3):

[0034]

[0035] (3-4): Fuse the edge tracking map and the edge binary map to generate the fused result map. The fused result map is denoted as... The edge binary graph is denoted as Edge tracing graph notation is The threshold of the OTSU algorithm is denoted as The fusion result can be calculated using formula (4):

[0036]

[0037] (3-5): Generate the final edge fusion result map and the complete edge tracking result map based on the fusion result map. The final edge fusion result map is binarized using a threshold segmentation algorithm, and the edge tracking result map is refined using a skeleton thinning algorithm to obtain the final ground feature edge extraction result.

[0038] (3-6): The refined edge tracking results are used as edge tracking samples again, and multiple training sessions are conducted to obtain better edge fusion results.

[0039] Preferably, in step (1-2), the remote sensing images are uniformly cropped to a size of 1000*1000 pixels; in step (1-3), the edge-detected remote sensing images are uniformly filled to a size of 1024*1024 pixels.

[0040] Preferably, step (1-8) sets the training hyperparameters, specifically including:

[0041] The training parameters for DLinkNet34_3plus are set as follows: number of iterations = 400, batch_size = 4, learning rate update strategy = step, learning rate update step size = [100, 200, 300, 400], initial learning rate = 0.0002, and learning rate update coefficient = 0.2.

[0042] The training parameters for RoadTracer are set as follows: number of iterations = 400, number of iterations per round = 1000, search window = 256, batch_size = 1, learning rate update strategy = step, learning rate update step size = [100, 200, 300, 400], initial learning rate = 0.00001, and learning rate update coefficient = 0.1.

[0043] Preferably, the corner detection algorithm described in step (2-2-1) has a maximum number of corners = 100, a quality level = 0.1, and a minimum corner distance = 200.

[0044] The general edge features learned in the model used in step 2 can meet the subsequent segmentation requirements; for the tracking task, the tracking angle and component features learned in the model used in step 2 need to be re-prepared for special remote sensing images or special ground features, and training samples need to be prepared to learn different features.

[0045] The edge detection results generated by iterative tracking are used to repair imperfect results. The edge tracking prediction map is usually complete and continuous, which effectively compensates for the discontinuous edge information caused by factors such as trees, building occlusion, and noise in edge detection. The fusion algorithm combines the two types of edge information, and through tracking-assisted edge detection, it completes the edges that are not in the edge tracking results and automatically infers the edge width.

[0046] The edge tracking algorithm in this invention primarily employs a multi-starting-point tracking strategy. The starting point is automatically generated through the intersection of ground feature boundaries and embedded in the RoadTracer to learn ground feature boundary information from high-resolution remote sensing images for subsequent reconstruction of complete ground feature boundaries. The RoadTracer has two important outputs: an action component determining walking or stopping and an angle component determining the search direction. The training process is optimized using the squared loss of angle and action, as well as the cross-entropy loss between the predicted thumbnail and the ground reality image. In actual edge tracking, automatically discovering and tracking edges based on ground feature edge features can overcome the adverse effects of weak boundaries, unclear boundary information, and tall occlusions, connecting ground feature edges to form complete closed surfaces. However, due to the influence of the search starting point and the current tracking point, it may lack completeness. Therefore, the current problem to be solved is how to combine the ground feature boundaries extracted by segmentation and tracking methods, allowing them to constrain each other and complement each other's advantages.

[0047] This invention employs a deep learning-based multi-stage framework for ground feature edge detection and repair. The framework comprises three stages: edge detection, edge tracking, and edge fusion. The edge fusion algorithm combines the results of edge detection and tracking to aggregate the semantic and topological information of ground feature edges, generating the final, refined edges. This invention attempts to combine the advantages of both edge detection and edge tracking algorithms. Building upon the results of discontinuous edge extraction, it uses edge tracking to fuse the edge detection and tracking results, thus completing the repair of broken ground feature edges.

[0048] Due to the adoption of the above technical solution, the present invention has the following advantages and beneficial effects:

[0049] 1. Discontinuous edge parts in the edge result map generated by the edge detection algorithm can be repaired by using the topological information provided by the edge tracking algorithm. This can partially replace or assist manual production and improve the accuracy and efficiency of edge extraction.

[0050] 2. Edge binary maps help connect broken edge tracking maps, or create edge completion edge tracking maps in places where they do not exist. By repeatedly inputting and fusing edge tracking maps, the accuracy of edge tracking models can be improved.

[0051] 3. This invention uses a deep neural network algorithm in machine learning. It can extract and repair the boundaries of various types of land features by simply changing the training samples. It does not require redesigning the algorithm and has strong reusability and robustness. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the method of the present invention.

[0053] Figure 2 This is an example image of the satellite imagery farmland edge detection training model of the present invention.

[0054] Figure 3 This is an example image of a satellite imagery farmland edge tracking training model of the present invention.

[0055] Figure 4 This is an example image of satellite imagery and corresponding farmland edge prediction results of the present invention.

[0056] Figure 5 This is an example image of satellite imagery and corresponding starting point for tracing the edge of cultivated land, as presented in this invention.

[0057] Figure 6 This is an example image of satellite imagery and corresponding farmland edge tracking and prediction results of the present invention.

[0058] Figure 7 This is a comparative example of the satellite image overlay and farmland edge fusion results of the present invention.

[0059] Figure 8 This is a comparative example of the results of satellite image overlay and farmland edge fusion and refinement according to the present invention.

[0060] Figure 9 This is an example diagram comparing the edge fusion result and the edge detection result of the present invention. Detailed Implementation

[0061] Figure 1 This is a schematic diagram of a high-resolution remote sensing image edge repair method based on edge tracking.

[0062] Reference Figure 1 As shown, a preferred embodiment of the present invention is provided, including the following steps:

[0063] Step 1: Design or modify a deep convolutional neural network based on the remote sensing target extraction task, create samples, and train to obtain an edge detection model and an edge tracking model.

[0064] The detailed steps for step 1 are as follows:

[0065] (1-1) Acquiring high-resolution remote sensing images. Optical satellite remote sensing data with visible-near-infrared sensors or aerial remote sensing data with general optical cameras can be used directly or fused with panchromatic images, depending on the resolution requirements.

[0066] (1-2) Cropping remote sensing images. Select the area where typical targets are located in the production area and crop the remote sensing images to a size of 1000*1000 pixels (the cropping range should have a certain degree of randomness and broad coverage according to the size of the production area).

[0067] (1-3) Cropping edge-tracking remote sensing images: Select the area where typical targets are located in the production area and uniformly fill the edge detection remote sensing images to a size of 1024*1024 pixels.

[0068] (1-4) Creating Edge Detection Samples. Use ArcGIS or other GIS software to draw edge detection training samples. In this embodiment, it is necessary to label the edges of cultivated land to obtain the corresponding .shp files. Generate the corresponding cultivated land samples according to the requirements of the edge detection model, such as... Figure 2 As shown.

[0069] (1-5) Creating Edge Tracking Samples. Convert the farmland samples corresponding to edge detection into .graph files. Input the farmland sample .graph files into the random starting point algorithm to generate a random starting point .json file. Edge tracking samples are as follows: Figure 3 As shown.

[0070] (1-6) Generally, more than 100 images and corresponding annotations are collected as training samples according to the task requirements. If detection accuracy is required, test samples can be prepared separately.

[0071] (1-7) Design of deep convolutional neural networks. In order to train the edge detection model and the edge tracking model, DLinkNet34_3plus and RoadTracer neural networks are selected in this invention. They can be replaced or modified accordingly depending on the production target network.

[0072] (1-8) Set training hyperparameters. Configure hyperparameters, and the specific values ​​after model tuning are shown below.

[0073] The training parameters for DLinkNet34_3plus are set as follows: number of iterations = 400, batch_size = 4, learning rate update strategy = step, learning rate update step size = [100, 200, 300, 400], initial learning rate = 0.0002, and learning rate update coefficient = 0.2.

[0074] The training parameters for RoadTracer are set as follows: number of iterations = 400, number of iterations per round = 1000, search window = 256, batch_size = 1, learning rate update strategy = step, learning rate update step size = [100, 200, 300, 400], initial learning rate = 0.00001, and learning rate update coefficient = 0.1.

[0075] (1-9) Input samples and train the model. Input the training samples into the DLinkNet34_3plus model and train it according to the hyperparameters described in (1-8) to obtain an edge detection model that can extract farmland edges; input the training samples into RoadTracer and train it according to the hyperparameters described in (1-8) to obtain an edge tracking model that can extract farmland edges.

[0076] Step 2: Input the remote sensing image into the edge detection model, that is, use the model to extract the boundaries of ground features, thereby obtaining the edge intensity map of ground features; at the same time, input the remote sensing image into the edge tracking model, that is, use the model to extract the edge tracking contour of ground features, thereby obtaining the edge tracking map of ground features.

[0077] The detailed steps for step 2 are as follows:

[0078] (2-1) Input high-resolution remote sensing imagery into the edge detection model to obtain farmland edge intensity map, such as... Figure 4 As shown.

[0079] (2-2) Input high-resolution remote sensing imagery into the edge tracking model to obtain farmland tracking maps, including the following two steps:

[0080] (2-2-1) Input the edge detection intensity map into the corner detection algorithm to detect vertices in the cultivated land sample that produce angular offsets. The maximum number of corners = 100, quality level = 0.1, minimum corner distance = 200. Output the corner pixel coordinates as a .txt file, which serves as the starting point for the multi-starting point tracking algorithm. Output a corner detection schematic diagram, such as... Figure 5 As shown.

[0081] set up Let w(x, y) represent the pixel grayscale value at point (x, y) in the image, and w(x, y) represent the sliding window for Harris corner detection centered at (x, y), i.e., the weighting function. When the window moves by (Δx, Δy), the mathematical changes in the pixel grayscale values ​​in the window before and after the sliding are shown in formula (1):

[0082]

[0083] (2-2-2) Input the high-resolution remote sensing image prediction samples, the random starting point .json file, and the corner detection .txt file into the edge tracking model, and use a post-processing algorithm to obtain the disconnected farmland edge tracking .graph file. Convert the .graph file into a binary image, such as... Figure 6 As shown.

[0084] Step 3: Based on the edge intensity map of ground features obtained from edge detection and the discontinuous edge tracking map of ground features obtained from edge tracing, the two results are overlaid and analyzed. The location of the broken line is determined by the fusion algorithm, and the edge is extended according to the tracking results to complete the broken line repair work. Finally, the edge detection result map and the complete edge tracking result map are generated.

[0085] The detailed steps for step 3 are as follows:

[0086] (3-1): The farmland edge intensity map obtained from (3) is a grayscale image with pixel values ​​ranging from 0 to 255. The edge intensity map is converted into a binary image according to the OTSU thresholding algorithm. The inter-class variance calculation of the OTSU thresholding algorithm is shown in formula (2):

[0087]

[0088] , This represents the probability of grayscale values ​​being less than and greater than this threshold. , The average gray level of each part. Given the overall grayscale mean of the image, extract the threshold Thresh corresponding to the maximum inter-class variance; this value is the optimal threshold.

[0089] (3-2): Detect discontinuous parts based on the edge binary map and the edge tracing map. Specifically, divide the vertices and edges of the edge tracing map into multiple segments. If the edge detection binary map and the edge tracing map only partially cover each other, then the part is considered discontinuous, and the next step is performed on the part.

[0090] (3-3): Automatically infer edge width in discontinuous parts. Specifically, based on experience, a buffer width is created and denoted as Buffer, the edge binary map is denoted as Seg, the intersection of Buffer and edge binary map is denoted as area, the intersection of the single pixel width on the edge tracing map and edge binary map is denoted as length, and num(·) counts the number of pixels in the intersection area. Therefore, the edge width can be calculated using formula (3):

[0091]

[0092] (3-4): Fuse the edge tracing map and the edge binary map to generate the fused result map. The fused result map is denoted as... The edge binary graph is denoted as Edge tracing graph is noted as The threshold of the OTSU algorithm is denoted as The fusion result can be calculated using formula (4):

[0093]

[0094] (3-5): Generate the final edge fusion result map and the complete edge tracing result map based on the fusion result, such as... Figure 7 As shown. The final edge fusion result image is binarized using the OTSU algorithm, and a skeleton extraction algorithm is used to refine the edge fusion result image to obtain a complete edge tracking result image, as shown. Figure 8 As shown. The edge fusion result map and the edge intensity map are compared, as shown. Figure 9 As shown.

[0095] (3-6): The refined edge tracking results are used as edge tracking samples again, and multiple training sessions are conducted to obtain better edge fusion results.

[0096] The implementation of the present invention is not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A high-resolution remote sensing edge repair method based on edge tracking, characterized in that... It includes the following steps: Step 1: Design or modify a deep convolutional neural network based on the remote sensing target extraction task, create samples, and train the edge detection model and edge tracking model; specifically including: (1-1) Acquire high-resolution remote sensing images; use optical satellite remote sensing data with visible light-near infrared sensors or aerial remote sensing data with general optical cameras. Depending on the resolution requirements, multispectral images or panchromatic images can be used directly. (1-2) Cropping remote sensing images; select remote sensing images of typical targets within the production area for cropping; (1-3) Cropping edge-tracking remote sensing images; filling the edge detection remote sensing image edges within the range of typical targets in the production area; (1-4) Create edge detection samples; use ArcGIS or other GIS software to draw edge detection training samples, mark the edges of cultivated land, and generate corresponding cultivated land samples according to the requirements of the edge detection model; (1-5) Create edge tracking samples; convert the farmland samples corresponding to edge detection into .graph files, input the farmland sample .graph files into the random starting point algorithm, and generate a random starting point .json file; (1-6) Collect images and corresponding annotations as training samples according to task requirements. When detecting accuracy, prepare test samples separately. (1-7) Design a deep convolutional neural network; In order to train the edge detection model and the edge tracking model, DLinkNet34_3plus and RoadTracer neural networks are used. The network can be replaced or modified accordingly according to the production target network. (1-8) Set the training hyperparameters; (1-9) Input samples and train the model; input the training samples into the DLinkNet34_3plus model and train it according to the hyperparameters to obtain an edge detection model that can extract the edge of farmland; input the training samples into RoadTracer and train it according to the hyperparameters to obtain an edge tracking model that can extract the edge of farmland. Step 2: Input the remote sensing image into the edge detection model, that is, use the model to extract the boundaries of ground features, thereby obtaining the edge intensity map of ground features; at the same time, input the remote sensing image into the edge tracing model, that is, use the model to extract the edge tracing contours of ground features, thereby obtaining the edge tracing map of ground features; specifically including: (2-1) Input high-resolution remote sensing images into the edge detection model to obtain ground feature edge intensity maps; (2-2) Input the high-resolution remote sensing image into the edge tracking model to obtain the ground feature edge tracking map, including the following two steps: (2-2-1) Input the edge intensity map of the land feature into the corner detection algorithm to detect the vertices in the cultivated land sample that produce angular offsets, output the pixel coordinates of the corner points as a .txt file, use it as the starting point of the multi-start point tracking algorithm, and output the corner detection schematic diagram; set up Represents the pixel grayscale value at point (x, y) in the image. The sliding window representing the Harris corner detection centered at (x, y) is the weighting function. When the window moves by (△x, △y), the mathematical changes in the grayscale of the pixels in the window before and after the sliding are shown in formula (1): (2-2-2) Input the high-resolution remote sensing image prediction samples, random starting point .json file, and corner detection .txt file into the edge tracking model, and use the post-processing algorithm to obtain the disconnected farmland edge tracking .graph file; Step 3: Based on the edge intensity map of ground features obtained from edge detection and the discontinuous edge tracking map of ground features obtained from edge tracing, the two results are overlaid and analyzed. A fusion algorithm is used to determine the location of the broken line, and edge extension is performed based on the tracking results to complete the broken line repair work. Finally, an edge detection result map and a complete edge tracking result map are generated; specifically including: (3-1): The edge intensity map of the ground features obtained from step 2 is a grayscale image with pixel values ​​ranging from 0 to 255. The edge intensity map is converted into a binary image according to the threshold segmentation algorithm. The inter-class variance calculation of the OTSU threshold algorithm is shown in formula (2): , This represents the probability of grayscale values ​​being less than and greater than this threshold. , The average gray level of each part. Given the overall grayscale mean of the image, extract the threshold Thresh corresponding to the maximum inter-class variance; this value is the optimal threshold. (3-2): Detect discontinuous parts based on the edge binary map and the edge tracing map. Specifically, divide the vertices and edges of the edge tracing map into multiple segments. If the edge binary map and the edge tracing map only partially cover each other, then the part is considered discontinuous, and the next operation is performed on the part. (3-3): Automatically infer edge width in discontinuous sections; specifically, create a buffer width based on experience, denoted as Buffer, and an edge binary map, denoted as . Let area be the intersection of the Buffer and the edge binary map, length be the intersection of the single pixel width on the edge tracing map and the edge binary map, and num(·) count the number of pixels in the intersection region; therefore, the edge width can be calculated by formula (3): (3-4): Fuse the edge tracking map and the edge binary map to generate the fused result map; the fused result map is denoted as... The edge binary graph is denoted as Edge tracing graph notation is The threshold of the OTSU algorithm is denoted as The fusion result can be calculated using formula (4): (3-5): Generate the final edge fusion result map and the complete edge tracking result map based on the fusion result map; the final edge fusion result map is binarized using a threshold segmentation algorithm, and the edge tracking result map is refined using a skeleton thinning algorithm to obtain the final ground feature edge extraction result; (3-6): The refined edge tracking results are used as edge tracking samples again, and multiple training sessions are conducted to obtain better edge fusion results.

2. The high-resolution remote sensing edge repair method according to claim 1, characterized in that: Step (1-2) uniformly crop the remote sensing images to a size of 1000*1000 pixels; Step (1-3) uniformly fill the edge-detected remote sensing images to a size of 1024*1024 pixels.

3. The high-resolution remote sensing edge repair method according to claim 1, characterized in that: Steps (1-8) involve setting the training hyperparameters, specifically including: The training parameters for DLinkNet34_3plus are set as follows: number of iterations = 400, batch_size = 4, learning rate update strategy = step, learning rate update step size = [100, 200, 300, 400], initial learning rate = 0.0002, and learning rate update coefficient = 0.

2. The training parameters for RoadTracer are set as follows: number of iterations = 400, number of iterations per round = 1000, search window = 256, batch_size = 1, learning rate update strategy = step, learning rate update step size = [100, 200, 300, 400], initial learning rate = 0.00001, and learning rate update coefficient = 0.

1.

4. The high-resolution remote sensing edge repair method according to claim 1, characterized in that: The corner detection algorithm described in step (2-2-1) has a maximum number of corners = 100, a quality level = 0.1, and a minimum corner distance = 200.

5. The high-resolution remote sensing edge repair method according to claim 1, characterized in that: For general segmentation tasks, step 2 uses an edge detection model, and the learned general edge features can meet the subsequent segmentation requirements. For tracking tasks, step 2 uses an edge tracking model, and the learned tracking angle and component features need to be re-prepared for special remote sensing images or special ground features, and training samples are used to learn different features.

Images