Remote sensing image land development category segmentation method and system based on edge detection

Abstract

The invention relates to the technical field of satellite remote sensing image segmentation, in particular to a remote sensing image land development category segmentation method and system based on edge detection, and the method specifically comprises the following steps: constructing a land development category segmentation model based on a satellite remote sensing image, and inputting a satellite remote sensing land image shot by satellite remote sensing into the model; obtaining a land development category segmentation result; the model comprises a satellite image feature reconstruction module, a super-resolution satellite image detail enhancement module, a multi-scale convolution construction super-resolution satellite image detail enhancement module, an edge local perception attention module, a satellite image global capture module and a satellite image land category segmentation module. According to the method, the problems of fuzzy ground feature boundary, difficulty in multi-scale feature fusion, difficulty in collaborative modeling of local details and global semantic information and the like when the satellite remote sensing land image is processed can be solved, and the segmentation precision of the satellite remote sensing image on the development land category is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of satellite remote sensing image segmentation, and in particular to a method and system for segmenting land development categories in remote sensing images based on edge detection. Background Art

[0002] With the development of deep learning and high-resolution satellite remote sensing technology, land classification using satellite remote sensing image segmentation has become a key method for smart city construction and natural resource management. Satellite remote sensing imagery can accurately segment dozens of land types, including construction land, cultivated land, forest land, and water areas. Its application scenarios cover national land space planning, crop planting monitoring, ecological and environmental assessments, and disaster loss quantification. Satellite remote sensing imagery land segmentation technology can dynamically monitor macro-trends such as urban expansion, non-grain conversion of cultivated land, and changes in forest cover, assisting government departments in formulating strategies for optimizing land resource allocation. Compared to traditional manual demarcation methods, deep learning-based satellite remote sensing land segmentation methods can increase classification efficiency by nearly 100 times, significantly reducing the cost of high-precision land surveys and providing highly timely and spatially granular data support for global climate change research and the implementation of sustainable development goals.

[0003] Although satellite remote sensing land segmentation technology based on deep learning has made significant progress, its application still faces many problems: first, the complex landforms place higher demands on the model's feature extraction capabilities, and the existing network architecture often has blurred edges or missed small targets; second, dynamic changes in the environment will significantly reduce the generalization performance of the model, and frequent data retraining is required to maintain accuracy.

[0004] Therefore, the present invention proposes a remote sensing image land development category segmentation method and system based on edge detection to solve the above problems. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention develops a method and system for land development category segmentation in remote sensing images based on edge detection. By solving the problems existing in the processing of satellite remote sensing land images, such as blurred boundaries of land objects, difficulty in fusing multi-scale features, and difficulty in collaborative modeling of local details and global semantic information, the present invention can effectively improve the accuracy of land category segmentation in satellite remote sensing images.

[0006] On the one hand, the technical solution to the technical problem of the present invention is a method for land development category segmentation of remote sensing images based on edge detection, which constructs a land development category segmentation model based on satellite remote sensing images, inputs satellite remote sensing land images taken by satellite remote sensing into the model, and obtains land development category segmentation results, including the following steps: S1. Construct satellite image feature reconstruction module: This module includes depthwise separable convolution operations, The activation function, satellite image feature space reconstruction module and downsampling layer input the satellite remote sensing land image taken by satellite remote sensing into the satellite image feature reconstruction module and output the satellite image feature reconstruction features; S2. Constructing a super-resolution satellite image detail enhancement module: Constructing a super-resolution satellite image detail enhancement module based on multi-scale convolution, inputting the satellite image feature reconstruction features into the super-resolution satellite image detail enhancement module for super-resolution detail enhancement, and outputting super-resolution satellite image detail enhancement features; S3. Constructing an edge local perception attention module: This module includes an edge local perception attention module, which inputs the super-resolution satellite image detail enhancement features into the edge local perception attention module to capture local features and output edge local perception attention features; S4. Constructing a satellite image global capture module: This module includes a satellite image global capture module, which inputs the edge local perception attention features into the satellite image global capture module to enhance the global correlation of the features and outputs the satellite image global features; S5. Construct satellite image land category segmentation module: combine edge local perception attention features with satellite image global features, perform feature splicing operation, output fusion local global information features, and then input them into satellite image land category segmentation module to output satellite image land category segmentation image .

[0007] S1 is as follows: S1.1. Satellite remote sensing land images The convolution kernel size input to the satellite image feature reconstruction module is The depth-wise separable convolutional layers and Activation function, output features ; S1.2、Characterize Input into the satellite image feature space reconstruction module in the satellite image feature reconstruction module, first change the feature by transposition operation The spatial structure of the output is transformed after the transposition operation ; Transformation features After the convolution kernel size is The convolution layer and feature segmentation Operation, transform features The number of channels divided along the channel direction is the transformation feature Sub-features of and ; (1) For sub-features , the feature is convolution kernel size is Depthwise separable convolution operations, Non-linear activation function operation, Feature reconstruction operations, Feature scanning operation, Normalization operation, and reuse Reconstruction operation, output processed sub-features , the first time passing The feature reconstruction operation flattens the features into a one-dimensional vector, and the second pass The feature reconstruction operation restores the features into two-dimensional vectors; (2) For sub-features , after the convolution kernel size is The convolution layer and convolution kernel size are The depthwise separable convolutional layers and Non-linear activation function operation, outputting processed sub-features ; (3) The processed sub-features and Splicing fusion, after convolution kernel size is The convolution layer outputs features ; Then the features and features Splicing and fusion, output features ; S1.3、Characterize Input to the downsampling layer for downsampling operation, and output satellite image feature reconstruction features .

[0008] S2 is as follows: The super-resolution satellite image detail enhancement module based on multi-scale convolution specifically includes three parallel convolution layers to reconstruct the satellite image features. Input into the super-resolution satellite image detail enhancement module; S2.1. Reconstructing satellite image features Input into three parallel convolutional layers respectively; For the convolutional layer of the first branch: features After the convolution kernel size is The depth-wise separable convolutional layer outputs features ;feature After the convolution kernel size is The depth-wise separable convolutional layer outputs features ; For the convolutional layer of the second branch: features After the convolution kernel size is The convolutional layer and Non-linear activation function, output features ; For the convolutional layer of the third branch: features After the convolution kernel size is The depth of the separable convolution layer, the features obtained are ,feature Splicing and fusion, output features ;feature After the convolution kernel size is The depth-wise separable convolutional layer outputs features ; The features output by the convolutional layers of the three branches ,feature and features Perform splicing and fusion to output features ,By splicing features through multi-scale convolutional layers, the multi-scale expression capability of ,detail features of super-resolution satellite images can be enhanced; S2.2, the characteristics After the global average pooling layer and the fully connected layer, the feature Fusion, output channel attention features ; Then the features After the convolution kernel size is After the convolution layer and the feature Fusion, output spatial attention features ; Then the channel attention feature and spatial attention features Splicing and fusion, and after convolution kernel size is The convolutional layer outputs the super-resolution satellite image detail enhancement features .

[0009] S3 is as follows: Enhance the details of super-resolution satellite images Input to the edge local perception attention module, the edge local perception attention module consists of two branches, The two branches of the edge local perception attention module are respectively input to extract local channel features and local spatial features, and then the two features are spliced ​​and fused to obtain the edge local perception attention features. ; S3.1. Enhance the details of super-resolution satellite images Input into the first branch of the edge local perception attention module, the feature After the global average pooling layer and the convolution kernel size is The convolutional layers and Activation function, output features , the features and features Perform broadcast multiplication and then combine the output features with the features Perform feature splicing and finally output local channel features ; S3.2. Enhance the details of super-resolution satellite images Input to the second branch of the edge local perception attention module, feature After the convolution kernel size is The convolution layer outputs features , then the feature Input to three parallel depth-separable convolutional layers, the convolution kernel size is The depth of the separable convolution layer and the convolution kernel size are The depth of the separable convolution layer and the convolution kernel size is The depth-wise separable convolution layer is then used to concatenate the output features of the three parallel depth-wise separable convolution layers. , and then the features After the convolution kernel size is The convolutional layers and Activation function, output features , and finally the features and features Perform broadcast multiplication and then combine the output features with the features Perform feature splicing and finally output local spatial features ; S3.3, local channel features and local spatial features Splicing and fusion, output edge local perception attention features .

[0010] The S4 steps are as follows: Enhance the details of super-resolution satellite images Input to the satellite image global capture module, the satellite image global capture module includes two branches, the feature Input the two branches of the satellite image global capture module respectively to obtain the first branch feature and the second branch feature, and then combine the two branch features with the feature Perform stitching and fusion to obtain the global features of satellite images ; S4.1. Enhance the details of super-resolution satellite images Input into the first branch structure of the satellite image global capture module, first Perform dimension transposition operation to transform the features Rotate along the height dimension and output features , then the feature The input to the convolution kernel size is The depth of the separable convolution layer and batch normalization layer, the output features are combined with the features Perform feature splicing and output features ,feature Then pass The activation function operation and convolution kernel size are The convolution layer outputs features ; The features After the convolution kernel size is The depth of the separable convolution layer and batch normalization layer are used to combine the output features with the features Perform feature splicing and output features , the features After the convolution kernel size is The depth-wise separable convolutional layer, batch normalization layer and convolution kernel size are The convolution layer outputs the first branch feature ; S4.2. Enhance the details of super-resolution satellite images Input into the second branch structure of the satellite image global capture module, first Perform dimension transposition operation to transform the features Rotate along the width dimension and output features , then the feature The input to the convolution kernel size is The depth of the separable convolution layer and batch normalization layer are combined with the obtained output to enhance the features Perform feature splicing and output features , and then the features go through The activation function operation and the convolution kernel size are The convolution layer outputs features ; Enhance the features After the convolution kernel size is The depth of the separable convolution layer and batch normalization layer are used to combine the output features with the features Perform feature splicing and output features ; The feature After the convolution kernel size is The depth-wise separable convolutional layer, batch normalization layer and convolution kernel size are The convolution layer outputs the second branch feature ; S4.3, the first branch feature , the second branch features and features Stitching and fusion to output global features of satellite images .

[0011] S5 is as follows: The edge local perception attention feature Global features of satellite images Perform splicing and fusion to output fusion local and global information features , and then the features Input to the satellite image land category segmentation module, after the convolution kernel size is The depth-wise separable convolutional layer and upsampling operation transform the features The resolution size is restored to match the image Same, use the image after restoring the resolution The extreme point positioning operation locates the specific land category to which each pixel in the image belongs, through The color mapping operation maps the specific land category to which the pixel belongs to different colors to generate a satellite image land category segmentation image. .

[0012] On the other hand, the present invention also provides a remote sensing image land development category segmentation system based on edge detection, including a module for executing processing instructions for each step in a remote sensing image land development category segmentation method based on edge detection.

[0013] The effects provided in the summary of the invention are only the effects of the embodiments, rather than all the effects of the invention. The above technical solution has the following advantages or beneficial effects: The present invention constructs a land development category segmentation model based on satellite remote sensing images. It processes the input satellite remote sensing land images through a combination of five modules. Specifically, through the synergistic effect of the satellite image feature reconstruction module and the super-resolution detail enhancement module, it can enhance edge texture details while retaining global structural information, effectively solving the problems of blurred ground object boundaries and missed detection of small targets in traditional methods. In addition, the dynamic capture of subtle features by the edge local perception attention module and the multi-dimensional transposed feature association modeling of the global capture module are combined to achieve a deep fusion of local details and large-scale land cover semantics, significantly improving the generalization ability of complex scenes. This requires overcoming the collaborative conflicts caused by differences in feature dimensions and resolution between modules, dynamically weighting key information through channel / spatial attention mechanisms to eliminate redundancy, and balancing the surge in computational complexity brought about by super-resolution multi-scale convolution and global dimensional transposition operations. Specifically, through the progressive architecture of "feature reconstruction → detail enhancement → local perception → global modeling → classification decision", it breaks through the limitations of the unidirectional information flow of traditional segmentation networks, innovatively designs the interaction mechanism between modules, and realizes step-by-step abstraction and feedback from low-level pixels to high-level semantics.

[0014] In summary, the present invention discloses a remote sensing image land category segmentation method and system based on edge detection, which reconstructs the features of effective information of land images through a satellite image feature reconstruction module; significantly enhances the representation ability of the boundary texture details of ground objects in remote sensing images through a super-resolution satellite image detail enhancement module; captures subtle edge features in the image through an edge local perception attention module; and solves the problem of semantic information irrelevance in the classification of large-scale land cover types in traditional methods through a satellite image global capture module; ultimately, it can effectively improve the segmentation accuracy of satellite remote sensing images for land categories. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0016] Figure 1 Schematic diagram of the overall structure of the method of the present invention.

[0017] Figure 2 The figure shows the comparison results of segmentation effects.

[0018] Figure 3 This is a structural diagram of the satellite image feature reconstruction module of the present invention.

[0019] Figure 4 This is a structural diagram of the super-resolution satellite image detail enhancement module of the present invention.

[0020] Figure 5 This is a structural diagram of the edge local perception attention module of the present invention.

[0021] Figure 6 This is a structural diagram of the satellite image global capture module of the present invention.

[0022] Figure 7 This is a structural diagram of the satellite image land category segmentation module of the present invention. DETAILED DESCRIPTION

[0023] In order to clearly illustrate the technical features of this solution, the present invention is described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0024] Example 1 A land development category segmentation method for remote sensing images based on edge detection is proposed. A land development category segmentation model based on satellite remote sensing images is constructed. Satellite remote sensing land images taken by satellite remote sensing are input into the model to obtain land development category segmentation results. The method includes the following steps: S1. Construct satellite image feature reconstruction module: This module includes depthwise separable convolution operations, The activation function, satellite image feature space reconstruction module and downsampling layer input the satellite remote sensing land image taken by satellite remote sensing into the satellite image feature reconstruction module and output the satellite image feature reconstruction features; S2. Constructing a super-resolution satellite image detail enhancement module: Constructing a super-resolution satellite image detail enhancement module based on multi-scale convolution, inputting the satellite image feature reconstruction features into the super-resolution satellite image detail enhancement module for super-resolution detail enhancement, and outputting super-resolution satellite image detail enhancement features; S3. Constructing an edge local perception attention module: This module includes an edge local perception attention module, which inputs the super-resolution satellite image detail enhancement features into the edge local perception attention module to capture local features and output edge local perception attention features; S4. Constructing a satellite image global capture module: This module includes a satellite image global capture module, which inputs the edge local perception attention features into the satellite image global capture module to enhance the global correlation of the features and outputs the satellite image global features; S5. Construct satellite image land category segmentation module: combine edge local perception attention features with satellite image global features, perform feature splicing operation, output fusion local global information features, and then input them into satellite image land category segmentation module to output satellite image land category segmentation image .

[0025] In a specific implementation manner, S1 is specifically as follows: S1.1. Satellite remote sensing land images The convolution kernel size input to the satellite image feature reconstruction module is The depth-wise separable convolutional layers and Activation function, output features , the calculation formula is: , in, Indicates that the activation function performs nonlinear processing on the features. Indicates that the convolution kernel size is The depth-wise separable convolutional layer, , Indicates the height of satellite remote sensing land image, Indicates the width of satellite remote sensing land image, Indicates the number of channels of satellite remote sensing land images; S1.2, the characteristics Input into the satellite image feature space reconstruction module in the satellite image feature reconstruction module, first change the feature by transposition operation The spatial structure of The spatial structure of can facilitate the subsequent reconstruction of feature information in different directions, and the transformed features are output after the transposition operation. ; Transformation features After the convolution kernel size is The convolution layer and feature segmentation Operation, transform features The number of channels divided along the channel direction is the transformation feature Sub-features of and , the calculation formula is: , , in, represents the feature transposition operation, represents the feature segmentation operation, Indicates that the convolution kernel size is Convolutional layers; (1) For sub-features , the feature is convolution kernel size is Depthwise separable convolution operations, Non-linear activation function operation, Feature reconstruction operations, Feature scanning operation, Normalization operation, and reuse Reconstruction operation, output processed sub-features , the first time passing The feature reconstruction operation flattens the features into a one-dimensional vector, and the second pass The feature reconstruction operation restores the feature into a two-dimensional vector, and the calculation formula is: , in, Indicates that the convolution kernel size is The depth-wise separable convolutional layer, express Non-linear activation function operation, the first The purpose of the feature reconstruction operation is to flatten the features into a one-dimensional vector. The feature scanning operation is used to process the long-distance dependencies between the flattened sequence features. Represents the normalization operation, after the second The purpose of the feature reconstruction operation is to restore the features into a two-dimensional vector; (2) For sub-features , after the convolution kernel size is The convolution layer and convolution kernel size are The depthwise separable convolutional layers and Non-linear activation function operation, outputting processed sub-features , the calculation formula is: , in, express Non-linear activation function operation, Indicates that the convolution kernel size is The depth-wise separable convolutional layer, Indicates that the convolution kernel size is Convolutional layers; (3) The processed sub-features and Splicing fusion, after convolution kernel size is The convolution layer outputs features ; Then the features and features Splicing and fusion, output features , the calculation formula is: , , in, Represents a splicing operation; S1.3、Characterize Input to the downsampling layer for downsampling operation, and output satellite image feature reconstruction features , the calculation formula is: , in, Represents the downsampling operation on the feature.

[0026] In a specific implementation manner, S2 is specifically as follows: The super-resolution satellite image detail enhancement module based on multi-scale convolution specifically includes three parallel convolution layers to reconstruct the satellite image features. Input into the super-resolution satellite image detail enhancement module; S2.1. Reconstructing satellite image features Input into three parallel convolutional layers respectively; For the convolutional layer of the first branch: features After the convolution kernel size is The depth-wise separable convolutional layer outputs features ;feature After the convolution kernel size is The depth-wise separable convolutional layer outputs features , the calculation formula is: , , in, Indicates that the convolution kernel size is The depth-wise separable convolutional layer, Indicates that the convolution kernel size is Depthwise separable convolutional layers; For the convolutional layer of the second branch: features After the convolution kernel size is The convolutional layer and Non-linear activation function, output features , the calculation formula is: , in, Indicates that the convolution kernel size is Convolutional layers; For the convolutional layer of the third branch: features After the convolution kernel size is The depth of the separable convolution layer, the features obtained are ,feature Splicing and fusion, output features ;feature After the convolution kernel size is The depth-wise separable convolutional layer outputs features , the calculation formula is: , , in, Represents a splicing operation, Indicates that the convolution kernel size is The depth-wise separable convolutional layer, Indicates that the convolution kernel size is Depthwise separable convolutional layers; The features output by the convolutional layers of the three branches ,feature and features Perform splicing and fusion to output features By splicing features in multi-scale convolutional layers, the multi-scale expression capability of the detail features of super-resolution satellite images can be enhanced. The calculation formula is: ; S2.2, the characteristics After the global average pooling layer and the fully connected layer, the feature Fusion, output channel attention features ; Then the features After the convolution kernel size is After the convolution layer and the feature Fusion, output spatial attention features ; Then the channel attention feature and spatial attention features Splicing and fusion, and after convolution kernel size is The convolutional layer outputs the super-resolution satellite image detail enhancement features , the calculation formula is: , , , in, represents broadcast multiplication across feature dimensions, represents the global average pooling operation, represents the fully connected layer, Indicates that the convolution kernel size is The convolutional layer, Represents a splicing operation.

[0027] In a specific implementation, S3 is as follows: Enhance the details of super-resolution satellite images Input to the edge local perception attention module, the edge local perception attention module consists of two branches, The two branches of the edge local perception attention module are respectively input to extract local channel features and local spatial features, and then the two features are spliced ​​and fused to obtain the edge local perception attention features. ; S3.1. Enhance the details of super-resolution satellite images Input into the first branch of the edge local perception attention module, the feature After the global average pooling layer and the convolution kernel size is The convolutional layers and Activation function, output features , the features and features Perform broadcast multiplication and then combine the output features with the features Perform feature splicing and finally output local channel features , the calculation formula is: , , in, express Activation function operation, represents the global average pooling operation, represents the feature concatenation operation, An operator representing broadcast multiplication; S3.2. Enhance the details of super-resolution satellite images Input to the second branch of the edge local perception attention module, feature After the convolution kernel size is The convolution layer outputs features , then the feature Input to three parallel depth-separable convolutional layers, the convolution kernel size is The depth of the separable convolution layer and the convolution kernel size are The depth of the separable convolution layer and the convolution kernel size is The depth-wise separable convolution layer is then used to concatenate the output features of the three parallel depth-wise separable convolution layers. , and then the features After the convolution kernel size is The convolutional layers and Activation function, output features , and finally the features and features Perform broadcast multiplication and then combine the output features with the features Perform feature splicing and finally output local spatial features , the calculation formula is: , , , in, Indicates that the convolution kernel size is The convolutional layer, 、 、 They represent the convolution kernel size respectively. The depth of the separable convolution layer and the convolution kernel size are The depth of the separable convolution layer and the convolution kernel size are The depth-wise separable convolutional layer, An operator representing broadcast multiplication; S3.3, local channel features and local spatial features Splicing and fusion, output edge local perception attention features , the calculation formula is: .

[0028] In a specific implementation, step S4 is as follows: Enhance the details of super-resolution satellite images Input to the satellite image global capture module, the satellite image global capture module includes two branches, the feature Input the two branches of the satellite image global capture module respectively to obtain the first branch feature and the second branch feature, and then combine the two branch features with the feature Perform stitching and fusion to obtain the global features of satellite images ; S4.1. Enhance the details of super-resolution satellite images Input into the first branch structure of the satellite image global capture module, first Perform dimension transposition operation to transform the features Rotate along the height dimension and output features , then the feature The input to the convolution kernel size is The depth of the separable convolution layer and batch normalization layer, the output features are combined with the features Perform feature splicing and output features ,feature Then pass The activation function operation and convolution kernel size are The convolution layer outputs features ; The features After the convolution kernel size is The depth of the separable convolution layer and batch normalization layer are used to combine the output features with the features Perform feature splicing and output features , the features After the convolution kernel size is The depth-wise separable convolutional layer, batch normalization layer and convolution kernel size are The convolution layer outputs the first branch feature , the calculation formula is: , , , , , in, Represents the transposition operation along the feature height dimension, feature The order of dimensions is , Representation characteristics The number of channels, Representation characteristics height, Representation characteristics The width of the corresponding dimension is ,feature The dimension order is ; represents the batch normalization layer, Indicates that the convolution kernel size is The depth-wise separable convolutional layer, Indicates that the convolution kernel size is Convolutional layers; S4.2. Enhance the details of super-resolution satellite images Input into the second branch structure of the satellite image global capture module, first Perform dimension transposition operation to transform the features Rotate along the width dimension and output features , then the feature The input to the convolution kernel size is The depth of the separable convolution layer and batch normalization layer are combined with the obtained output to enhance the features Perform feature splicing and output features , and then the features go through The activation function operation and the convolution kernel size are The convolution layer outputs features ; Enhance the features After the convolution kernel size is The depth of the separable convolution layer and batch normalization layer are used to combine the output features with the features Perform feature splicing and output features ; The feature After the convolution kernel size is The depth-wise separable convolutional layer, batch normalization layer and convolution kernel size are The convolution layer outputs the second branch feature , the calculation formula is: , , , , , in, Represents the transposition operation along the feature width dimension, feature The dimension order is ; represents the batch normalization layer, Indicates that the convolution kernel size is The depth-wise separable convolutional layer, Indicates that the convolution kernel size is Convolutional layers; S4.3, the first branch feature , the second branch features and features Stitching and fusion to output global features of satellite images , the calculation formula is: .

[0029] In a specific implementation, S5 is as follows: The edge local perception attention feature Global features of satellite images Perform splicing and fusion to output fusion local and global information features , and then the features Input to the satellite image land category segmentation module, after the convolution kernel size is The depth-wise separable convolutional layer and upsampling operation transform the features The resolution size is restored to match the image Same, use the image after restoring the resolution The extreme point positioning operation locates the specific land category to which each pixel in the image belongs, through The color mapping operation maps the specific land category to which the pixel belongs to different colors to generate a satellite image land category segmentation image. , the calculation formula is: , , in, Represents an upsampling operation.

[0030] Example 2 A remote sensing image land development category segmentation system based on edge detection comprises a module for executing processing instructions of each step in a remote sensing image land development category segmentation method based on edge detection.

[0031] Example 3 In order to verify the effectiveness of the edge detection-based land development category segmentation method for remote sensing images proposed in the present invention in the field of land category segmentation, the edge detection-based land development category segmentation method for remote sensing images proposed in the present invention and the existing remote sensing image land category segmentation method are verified under the same experimental environment and conditions, and the obtained land category segmentation results are compared and analyzed to verify the effectiveness of the method proposed in the present invention in the field of satellite remote sensing land category segmentation.

[0032] Among the existing methods with high land class share, three currently mainstream segmentation methods are selected as comparison models for the method of the present invention. The three comparison methods selected are: nnUNet model (neural network union network), nnFormer model (non-local attention driven Transformer network) and SwinUNETR model (a network composed of Swin Transformer and 3D convolution module); The evaluation indicators are: F1-score (F1 score), mIoU (mean intersection over union) and OA (overall accuracy); among them, F1-score measures the comprehensive performance of the land category segmentation task by harmonizing the average precision and recall rate. This indicator is suitable for remote sensing scenes with uneven distribution of land development categories and can comprehensively reflect the robustness of the proposed method for small sample segmentation; mIoU can measure the overlap between the predicted segmented area and the true labeled area, and can quantify the accuracy of the proposed method in the boundary segmentation of multi-category land development targets (such as farmland, residential areas, and industrial areas); OA can intuitively evaluate the overall performance of the proposed method for land development category segmentation in complex remote sensing images; The DeepGlobe dataset was selected as the dataset for the experiment. It is a high-resolution satellite remote sensing public dataset for global land cover analysis and building detection. It covers a variety of land development-related scenes such as cities, farmlands, forests, and water bodies, and is suitable for land development category segmentation and change detection tasks. The images in the DeepGlobe dataset cover different geographical regions, seasonal characteristics, and lighting conditions, which can effectively verify the segmentation performance of the proposed method in complex surface cover scenarios. In the experiment, the DeepGlobe dataset was divided into training set, validation set, and test set according to official standards, with a ratio of 8:1:1. This can ensure the effectiveness of model training and verify the generalization ability of the proposed method in cross-regional land development category segmentation. The comparative experimental results of land development category segmentation based on micro remote sensing images are shown in Table 1. The experimental results show that the proposed method has significantly better comprehensive performance than the mainstream comparison model in the land development category segmentation task, with an F1 score of 89.74%, a mean intersection over union (mIoU) of 85.15%, and an overall accuracy (OA) of 89.06%. These three indicators are improved by 3.86%, 2.95%, and 5.38% respectively compared with the best comparison model SwinUNETR. Table 1 Comparison results of the proposed method in the dataset Figure 2 The segmentation effect of the method proposed in this invention is compared with other methods more intuitively. Figure 2The medium purple areas represent buildings, the black areas represent trees, and the light blue areas represent low vegetation areas. Figure 2 It can be seen that the method proposed in the present invention is closer to the actual land remote sensing image classification area in actual segmentation effect; Therefore, the above experimental results can prove the effectiveness and advancement of the method proposed in the present invention in the field of land development category segmentation in satellite remote sensing images.

[0033] Although the above describes the specific implementation methods of the invention in conjunction with the accompanying drawings, it does not limit the scope of protection of the invention. Based on the technical solution of the present invention, various modifications or variations that can be made by those skilled in the art without creative work are still within the scope of protection of the present invention.

Claims

1. A land development category segmentation method for remote sensing images based on edge detection, characterized by: Construct a land development category segmentation model based on satellite remote sensing images, input the satellite remote sensing land images taken by satellite remote sensing into the model, and obtain the land development category segmentation results, including the following steps: S1. Construct satellite image feature reconstruction module: This module includes depthwise separable convolution operations, The activation function, satellite image feature space reconstruction module and downsampling layer input the satellite remote sensing land image taken by satellite remote sensing into the satellite image feature reconstruction module and output the satellite image feature reconstruction features; S2. Constructing a super-resolution satellite image detail enhancement module: Constructing a super-resolution satellite image detail enhancement module based on multi-scale convolution, inputting the satellite image feature reconstruction features into the super-resolution satellite image detail enhancement module for super-resolution detail enhancement, and outputting super-resolution satellite image detail enhancement features; S3. Constructing an edge local perception attention module: This module includes an edge local perception attention module, which inputs the super-resolution satellite image detail enhancement features into the edge local perception attention module to capture local features and output edge local perception attention features; S4. Constructing a satellite image global capture module: This module includes a satellite image global capture module, which inputs the edge local perception attention features into the satellite image global capture module to enhance the global correlation of the features and outputs the satellite image global features; S5. Construct satellite image land category segmentation module: combine edge local perception attention features with satellite image global features, perform feature splicing operation, output fusion local global information features, and then input them into satellite image land category segmentation module to output satellite image land category segmentation image .

2. The method for land development category segmentation based on edge detection in remote sensing images according to claim 1, wherein: S1 is as follows: S1.

1. Satellite remote sensing land images The convolution kernel size input to the satellite image feature reconstruction module is The depth-wise separable convolutional layers and Activation function, output features ; S1.2、Characterize Input into the satellite image feature space reconstruction module in the satellite image feature reconstruction module, first change the feature by transposition operation The spatial structure of the output is transformed after the transposition operation ; Transformation Features After the convolution kernel size is The convolution layer and feature segmentation Operation, transform features The number of channels divided along the channel direction is the transformation feature Sub-features of and ; (1) For sub-features , the feature is convolution kernel size is Depthwise separable convolution operations, Non-linear activation function operation, Feature reconstruction operations, Feature scanning operation, Normalization operation, and reuse Reconstruction operation, output processed sub-features , the first time passing The feature reconstruction operation flattens the features into a one-dimensional vector, and the second pass The feature reconstruction operation restores the features into two-dimensional vectors; (2) For sub-features , after the convolution kernel size is The convolution layer and convolution kernel size are The depthwise separable convolutional layers and Non-linear activation function operation, outputting processed sub-features ; (3) The processed sub-features and Splicing fusion, after convolution kernel size is The convolution layer outputs features ; Then the features and features Splicing and fusion, output features ; S1.3、Characterize Input to the downsampling layer for downsampling operation, and output satellite image feature reconstruction features .

3. The method for land development category segmentation based on edge detection in remote sensing images according to claim 2, wherein: S2 is as follows: The super-resolution satellite image detail enhancement module based on multi-scale convolution specifically includes three parallel convolution layers to reconstruct the satellite image features. Input into the super-resolution satellite image detail enhancement module; S2.

1. Reconstructing satellite image features Input into three parallel convolutional layers respectively; For the convolutional layer of the first branch: features After the convolution kernel size is The depth-wise separable convolutional layer outputs features ;feature After the convolution kernel size is The depth-wise separable convolutional layer outputs features ; For the convolutional layer of the second branch: features After the convolution kernel size is The convolutional layer and Non-linear activation function, output features ; For the convolutional layer of the third branch: features After the convolution kernel size is The depth of the separable convolution layer, the features obtained are ,feature Splicing and fusion, output features ;feature After the convolution kernel size is The depth-wise separable convolutional layer outputs features ; The features output by the convolutional layers of the three branches ,feature and features Perform splicing and fusion to output features ,By splicing features through multi-scale convolutional layers, the multi-scale expression capability of ,detail features of super-resolution satellite images can be enhanced; S2.2, the characteristics After the global average pooling layer and the fully connected layer, the feature Fusion, output channel attention features ; Then the features After the convolution kernel size is After the convolution layer and the feature Fusion, output spatial attention features ; Then the channel attention feature and spatial attention features Splicing and fusion, and after convolution kernel size is The convolutional layer outputs the super-resolution satellite image detail enhancement features .

4. The method for land development category segmentation based on edge detection in remote sensing images according to claim 3, wherein S3 The details are as follows: Enhance the details of super-resolution satellite images Input to the edge local perception attention module, the edge local perception attention module consists of two branches, The two branches of the edge local perception attention module are respectively input to extract local channel features and local spatial features, and then the two features are spliced ​​and fused to obtain the edge local perception attention features. ; S3.

1. Enhance the details of super-resolution satellite images Input into the first branch of the edge local perception attention module, the feature After the global average pooling layer and the convolution kernel size is The convolutional layers and Activation function, output features , the features and features Perform broadcast multiplication and then combine the output features with the features Perform feature splicing and finally output local channel features ; S3.

2. Enhance the details of super-resolution satellite images Input to the second branch of the edge local perception attention module, feature After the convolution kernel size is The convolution layer outputs features , then the feature Input to three parallel depth-separable convolutional layers, the convolution kernel size is The depth of the separable convolution layer and the convolution kernel size are The depth of the separable convolution layer and the convolution kernel size is The depth-wise separable convolution layer is then used to concatenate the output features of the three parallel depth-wise separable convolution layers. , and then the features After the convolution kernel size is The convolutional layers and Activation function, output features , and finally the features and features Perform broadcast multiplication and then combine the output features with the features Perform feature splicing and finally output local spatial features ; S3.3, local channel features and local spatial features Splicing and fusion, output edge local perception attention features .

5. The method for land development category segmentation based on edge detection in remote sensing images according to claim 4 is characterized in that: The S4 steps are as follows: Enhance the details of super-resolution satellite images Input to the satellite image global capture module, the satellite image global capture module includes two branches, the feature Input the two branches of the satellite image global capture module respectively to obtain the first branch feature and the second branch feature, and then combine the two branch features with the feature Perform stitching and fusion to obtain the global features of satellite images ; S4.

1. Enhance the details of super-resolution satellite images Input into the first branch structure of the satellite image global capture module, first Perform dimension transposition operation to transform the features Rotate along the height dimension and output features , then the feature The input to the convolution kernel size is The depth of the separable convolution layer and batch normalization layer, the output features are combined with the features Perform feature splicing and output features ,feature Then pass The activation function operation and convolution kernel size are The convolution layer outputs features ; The features After the convolution kernel size is The depth of the separable convolution layer and batch normalization layer are used to combine the output features with the features Perform feature splicing and output features , the features After the convolution kernel size is The depth-wise separable convolutional layer, batch normalization layer and convolution kernel size are The convolution layer outputs the first branch feature ; S4.

2. Enhance the details of super-resolution satellite images Input into the second branch structure of the satellite image global capture module, first Perform dimension transposition operation to transform the features Rotate along the width dimension and output features , then the feature The input to the convolution kernel size is The depth of the separable convolution layer and batch normalization layer are combined with the obtained output to enhance the features Perform feature splicing and output features , and then the features go through The activation function operation and the convolution kernel size are The convolution layer outputs features ; Enhance the features After the convolution kernel size is The depth of the separable convolution layer and batch normalization layer are used to combine the output features with the features Perform feature splicing and output features ; The feature After the convolution kernel size is The depth-wise separable convolutional layer, batch normalization layer and convolution kernel size are The convolution layer outputs the second branch feature ; S4.3, the first branch feature , the second branch features and features Stitching and fusion to output global features of satellite images .

6. The method for land development category segmentation based on edge detection in remote sensing images according to claim 5, wherein: S5 is as follows: The edge local perception attention feature Global features of satellite images Perform splicing and fusion to output fusion local and global information features , and then the features Input to the satellite image land category segmentation module, after the convolution kernel size is The depth-wise separable convolutional layer and upsampling operation transform the features The resolution size is restored to match the image Same, use the image after restoring the resolution The extreme point positioning operation locates the specific land category to which each pixel in the image belongs, through The color mapping operation maps the specific land category to which the pixel belongs to different colors to generate a satellite image land category segmentation image. .

7. A remote sensing image land development category segmentation system based on edge detection, characterized by: Including for executing claims A module for processing instructions for each step in the method for land development category segmentation in remote sensing images based on edge detection as described in any one of the claims 1 to 5.

Images