Remote sensing image adaptive clustering segmentation method based on deep learning

Abstract

The invention discloses a remote sensing image adaptive clustering segmentation method based on deep learning, and the method comprises the following steps: obtaining a to-be-processed remote sensing image, and extracting image features through a feature extraction network; performing feature enhancement processing on the image features by using a feature enhancement network to obtain enhanced features; based on the enhanced features, constructing a weight matrix and generating a weighted graph; and processing the weighted graph through an adaptive clustering segmentation network to obtain a segmentation result of the remote sensing image to be processed. According to the method, deep learning is utilized to enhance remote sensing image features, and a weighted graph is constructed to realize adaptive clustering segmentation. Through optimization and two-stage segmentation during testing, the remote sensing image segmentation efficiency and precision are improved, and the method can be suitable for remote sensing image processing tasks in complex scenes.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing image segmentation, and more specifically to a remote sensing image adaptive clustering segmentation method based on deep learning. Background Art

[0002] With the development of remote sensing technology, remote sensing images are becoming increasingly important in many fields such as environmental monitoring, urban planning, and agricultural management. Remote sensing images not only provide rich spatial information, but also contain detailed ground features and semantic information, providing an important data source for decision support systems. However, due to the complex background and diverse targets of remote sensing images, which are easily affected by factors such as lighting and climate, automatic recognition and segmentation of targets still face many challenges.

[0003] Traditional pixel-based segmentation methods, such as threshold segmentation and edge detection, rely on low-level features such as color and texture, and are easily affected by noise and illumination changes, resulting in frequent mis-segmentation. Such methods are difficult to capture high-level spatial and semantic information, limiting their accuracy and reliability in practical applications.

[0004] Although region-based segmentation methods, such as the watershed algorithm and the region growing method, can overcome the influence of noise to a certain extent, they are often limited by the changes in target shape and size when processing complex scenes, and the computational complexity is high, making it difficult to achieve real-time processing. In addition, they usually require manual setting of parameters or rules, have poor adaptability, and are insufficiently responsive to the diverse needs of remote sensing image processing.

[0005] Existing deep learning models, such as U-Net and Mask R-CNN, perform well in instance segmentation tasks, but they are highly dependent on large-scale labeled data for training. When data is scarce or the quality of annotations is not high, the model is prone to overfitting and its generalization ability is weakened. At the same time, existing graph representation clustering methods face the problem of insufficient flexibility in remote sensing image processing. For example, the preset requirements for the number of clusters and the restrictions on weight processing in graph construction cannot fully adapt to the complex and changeable characteristics of remote sensing images.

[0006] Therefore, how to design an adaptive clustering segmentation method for remote sensing images based on deep learning, improve the efficiency and accuracy of remote sensing image segmentation, reduce dependence on large-scale labeled data, and be applicable to remote sensing image processing tasks in complex scenes is an urgent problem that technicians in this field need to solve. Summary of the invention

[0007] In view of this, the present invention provides a remote sensing image adaptive clustering segmentation method based on deep learning, which uses deep learning to enhance remote sensing image features and constructs a weighted graph to achieve adaptive clustering segmentation. Through test-time optimization and two-stage segmentation, the efficiency and accuracy of remote sensing image segmentation are improved, which can be applied to image processing tasks in complex scenes.

[0008] In order to achieve the above object, the present invention adopts the following technical solution:

[0009] A remote sensing image adaptive clustering segmentation method based on deep learning, comprising the following steps:

[0010] S1, obtaining the remote sensing image to be processed, and extracting image features through a feature extraction network;

[0011] S2, using the feature enhancement network to perform feature enhancement processing on the image features to obtain enhanced features;

[0012] S3. Based on the enhanced features, construct a weight matrix and generate a weighted graph;

[0013] S4. Process the weighted graph through an adaptive clustering segmentation network to obtain a segmentation result of the remote sensing image to be processed.

[0014] Furthermore, the S1 includes:

[0015] S11, for the remote sensing image I to be processed with a size of m×n and a number of channels c, the initialization layer is divided into patches, and flatten and map each patch to a fixed-dimensional vector, generating initial features through linear transformation; where p represents the patch size;

[0016] S12, normalizing the initial features through the first normalization layer;

[0017] S13, combine the multi-head attention mechanism to process the normalized features and perform residual connection to capture the dependencies between different patches;

[0018] S14, normalizing the features processed by the multi-head attention mechanism through the second normalization layer;

[0019] S15. Use a multi-layer perceptron MLP to extract features after the second normalization layer is processed, and perform residual connection to generate image features O.

[0020] Furthermore, the S2 includes:

[0021] S21, use global average pooling to perform local average pooling on the image feature O to obtain a feature vector F containing global information gp ;

[0022]

[0023] in, represents the normalization factor, O ij (c) represents the value of image feature O at position (i, j) and channel c;

[0024] S22, based on the feature vector F gp , feature extraction is performed through the first fully connected layer and the second fully connected layer, and the output F is obtained respectively. fc1 、F fc2 ;

[0025] F fc1 =ReLU(W1·F gp +b1)

[0026] F fc2 =ReLU(W2·F fc1 +b2)

[0027] Where W1 and W2 represent the weights of the first fully connected layer and the second fully connected layer, respectively; b1 and b2 represent the biases of the first fully connected layer and the second fully connected layer, respectively;

[0028] S23, output F of the second fully connected layer through the expansion layer fc2 Expand to the same spatial size as the image feature O to obtain the expanded feature F fc2 ′;

[0029] S24, combined with the fusion layer, the feature F fc2 ′ is added element-wise to the image feature O to generate the enhanced feature f.

[0030] Furthermore, in S3, a weight matrix is ​​constructed, including:

[0031] Based on the enhanced feature f, a weight matrix W of the weighted graph G is constructed; where G = (V, E), the node set V represents all patches, and the edge set E represents the similarity between different patches;

[0032] W = ff T

[0033] Among them, f T represents the transpose of the enhanced feature f.

[0034] Furthermore, constructing the weight matrix also includes:

[0035] The weight matrix W is subjected to zero threshold processing; the weight matrix W' after zero threshold processing is expressed as:

[0036]

[0037] in, is the indicator function, when ff T >0, is 1, otherwise, is 0.

[0038] Furthermore, constructing the weight matrix also includes:

[0039] Combined with the sensitivity hyperparameter α, the weight matrix W after zero threshold processing is adjusted; the adjusted weight matrix W" is expressed as:

[0040]

[0041] Among them, α∈[1,+∞).

[0042] Furthermore, in S4, the weighted graph is processed by an adaptive clustering segmentation network, including:

[0043] S411, performing convolution operation on the weighted graph through the graph convolution network GNN to obtain the node feature matrix f′;

[0044] f′=GNN(f,W′|Θ GNN )

[0045] Among them, Θ GNN Represents the parameters of the graph convolutional network GNN;

[0046] S412, processing the node feature matrix f′ by a multi-layer perceptron MLP to obtain a cluster allocation matrix S′;

[0047] S'=MLP(f'|Θ MLP )

[0048] Among them, Θ MLP Represents the parameters of the multi-layer perceptron MLP;

[0049] S413, based on the cluster allocation matrix S', perform foreground-background segmentation in combination with the cluster segmentation layer; the foreground-background segmentation includes:

[0050] Set the number of clusters k=2, cluster the nodes in the cluster assignment matrix S by using the normalized cut algorithm, and obtain the first clustering result;

[0051] Based on the first clustering result, the nodes in the image are classified into foreground and background, and a binary segmentation map is generated.

[0052] Furthermore, the loss function under the normalized cut algorithm is:

[0053]

[0054] Where Tr(·) represents the matrix trace, S′ T represents the transpose of the matrix S', D represents the diagonal matrix of the weight matrix W', ||·|| F represents the Frobenius norm of the matrix, K represents the number of clusters, and I K represents the K-dimensional identity matrix.

[0055] Furthermore, in S4, the weighted graph is processed by an adaptive clustering segmentation network, and further includes:

[0056] S421, performing convolution operation on the weighted graph through the graph convolution network GNN to obtain the node feature matrix f";

[0057] f”=GNN(f,W”|Θ GNN )

[0058] S422, processing the node feature matrix f" through a multi-layer perceptron MLP to obtain a cluster allocation matrix S";

[0059] S' = MLP(f'|Θ MLP )

[0060] S423, based on the cluster allocation matrix S', foreground semantic part segmentation is performed in combination with the cluster segmentation layer; the foreground semantic part segmentation includes:

[0061] Clustering the nodes in the foreground area by using a related clustering algorithm to obtain a second clustering result;

[0062] Based on the second clustering result, the nodes in the foreground area are classified into multiple semantic parts to generate a multi-category segmentation map.

[0063] Furthermore, the loss function under the relevant clustering algorithm is:

[0064] L CC =-Tr(W”S”S” T )

[0065] Among them, S T represents the transpose of the matrix S'.

[0066] It can be seen from the above technical solution that compared with the prior art, the technical solution of the present invention has the following advantages:

[0067] Beneficial effects:

[0068] 1. This method achieves efficient and accurate segmentation of remote sensing images through feature extraction, feature enhancement, weighted graph construction, and adaptive clustering segmentation network steps. It can effectively improve the automation level of remote sensing image processing and provide high-quality basic data for subsequent related specific applications.

[0069] 2. A method combining feature extraction network and feature enhancement network is introduced to obtain richer and more representative image features. It not only helps to capture the dependencies between different patches, but also enhances the expression ability of local features, thus providing a more reliable data basis for subsequent clustering segmentation.

[0070] 3. In the process of constructing the weight matrix and generating the weighted graph, the noise interference is effectively reduced and the similarity between different patches is highlighted by performing zero threshold processing on the weight matrix and combining sensitivity hyperparameter adjustment. This ensures that the foreground and background parts of the image can be accurately identified when performing adaptive clustering segmentation, and it also has good adaptability and robustness for semantic segmentation in complex scenes.

[0071] 4. The adaptive clustering segmentation network can not only achieve efficient node feature extraction and cluster allocation through the combination of graph convolutional network GNN and multi-layer perceptron MLP, but also flexibly adjust the clustering strategy according to different task requirements, such as foreground-background binary segmentation or foreground semantic part segmentation. It not only meets the needs of diversified applications, but also ensures the quality and accuracy of the segmentation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0073] Figure 1 A flow chart of a remote sensing image adaptive clustering segmentation method based on deep learning provided by an embodiment of the present invention;

[0074] Figure 2 A schematic diagram of a process of extracting image features through a feature extraction network provided by an embodiment of the present invention;

[0075] Figure 3 A schematic diagram of a process of performing feature enhancement processing on image features using a feature enhancement network provided in an embodiment of the present invention;

[0076] Figure 4 A schematic diagram of a process of processing a weighted graph through an adaptive clustering segmentation network provided in an embodiment of the present invention;

[0077] Figure 5 A schematic diagram of a remote sensing image segmentation process for a specific target, an airplane, provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0078] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0079] Embodiment 1;

[0080] like Figure 1 As shown, this embodiment provides a remote sensing image adaptive clustering segmentation method based on deep learning, comprising the following steps:

[0081] S1, obtaining the remote sensing image to be processed, and extracting image features through a feature extraction network;

[0082] S2, using the feature enhancement network to perform feature enhancement processing on the image features to obtain enhanced features;

[0083] S3. Based on the enhanced features, construct a weight matrix and generate a weighted graph;

[0084] S4. Process the weighted graph through an adaptive clustering segmentation network to obtain a segmentation result of the remote sensing image to be processed.

[0085] This method extracts and enhances remote sensing image features through deep learning, constructs a weighted graph that reflects the similarity between regions, and realizes adaptive cluster segmentation without pre-defining the number of clusters. It adopts a test-time optimization strategy and improves the accuracy of foreground object details and background segmentation through two-stage segmentation, which significantly improves the efficiency and accuracy of remote sensing image segmentation and can be applied to image processing tasks in complex scenes.

[0086] The following further describes each step in the above method in detail;

[0087] like Figure 2 As shown, the remote sensing image to be processed is obtained, and the image features are extracted through the feature extraction network; specifically, it includes:

[0088] S11, for the remote sensing image I to be processed with a size of m×n and a number of channels c, the initialization layer is divided into patches, and flatten and map each patch to a fixed-dimensional vector, generating initial features through linear transformation; where p represents the patch size;

[0089] S12, normalizing the initial features through the first normalization layer;

[0090] S13, combine the multi-head attention mechanism to process the normalized features and perform residual connection to capture the dependencies between different patches;

[0091] S14, normalizing the features processed by the multi-head attention mechanism through the second normalization layer;

[0092] S15. Use a multi-layer perceptron MLP to extract features after the second normalization layer is processed, and perform residual connection to generate image features O.

[0093] This step effectively improves the feature expression ability of remote sensing images through a meticulous feature extraction process. By dividing into patches, linear transformation, normalization processing, and combining a multi-head attention mechanism, it not only captures the local features of the image, but also establishes the dependency relationship between different patches, providing richer and more accurate feature information for subsequent feature enhancement and clustering segmentation.

[0094] like Figure 3 As shown, the feature enhancement network is used to perform feature enhancement processing on the image features to obtain enhanced features; specifically, it includes:

[0095] S21, use global average pooling to perform local average pooling on the image feature O to obtain a feature vector F containing global information gp ;

[0096]

[0097] in, represents the normalization factor, O ij (c) represents the value of image feature O at position (i, j) and channel c;

[0098] S22, based on the feature vector F gp , feature extraction is performed through the first fully connected layer and the second fully connected layer, and the output F is obtained respectively. fc1 、F fc2 ;

[0099] F fc1 =ReLU(W1·F gp +b1)

[0100] F fc2 =ReLU(W2·F fc1 +b2)

[0101] Where W1 and W2 represent the weights of the first fully connected layer and the second fully connected layer, respectively; b1 and b2 represent the biases of the first fully connected layer and the second fully connected layer, respectively;

[0102] S23, output F of the second fully connected layer through the expansion layer fc2 Expand to the same spatial size as the image feature O to obtain the expanded feature F fc2 ′;

[0103] S24, combined with the fusion layer, the feature F fc2 ′ is added element-wise to the image feature O to generate the enhanced feature f.

[0104] This step can significantly improve the expressiveness of image features. Global average pooling is used to capture global information, and features are further extracted and processed through the fully connected layer. Finally, the processed features are combined with the original features to generate enhanced features that contain both global context and local details. The enhanced features provide richer information for subsequent clustering segmentation, which helps to improve the accuracy and robustness of segmentation.

[0105] In this embodiment S3, based on the enhanced features, a weight matrix is ​​constructed and a weighted graph is generated; specifically,

[0106] Construct a weight matrix, including:

[0107] Based on the enhanced features f, a weight matrix W of the weighted graph G is constructed; where G = (V, E), the node set V represents all patches, and the edge set E represents the similarity between different patches;

[0108] W = ff T

[0109] Among them, f T represents the transpose of the enhanced feature f.

[0110] Furthermore, it also includes:

[0111] The weight matrix W is subjected to zero threshold processing; the weight matrix W' after zero threshold processing is expressed as:

[0112]

[0113] in, is the indicator function, when ff T >0, is 1, otherwise, is 0.

[0114] Furthermore, it also includes:

[0115] Combined with the sensitivity hyperparameter α, the weight matrix W after zero threshold processing is adjusted; the adjusted weight matrix W" is expressed as:

[0116]

[0117] Among them, α∈[1,+∞).

[0118] In the normalized cut algorithm, only positive weights are usually accepted. However, in the correlation clustering algorithm, negative weights carry important information for the clustering target and represent the repulsive relationship between nodes. Therefore, the negative values ​​in the weight matrix are set to zero through threshold processing, and the sensitivity hyperparameter is introduced to adjust the repulsive force in the clustering process, thereby affecting the number and quality of clusters.

[0119] The sensitivity hyperparameter allows controlling the repulsion between nodes during the clustering process, allowing the number of clusters to be automatically adjusted in the relevant clustering algorithm according to the intrinsic structure of the data. Among them, higher k values ​​correspond to more clusters, and lower k values ​​lead to a higher number of clusters, which strengthens the repulsion between nodes and allows more independent small clusters to form.

[0120] like Figure 4 As shown, the weighted graph is processed by an adaptive clustering segmentation network to obtain the segmentation result of the remote sensing image to be processed; specifically,

[0121] The weighted graph is processed through an adaptive clustering segmentation network, including:

[0122] S411, performing convolution operation on the weighted graph through the graph convolution network GNN to obtain the node feature matrix f′;

[0123] f′=GNN(f,W′|Θ GNN )

[0124] Among them, Θ GNN Represents the parameters of the graph convolutional network GNN;

[0125] S412, processing the node feature matrix f′ by a multi-layer perceptron MLP to obtain a cluster allocation matrix S′;

[0126] S'=MLP(f'|Θ MLP )

[0127] Among them, Θ MLP Represents the parameters of the multi-layer perceptron MLP;

[0128] S413, based on the cluster allocation matrix S', perform foreground-background segmentation in combination with the cluster segmentation layer; the foreground-background segmentation includes:

[0129] Set the number of clusters k=2, cluster the nodes in the cluster assignment matrix S by using the normalized cut algorithm, and obtain the first clustering result;

[0130] Based on the first clustering result, the nodes in the image are classified into foreground and background, and a binary segmentation map is generated.

[0131] Among them, the loss function under the normalized cutting algorithm is:

[0132]

[0133] Where Tr(·) represents the matrix trace, S′ T represents the transpose of the matrix S', D represents the diagonal matrix of the weight matrix W', ||·|| F represents the Frobenius norm of the matrix, K represents the number of clusters, and I K represents the K-dimensional identity matrix.

[0134] In this loss function, the first term Promotes clustering of strongly connected components. Encourage orthogonality and similar size between cluster assignments. Used to adjust the overall balance.

[0135] Furthermore, the weighted graph is processed by an adaptive clustering segmentation network, which also includes:

[0136] S421, performing convolution operation on the weighted graph through the graph convolution network GNN to obtain the node feature matrix f";

[0137] f”=GNN(f,W”|Θ GNN )

[0138] S422, processing the node feature matrix f" through a multi-layer perceptron MLP to obtain a cluster allocation matrix S";

[0139] S' = MLP(f'|Θ MLP )

[0140] S423, based on the cluster allocation matrix S', foreground semantic part segmentation is performed in combination with the cluster segmentation layer; the foreground semantic part segmentation includes:

[0141] Clustering the nodes in the foreground area by using a related clustering algorithm to obtain a second clustering result;

[0142] Based on the second clustering result, the nodes in the foreground area are classified into multiple semantic parts to generate a multi-category segmentation map.

[0143] Among them, the loss function under the relevant clustering algorithm is:

[0144] L CC =-Tr(W”S”S” T )

[0145] Among them, S T represents the transpose of the matrix S'.

[0146] In this step, during the semantic part segmentation phase, the test-time optimization strategy is used. In conventional neural network image segmentation, a model is first trained with a large number of labeled data sets, and then the model is used to predict new samples. Under the test-time optimization strategy, when processing each image, the model weights are optimized based on the processing results of the previous image to better adapt to the characteristics and segmentation requirements of the current image.

[0147] Even without large-scale labeled data, it can still effectively learn the features of new images and make more accurate segmentation. This not only improves processing efficiency, but also simplifies the workflow, allowing the model to more flexibly cope with different remote sensing image segmentation tasks.

[0148] The remote sensing image adaptive clustering segmentation method of this embodiment extracts and enhances image features through deep learning technology, constructs a weighted graph reflecting the similarity between regions, and realizes adaptive clustering segmentation without pre-defining the number of clusters. This method adopts a test-time optimization strategy and combines a two-stage segmentation process to significantly improve the efficiency and accuracy of remote sensing image segmentation. It is particularly suitable for image processing tasks in complex scenes and has good adaptability and scalability.

[0149] Embodiment 2;

[0150] In this embodiment, the remote sensing image adaptive clustering segmentation method based on deep learning is applied to the recognition and segmentation of a specific target, an airplane on an airport runway;

[0151] like Figure 5 As shown, the specific steps include:

[0152] 1) Obtain the remote sensing image to be processed and extract features;

[0153] Get a high-resolution remote sensing satellite image of an airport area taken from high altitude. The image may contain multiple aircraft parked on the runway or taxiway, as well as other structures in the background, such as related buildings, vegetation, etc.

[0154] The image is divided into patches of fixed size p×p (such as 16×16 pixels), and then each patch is flattened to a one-dimensional vector and mapped to an initial feature representation of a fixed dimension c (such as 768 dimensions) through a linear transformation. These initial features are processed using layer normalization to ensure that the feature distribution is stable; then a multi-head self-attention mechanism is used to capture the long-distance dependencies between different patches, which helps to emphasize the uniqueness and continuity of the aircraft shape.

[0155] Finally, the multi-layer perceptron (MLP) is used to further process the features and generate the final image features, which can well characterize the existence of the aircraft and its detailed features.

[0156] 2) Feature enhancement processing;

[0157] The image features are globally averaged and pooled to obtain a compact feature vector containing all channel information in the entire image. Two consecutive fully connected layers, combined with the ReLU activation function, are used to further extract key features that can better distinguish aircraft from other background elements.

[0158] The features processed by the fully connected layer are expanded back to the original spatial size and added element by element to the original features to generate enhanced features. This process allows the model to focus on local details and global context information at the same time.

[0159] 3) Construct a weight matrix and generate a weighted graph;

[0160] Based on the enhanced features, the similarity between each pair of patches is calculated to form a correlation matrix that reflects the strength of the association between them. This matrix serves as the basis for the weight matrix of the weighted graph, where nodes represent individual patches and edge weights reflect the relationship between patches.

[0161] Considering that negative correlation may also carry important clues about the location of the aircraft in some cases, in addition to a simple threshold truncation of negative weights, a sensitivity hyperparameter is introduced to adjust the negative values ​​in the weight matrix to better adapt to different application scenarios.

[0162] 4) Adaptive clustering segmentation to obtain results;

[0163] Use one or more layers of graph convolutional networks to convolve the weighted graph and obtain the node feature matrix. Here, the role of the graph convolutional network is to propagate feature information according to the graph structure and help the model understand the spatial relationship between different patches.

[0164] Perform cluster assignment and segmentation:

[0165] In foreground-background segmentation, the number of clusters k is set to 2, and the cluster assignment matrix S is processed by the normalized cut algorithm to classify the nodes in the image into foreground (aircraft) and background (non-aircraft). Since aircraft usually occupy a smaller proportion and have a more obvious boundary, this dichotomy can effectively distinguish targets.

[0166] In the foreground semantic part segmentation, the separated foreground area is further segmented using the relevant clustering algorithm. At this time, a larger k value (such as k = 4) may be set to identify different components of the aircraft, such as the fuselage, wings, etc., so as to generate a more detailed multi-category segmentation map.

[0167] During the entire process, the model will dynamically adjust its parameters based on the learning results of the previous image to adapt to the characteristics of the current image, ensuring that even aircraft with different shapes or postures can be accurately identified and segmented. In addition, by gradually contacting each new image, the model can continuously improve its performance without relying on large-scale pre-training data.

[0168] Through the above detailed process, this method not only improves the accuracy of aircraft recognition and segmentation, but also enhances the flexibility and generalization ability of the system, making it applicable to various complex remote sensing image environments. It does not need to predefine the number of clusters and can automatically adapt to targets of different sizes and locations, which is very suitable for processing scenes with multiple changing factors such as airports.

[0169] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0170] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A remote sensing image adaptive clustering segmentation method based on deep learning, characterized in that: The following steps are involved: S1, obtaining the remote sensing image to be processed, and extracting image features through a feature extraction network; S2, using the feature enhancement network to perform feature enhancement processing on the image features to obtain enhanced features; S3. Based on the enhanced features, construct a weight matrix and generate a weighted graph; S4. Process the weighted graph through an adaptive clustering segmentation network to obtain a segmentation result of the remote sensing image to be processed.

2. The method for adaptive clustering and segmentation of remote sensing images based on deep learning according to claim 1, characterized in that: Said S1 comprises: S11, for the remote sensing image I to be processed with a size of m×n and a number of channels c, the initialization layer is divided into patches, and flatten and map each patch to a fixed-dimensional vector, generating initial features through linear transformation; where p represents the patch size; S12, normalizing the initial features through the first normalization layer; S13, combine the multi-head attention mechanism to process the normalized features and perform residual connection to capture the dependencies between different patches; S14, normalizing the features processed by the multi-head attention mechanism through the second normalization layer; S15. Use a multi-layer perceptron MLP to extract features after the second normalization layer is processed, and perform residual connection to generate image features O.

3. The method for adaptive clustering and segmentation of remote sensing images based on deep learning according to claim 1, characterized in that: The S2 comprises: S21, use global average pooling to perform local average pooling on the image feature O to obtain a feature vector F containing global information gp ; in, represents the normalization factor, O ij (c) represents the value of image feature O at position (i, j) and channel c; S22, based on the feature vector F gp , feature extraction is performed through the first fully connected layer and the second fully connected layer, and the output F is obtained respectively. fc1 、F fc2 ; F fc1 =ReLU(W1·F gp +b1) <h2 style=";text-align:left;direction:ltr">F<h2 style=";text-align:left;direction:ltr"> fc2 <h2 style=";text-align:left;direction:ltr"> =ReLU(W2·F<h2 style=";text-align:left;direction:ltr"> fc1 <h2 style=";text-align:left;direction:ltr"> +b2) Where W1 and W2 represent the weights of the first fully connected layer and the second fully connected layer, respectively; b1 and b2 represent the biases of the first fully connected layer and the second fully connected layer, respectively; S23, output F of the second fully connected layer through the expansion layer fc2 Expand to the same spatial size as the image feature O to obtain the expanded feature F fc2 ′; S24, combined with the fusion layer, the feature F fc2 ′ is added element-wise to the image feature O to generate the enhanced feature f.

4. The method for adaptive clustering and segmentation of remote sensing images based on deep learning according to claim 1, characterized in that: In S3, a weight matrix is ​​constructed, including: Based on the enhanced features f, a weight matrix W of the weighted graph G is constructed; where G = (V, E), the node set V represents all patches, and the edge set E represents the similarity between different patches; W=ff T Among them, f T represents the transpose of the enhanced feature f.

5. The method for adaptive clustering and segmentation of remote sensing images based on deep learning according to claim 4, characterized in that: Constructing the weight matrix also includes: The weight matrix W is subjected to zero threshold processing; the weight matrix W' after zero threshold processing is expressed as: in, is the indicator function, when ff T >0, is 1, otherwise, is 0.

6. The method for adaptive clustering and segmentation of remote sensing images based on deep learning according to claim 4, characterized in that: Constructing the weight matrix also includes: Combined with the sensitivity hyperparameter α, the weight matrix W after zero threshold processing is adjusted; the adjusted weight matrix W" is expressed as: Among them, α∈[1,+∞).

7. The method for adaptive clustering and segmentation of remote sensing images based on deep learning according to claim 1, characterized in that: In S4, the weighted graph is processed by an adaptive clustering segmentation network, including: S411, performing convolution operation on the weighted graph through the graph convolution network GNN to obtain the node feature matrix f′; f′=GNN(f,W'|Θ GNN ) Among them, Θ GNN Represents the parameters of the graph convolutional network GNN; S412, processing the node feature matrix f′ by a multi-layer perceptron MLP to obtain a cluster allocation matrix S′; S'=MLP(f′|Θ MLP ) Among them, Θ MLP Represents the parameters of the multi-layer perceptron MLP; S413, based on the cluster allocation matrix S', perform foreground-background segmentation in combination with the cluster segmentation layer; the foreground-background segmentation includes: Set the number of clusters k=2, cluster the nodes in the cluster assignment matrix S by using the normalized cut algorithm, and obtain the first clustering result; Based on the first clustering result, the nodes in the image are classified into foreground and background, and a binary segmentation map is generated.

8. The method for adaptive clustering and segmentation of remote sensing images based on deep learning according to claim 7, characterized in that: The loss function under the normalized cut algorithm is: Where Tr(·) represents the matrix trace, S′ T represents the transpose of the matrix S', D represents the diagonal matrix of the weight matrix W', ||·|| F represents the Frobenius norm of the matrix, K represents the number of clusters, and I K represents the K-dimensional identity matrix.

9. The remote sensing image adaptive clustering segmentation method based on deep learning according to claim 7, characterized in that: In S4, the weighted graph is processed by an adaptive clustering segmentation network, and further includes: S421, performing convolution operation on the weighted graph through the graph convolution network GNN to obtain the node feature matrix f"; f”=GNN(f,W”|Θ GNN ) S422, processing the node feature matrix f" through a multi-layer perceptron MLP to obtain a cluster allocation matrix S"; S”=MLP(f′|Θ MLP ) S423, based on the cluster allocation matrix S', foreground semantic part segmentation is performed in combination with the cluster segmentation layer; the foreground semantic part segmentation includes: Clustering the nodes in the foreground area by using a related clustering algorithm to obtain a second clustering result; Based on the second clustering result, the nodes in the foreground area are classified into multiple semantic parts to generate a multi-category segmentation map.

10. The method for adaptive clustering and segmentation of remote sensing images based on deep learning according to claim 7, characterized in that: The loss function under the relevant clustering algorithm is: L CC =-Tr(W”S”S” T ) Among them, S T represents the transpose of the matrix S'.

Images