Remote sensing image change detection method and system based on lightweight structure perception

Abstract

The invention relates to the field of remote sensing image processing, in particular to a remote sensing image change detection method and system based on lightweight structure perception, and the method comprises the steps: carrying out the preliminary feature extraction of a first remote sensing image, and obtaining a first preliminary feature; performing deep feature extraction on the first preliminary feature to obtain a first deep semantic feature; performing preliminary feature extraction on the second remote sensing image to obtain a second preliminary feature; performing deep feature extraction on the second preliminary feature to obtain a second deep semantic feature; performing enhancement processing on the first and second deep semantic features to obtain enhanced features; performing feature splicing on the first and second deep semantic features and the enhanced features, performing preliminary feature fusion on the spliced features, performing deep fusion on the preliminarily fused features, and classifying the deep fused features to obtain a change detection result between the dual-temporal remote sensing images. The calculation complexity can be remarkably reduced while the change detection precision is guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical fields of remote sensing image processing and computer vision, and particularly to a remote sensing image change detection method and system based on lightweight structure perception. Background Art

[0002] Remote sensing image change detection aims to analyze dual-temporal remote sensing images of the same geographical area to extract surface change information. This technology has been widely applied in fields such as river channel evolution, land cover change analysis, and disaster assessment. In recent years, convolutional neural network (CNNs) technology based on deep learning has shown excellent performance in change detection tasks, such as fully convolutional network (FCN) and multi-scale feature fusion network based on attention mechanism. However, convolutional neural networks have limited ability in modeling long-range contextual dependencies in dual-temporal remote sensing images, which affects their detection performance in complex scenes.

[0003] Vision Transformer (ViT) has been gradually applied to change detection tasks in recent years due to its ability to capture long-range dependencies using self-attention mechanism and has shown good performance. However, the computational complexity of traditional Transformer networks grows quadratically in high-resolution remote sensing images, resulting in low training and inference efficiency. In addition, existing Transformer-based change detection methods usually ignore the importance of fine-grained features, leading to insufficient detection ability for small target changes and poor performance in edge integrity and internal compactness for large target changes. Therefore, how to design a lightweight change detection method that can efficiently capture long-range contextual information and enhance the expression of fine-grained change features is the key challenge in current research. Summary of the Invention

[0004] To solve the deficiencies of the prior art, the present invention provides a remote sensing image change detection method and system based on lightweight structure perception; by designing a cross-dimensional interaction self-attention module and a structure perception enhancement module with linear complexity, this method can significantly reduce the computational complexity while ensuring the change detection accuracy. This method can effectively improve the edge integrity and internal compactness of dual-temporal remote sensing image change detection, and is of great significance for fields such as river channel evolution, land use monitoring, and disaster assessment.

[0005] On the one hand, a remote sensing image change detection method based on lightweight structure perception is provided, including: Obtain a pair of dual-temporal remote sensing images, where the dual-temporal remote sensing images include: a first remote sensing image and a second remote sensing image; Input the dual-temporal remote sensing image into the trained remote sensing image change detection model to obtain the change detection result of the dual-temporal remote sensing image; Among them, the trained remote sensing image change detection model is used to perform preliminary feature extraction on the first remote sensing image to obtain the first preliminary feature; perform deep feature extraction on the first preliminary feature to obtain the first deep semantic feature; perform preliminary feature extraction on the second remote sensing image to obtain the second preliminary feature; perform deep feature extraction on the second preliminary feature to obtain the second deep semantic feature; perform enhancement processing on the first deep semantic feature and the second deep semantic feature to obtain the enhanced feature; perform feature splicing on the first deep semantic feature, the second deep semantic feature and the enhanced feature, perform preliminary feature fusion on the spliced feature, perform deep fusion on the preliminary fusion feature, and perform classification on the deep fusion feature to obtain the change detection result between the dual-temporal remote sensing images.

[0006] On the other hand, a remote sensing image change detection system based on lightweight structure perception is provided, including: An acquisition module, which is configured to: acquire a pair of dual-temporal remote sensing images, and the dual-temporal remote sensing images include: a first remote sensing image and a second remote sensing image; A detection module, which is configured to: input the dual-temporal remote sensing images into the trained remote sensing image change detection model to obtain the change detection result of the dual-temporal remote sensing images; Among them, the trained remote sensing image change detection model is used to perform preliminary feature extraction on the first remote sensing image to obtain the first preliminary feature; perform deep feature extraction on the first preliminary feature to obtain the first deep semantic feature; perform preliminary feature extraction on the second remote sensing image to obtain the second preliminary feature; perform deep feature extraction on the second preliminary feature to obtain the second deep semantic feature; perform enhancement processing on the first deep semantic feature and the second deep semantic feature to obtain the enhanced feature; perform feature splicing on the first deep semantic feature, the second deep semantic feature and the enhanced feature, perform preliminary feature fusion on the spliced feature, perform deep fusion on the preliminary fusion feature, and perform classification on the deep fusion feature to obtain the change detection result between the dual-temporal remote sensing images.

[0007] The above technical solution has the following advantages or beneficial effects: The present invention designs a cross-dimensional interaction self-attention module, which uses a linear complexity method to replace the traditional self-attention mechanism, effectively reducing the computational complexity and enhancing the ability to capture global information. In addition, the present invention proposes a structure-aware enhancement module, which extracts differences and aggregates details of changing features through a two-branch structure, realizing double enhancement of the edge integrity of large targets and the changes of small targets. Further, the present invention combines an attention-based feature fusion module to efficiently integrate the deep semantic features of dual-temporal remote sensing images and generate fine change detection results. The present invention can significantly improve the detection accuracy while reducing the consumption of computing resources, has excellent robustness and generalization ability, and has broad application prospects in the fields of land use monitoring, disaster assessment, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] The accompanying drawings forming a part of this specification are used to provide a further understanding of the present invention. The schematic embodiments and descriptions thereof of the present invention are used to explain the present invention and do not constitute an improper limitation of the present invention.

[0009] Figure 1 It is a flowchart of the method for the first embodiment; Figure 2 It is a schematic diagram of the network structure of the remote sensing image change detection model for the first embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0010] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0011] First Embodiment This embodiment provides a method for remote sensing image change detection based on lightweight structure awareness; As Figure 1 shown, the method for remote sensing image change detection based on lightweight structure awareness includes: S101: Obtain a pair of dual-temporal remote sensing images, where the dual-temporal remote sensing images include: a first remote sensing image and a second remote sensing image; S102: Input the dual-temporal remote sensing images into a trained remote sensing image change detection model to obtain the change detection result of the dual-temporal remote sensing images; Among them, the trained remote sensing image change detection model is used to perform preliminary feature extraction on the first remote sensing image to obtain the first preliminary feature; perform deep feature extraction on the first preliminary feature to obtain the first deep semantic feature; perform preliminary feature extraction on the second remote sensing image to obtain the second preliminary feature; perform deep feature extraction on the second preliminary feature to obtain the second deep semantic feature; perform enhancement processing on the first deep semantic feature and the second deep semantic feature to obtain the enhanced feature; perform feature splicing on the first deep semantic feature, the second deep semantic feature and the enhanced feature, perform preliminary feature fusion on the spliced features, perform deep fusion on the preliminary fusion features, and perform classification on the deep fusion features to obtain the change detection result between the two-temporal remote sensing images.

[0012] Further, for the step of obtaining a pair of two-temporal remote sensing images, the two-temporal remote sensing images include: a first remote sensing image and a second remote sensing image, and the two-temporal remote sensing images refer to remote sensing images taken at different time points in the same geographical area.

[0013] The input of the present invention is a pair of two-temporal remote sensing images, which respectively represent the same geographical area at two different time points. Let the input two-temporal images be and , and these images are used to capture surface change information, that is, to detect the changes that occur between two moments.

[0014] Further, in S102: input the two-temporal remote sensing images into the trained remote sensing image change detection model to obtain the change detection result of the two-temporal remote sensing images. Among them, the training process of the trained remote sensing image change detection model includes: Construct a training set, and the training set is several pairs of two-temporal remote sensing images with known change detection results; Input the training set into the remote sensing image change detection model to train the remote sensing image change detection model. When the loss function value of the remote sensing image change detection model no longer decreases, or the number of iterations reaches the set number of times, stop training to obtain the trained remote sensing image change detection model.

[0015] Further, as Figure 2 shown, the network structure of the remote sensing image change detection model includes: A first depthwise separable convolutional layer and a second depthwise separable convolutional layer; The input end of the first depthwise separable convolutional layer is used to input the first remote sensing image; the output end of the first depthwise separable convolutional layer is connected to the input end of the first cross-dimensional interaction self-attention module; The input end of the second depthwise separable convolutional layer is used to input the second remote sensing image; the output end of the second depthwise separable convolutional layer is connected to the input end of the second cross-dimensional interactive self-attention module; The output ends of the first cross-dimensional interactive self-attention module and the second cross-dimensional interactive self-attention module are both connected to the input end of the structure-aware enhancement module; The output ends of the first cross-dimensional interactive self-attention module, the second cross-dimensional interactive self-attention module, and the output end of the structure-aware enhancement module are all connected to the input end of the splicing unit; the output end of the splicing unit is also connected to the input end of the third depthwise separable convolutional layer; the output end of the third depthwise separable convolutional layer is connected to the input end of the third cross-dimensional interactive self-attention module; the output end of the third cross-dimensional interactive self-attention module is connected to the input end of the classification head; the output end of the classification head outputs the change detection result of the dual-temporal remote sensing image.

[0016] Furthermore, the preliminary feature extraction of the first remote sensing image to obtain the first preliminary feature is achieved through the first depthwise separable convolutional layer.

[0017] It should be understood that the first depthwise separable convolutional layer includes: Depth convolution. The difference between depth convolution and standard convolution is that the convolution kernel of depth convolution is in single-channel mode, and it is necessary to perform spatial filtering on each channel of the input feature map separately without information fusion between channels; Pointwise convolution. Pointwise convolution is essentially a dimension transformation using a 1×1 convolution kernel. The role of pointwise convolution is to aggregate the feature map output by depth convolution in the channel dimension to generate a new feature representation.

[0018] Depth convolution only performs spatial filtering, and pointwise convolution only performs channel fusion. The depthwise separable convolution composed of the two reduces the computational amount significantly while maintaining good feature expression ability.

[0019] It should be understood that in order to efficiently extract the preliminary features in the image, depthwise separable convolution is used as the feature extraction module, which reduces the consumption of computing resources while improving the efficiency of feature learning. The preliminarily extracted features lay the foundation for the execution of subsequent steps (such as the cross-dimensional interactive self-attention module and the structure-aware enhancement module).

[0020] Furthermore, the deep feature extraction of the first preliminary feature to obtain the first deep semantic feature is achieved through the first cross-dimensional interactive self-attention module.

[0021] The first cross-dimensional interactive self-attention module has three parallel branches, namely the channel-channel interactive attention branch, the channel-height interactive attention branch, and the channel-width interactive attention branch, which are used for information interaction between different dimensions of an image; Let the initial feature input to the first cross-dimensional interactive self-attention module be , where is the number of channels, and are the height and width of the feature map respectively; For the channel-channel attention branch, the input generates query Q, key K, and value V through a depthwise separable convolutional layer to enhance the local context modeling ability, and is formulated as: ; ; ; where, , and represent the depthwise separable convolutional operations for generating query Q, key K, and value V respectively; After obtaining query Q, key K, and value V, calculate the channel-channel interactive attention, and the calculation formula is as follows:

[0022] where , , Generate the channel-channel attention feature map through matrix multiplication, and pass through function and scaling factor to improve stability. The scaling factor is learnable and automatically updated following the training process; The operation process of the channel-height interactive attention branch:

[0023] The operation process of the channel-width interactive attention branch:

[0024] where, , represent the channel-height attention feature map and the channel-width attention feature map respectively, represents the parallel operation of max pooling and average pooling, and the results of the parallel operation are concatenated, represents a 1×1 convolution, and respectively represent the input tensors after dimensional transformation, corresponding to the feature maps in the height and width directions respectively, represents function, represents the parallel operations of max pooling and average pooling: Taking as an example, the max pooling operation retains the maximum value in the first channel, obtaining a max pooling result with dimensions 1×C×H. Similarly, the average pooling operation retains the average value in the first channel, obtaining an average pooling result with dimensions 1×C×H. Subsequently, the two pooling results are concatenated to obtain a concatenated result with dimensions 2×C×H.

[0025] Obtain the channel-height attention feature map and the channel-width attention feature map After that, generate the final global attention map , and the calculation method is as follows:

[0026] Among them, , , are adjustable weight hyperparameters used to balance the contributions of attentions in different dimensions.

[0027] After obtaining the final global attention map , apply it to the input feature to generate the updated feature representation :

[0028] Among them, is the deep feature.

[0029] Perform deep feature extraction on the first preliminary feature to obtain the first deep semantic feature denoted as , perform deep feature extraction on the first preliminary feature to obtain the first deep semantic feature denoted as ; And The generation processes of are the same as those of the deep feature

[0030] It should be understood that the computational complexity of the self-attention mechanism is , which is very expensive when processing high-resolution remote sensing images. The cross-dimensional interaction self-attention module adopted in the present invention introduces an interaction mechanism with linear complexity through depthwise separable convolution, average pooling, and max pooling, reducing the computational complexity to , this optimization makes the calculation more efficient while maintaining the ability to efficiently extract global information. Through multi-dimensional interaction and the optimized attention mechanism, this method can effectively capture the global context information in remote sensing images. Whether it is the small-area details or the large-scale global features in the image, the present invention can effectively model and learn through its self-attention mechanism. Finally, the present invention outputs an image representation after deep feature extraction. , this representation contains rich semantic information from the dual-temporal remote sensing images, providing deeper feature support for subsequent structure-aware enhancement and change detection.

[0031] It should be understood that the deep semantic features of the dual-temporal remote sensing images are extracted through the cross-dimensional interaction self-attention module. The cross-dimensional interaction self-attention module effectively reduces the computational complexity through the interaction attention mechanism with linear complexity while strengthening the global information extraction ability.

[0032] Furthermore, the preliminary feature extraction of the second remote sensing image to obtain the second preliminary features is achieved through the second depthwise separable convolutional layer.

[0033] It should be understood that the working process of the second depthwise separable convolutional layer is the same as that of the first depthwise separable convolutional layer.

[0034] Furthermore, the deep feature extraction of the second preliminary features to obtain the second deep semantic features is achieved through the second cross-dimensional interaction self-attention module.

[0035] It should be understood that the working process of the second cross-dimensional interaction self-attention module is the same as that of the first cross-dimensional interaction self-attention module.

[0036] Furthermore, the enhancement processing of the first deep semantic features and the second deep semantic features to obtain the enhanced features is achieved through the structure-aware enhancement module.

[0037] Furthermore, the structure-aware enhancement module includes: a detail aggregation branch and a difference extraction branch in parallel; The detail aggregation branch aggregates the common information of the first deep semantic features and the second deep semantic features to improve the edge integrity and the internal compactness of the changed areas; The detail aggregation branch includes two paths in parallel: a feature concatenation path and a feature summation path, obtaining the concatenated feature and the summation feature ; The feature concatenation path concatenates the first deep semantic features and the second deep semantic features in the channel dimension and uses a multi-layer perceptron to extract richer detail features. The calculation method is as follows:

[0038] Among them, represents a 3×3 convolution, represents a multi-layer perceptron, represents a concatenation operation along the channel dimension.

[0039] For the feature summation path, element-wise summation of the dual-temporal features is performed, and the details are enhanced through processing by a multi-layer perceptron. The calculation method is as follows:

[0040] Where represents the element-wise summation operation.

[0041] Finally, for the concatenated features and the summed features aggregation is performed, and the calculation method is as follows:

[0042] Among them, represents the output of the detailed aggregation branch.

[0043] For the difference extraction branch, aiming to generate complete and detailed feature changes, the differences in the dual-temporal features are refined, and the calculation algorithm is:

[0044] Among them, represents a 3×3 convolution, represents a multi-layer perceptron, represents an absolute value operation to ensure the stability of the difference features, represents the output value of the difference extraction branch; Finally, the outputs of the difference extraction branch and the detailed aggregation branch are fused to generate an enhanced feature representation : .

[0045] Through dual-branch processing, the structure-aware enhancement module effectively integrates the difference information and common details of the change features, not only improving the edge integrity of the change region but also enhancing the internal compactness, laying a foundation for accurate change detection in remote sensing images.

[0046] It should be understood that using the structure-aware enhancement module, dual-branch enhancement processing is performed on the first deep semantic feature and the second deep semantic feature. The difference extraction branch generates complete and detailed change features, and the detailed aggregation branch captures the common detailed features of the dual-temporal features, improving the edge integrity and internal compactness.

[0047] Further, perform feature concatenation on the first deep semantic feature, the second deep semantic feature, and the enhanced feature, perform preliminary feature fusion on the concatenated features, and perform deep fusion on the preliminary fused features, including:

[0048] Among them, represents the concatenation operation along the channel dimension, represents the third depthwise separable convolution operation, represents the third cross-dimensional interactive self-attention module, is the deep feature finally used for change detection.

[0049] Further, the feature concatenation of the first deep semantic feature, the second deep semantic feature, and the enhanced feature is implemented by a concatenation unit. The concatenation unit is used to achieve the serial concatenation of input values.

[0050] Further, the preliminary feature fusion of the concatenated features is implemented by a third depthwise separable convolutional layer. Among them, the working process of the third depthwise separable convolutional layer is the same as that of the first depthwise separable convolutional layer and the second depthwise separable convolutional layer.

[0051] Further, the deep fusion of the preliminary fused features is implemented by a third cross-dimensional interactive self-attention module. Among them, the working process of the third cross-dimensional interactive self-attention module is the same as that of the first cross-dimensional interactive self-attention module and the second cross-dimensional interactive self-attention module.

[0052] It should be understood that feature fusion is to concatenate the features with different semantic information obtained in the early stage, perform preliminary feature fusion through depthwise separable convolution, and then use the cross-dimensional interactive self-attention module to efficiently fuse the deep semantic features of two time phases, providing support for the subsequent fine detection of change regions.

[0053] Feature fusion is one of the key steps of the present invention. The two-time-phase features and extracted by the cross-dimensional interactive self-attention module in the early stage efficiently capture the global context information in the remote sensing image, and the enhanced features

[0054] The specific implementation method is as follows: The three parts of features are concatenated, and depthwise separable convolution is used to balance the local information and global semantic information in the features, preserving the detailed features in the fine-grained change regions, while enhancing the modeling ability for long-range dependencies. The cross-dimensional interactive self-attention module is utilized to perform interactive modeling on the channels and spatial dimensions (height and width) of the features, further capturing the cross-dimensional dependencies in the features.

[0055] The feature fusion process integrates multi-semantic dual-temporal features. Through an efficient computational method and cross-dimensional interaction mechanism, a feature representation that combines both global and detailed information is generated. This process significantly reduces the computational resource requirements through depthwise separable convolution and a self-attention mechanism with linear complexity, reducing the computational complexity of feature fusion while maintaining the fusion effect, ensuring that this method can efficiently process high-resolution remote sensing images.

[0056] Furthermore, the classification of the deep fusion features to obtain the change detection results between the dual-temporal remote sensing images is achieved through a classification head.

[0057] Using the classification head, the final features are mapped to the channels corresponding to the number of change categories to generate the change detection results. This method has the ability to detect high-precision boundaries and capture small target changes. After obtaining the final deep features The final features are then mapped to the channels corresponding to the number of change categories using the classification head to generate the change detection results.

[0058] Furthermore, the working process of the classification head is as follows: Let the number of change categories be , the classification head performs channel mapping on the final deep features through 1×1 convolution to obtain the features . The mapped features are applied to the classification of the change categories, and the change detection results are generated through per-pixel classification operations:

[0059]

[0060]

[0061] Among them, represents 1×1 convolution, The function normalizes the classification result of each pixel into a probability distribution, ensuring that the sum is 1. The function represents taking the category corresponding to the maximum probability. The predicted categories of all pixels are combined into a complete change detection map . In , the value of each pixel represents its corresponding change category.

[0062] The present invention is generally divided into: Initial feature extraction: Taking a pair of multi-temporal remote sensing images as input, and using depthwise separable convolution for initial feature extraction; Cross-dimensional feature extraction: Through the cross-dimensional interactive self-attention module, deep semantic features of multi-temporal remote sensing images are extracted. The cross-dimensional interactive self-attention module effectively reduces the computational complexity through an interactive attention mechanism with linear complexity, while enhancing the global information extraction ability.

[0063] Difference and detail enhancement: Using the structure-aware enhancement module, a two-branch enhancement process is performed on multi-temporal features. The difference extraction branch generates complete and detailed change features, and the detail aggregation branch captures the common detail features of multi-temporal features, improving edge integrity and internal compactness.

[0064] Feature fusion: The features with different semantic information obtained previously are concatenated, and initial feature fusion is performed through depthwise separable convolution. Subsequently, the cross-dimensional interactive self-attention module is used to efficiently fuse the deep semantic features of multi-temporal images, providing support for fine detection of subsequent change regions.

[0065] Change detection: Finally, using the classification head, the final features are mapped to channels with the number of change categories to generate change detection results. This method has the ability of high-precision boundary detection and small target change capture.

[0066] By designing the cross-dimensional interactive self-attention module, the present invention effectively reduces the computational complexity of the traditional self-attention mechanism and enhances the ability to capture global context information. Combined with the structure-aware enhancement module, it realizes the difference refinement and detail aggregation of multi-temporal features, significantly improving the edge integrity of large targets and the accuracy of small target change detection. The present invention not only performs excellently in improving detection accuracy, but also has significant advantages in computational resource consumption and the number of model parameters, effectively balancing the relationship between detection accuracy and computational complexity. This method provides a new and efficient solution for the research and application in the field of remote sensing image change detection, and is of great significance for river evolution, land use monitoring, and disaster assessment, etc.

[0067] Embodiment 2 This embodiment provides a remote sensing image change detection system based on a lightweight structure-aware, including: An acquisition module, which is configured to: acquire a pair of multi-temporal remote sensing images, and the multi-temporal remote sensing images include: a first remote sensing image and a second remote sensing image; A detection module, which is configured to: input the multi-temporal remote sensing images into a trained remote sensing image change detection model to obtain a change detection result of the multi-temporal remote sensing images; Among them, the trained remote sensing image change detection model is used to perform preliminary feature extraction on the first remote sensing image to obtain the first preliminary feature; perform deep feature extraction on the first preliminary feature to obtain the first deep semantic feature; perform preliminary feature extraction on the second remote sensing image to obtain the second preliminary feature; perform deep feature extraction on the second preliminary feature to obtain the second deep semantic feature; perform enhancement processing on the first deep semantic feature and the second deep semantic feature to obtain the enhanced feature; perform feature splicing on the first deep semantic feature, the second deep semantic feature and the enhanced feature, perform preliminary feature fusion on the spliced features, perform deep fusion on the preliminary fusion features, and perform classification on the deep fusion features to obtain the change detection result between the two-temporal remote sensing images.

[0068] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A remote sensing image change detection method based on lightweight structure perception, characterized in that include: Acquire a pair of dual-temporal remote sensing images, the dual-temporal remote sensing images comprising: a first remote sensing image and a second remote sensing image; Inputting the dual-temporal remote sensing image into the trained remote sensing image change detection model to obtain the change detection result of the dual-temporal remote sensing image; Among them, the trained remote sensing image change detection model is used to perform preliminary feature extraction on the first remote sensing image to obtain the first preliminary feature; perform deep feature extraction on the first preliminary feature to obtain the first deep semantic feature; perform preliminary feature extraction on the second remote sensing image to obtain the second preliminary feature; perform deep feature extraction on the second preliminary feature to obtain the second deep semantic feature; perform enhancement processing on the first deep semantic feature and the second deep semantic feature to obtain the enhanced feature; perform feature splicing on the first deep semantic feature, the second deep semantic feature and the enhanced feature, perform preliminary feature fusion on the spliced ​​feature, perform deep fusion on the preliminary fused feature, classify the deep fused feature, and obtain the change detection result between the dual-phase remote sensing images.

2. The remote sensing image change detection method based on lightweight structure perception according to claim 1, wherein The remote sensing image change detection model, the network structure includes: A first depthwise separable convolutional layer and a second depthwise separable convolutional layer; An input end of the first depthwise separable convolutional layer is used to input a first remote sensing image; an output end of the first depthwise separable convolutional layer is connected to an input end of a first cross-dimensional interactive self-attention module; An input end of the second depthwise separable convolutional layer is used to input a second remote sensing image; an output end of the second depthwise separable convolutional layer is connected to an input end of a second cross-dimensional interactive self-attention module; The output end of the first cross-dimensional interactive self-attention module and the output end of the second cross-dimensional interactive self-attention module are both connected to the input end of the structure perception enhancement module; The output end of the first cross-dimensional interactive self-attention module, the output end of the second cross-dimensional interactive self-attention module and the output end of the structure-aware enhancement module are all connected to the input end of the splicing unit; the output end of the splicing unit is also connected to the input end of the third depth-wise separable convolutional layer; the output end of the third depth-wise separable convolutional layer is connected to the input end of the third cross-dimensional interactive self-attention module; the output end of the third cross-dimensional interactive self-attention module is connected to the input end of the classification head; the output end of the classification head outputs the change detection result of the dual-temporal remote sensing image.

3. The method for remote sensing image change detection based on lightweight structure perception according to claim 2, characterized in that The first cross-dimensional interactive self-attention module has three parallel branches, namely channel-channel interactive attention branch, channel-height interactive attention branch and channel-width interactive attention branch, which are used to interact information between different dimensions of the image; Let the initial feature input to the first cross-dimensional interactive self-attention module be , where is the number of channels, and are the height and width of the feature map respectively; For the channel-channel attention branch, the input generates query Q, key K, and value V through the fourth depthwise separable convolutional layer to enhance the local context modeling ability, formulated as: ； ； ； Among them, , and respectively represent depthwise separable convolution operations for generating query Q, key K, and value V; After obtaining the query Q, key K and value V, the channel-channel interaction attention is calculated. The calculation formula is as follows: ； Among them , , Generate channel-channel attention feature maps through matrix multiplication , and through Function and scaling factor Improve stability, and the scaling factor Is learnable and automatically updated following the training process.

4. The remote sensing image change detection method based on lightweight structure perception according to claim 3, characterized in that, The operation flow of the channel-highly interactive attention branch is as follows: ； The operation flow of the channel-width interaction attention branch is: ； Among them, , respectively represent the channel-height attention feature map and the channel-width attention feature map, represents the parallel operations of max pooling and average pooling, and stitches the results of the parallel operations, represents a 1×1 convolution, and respectively represent the tensors obtained by dimension transformation of the input corresponding to the feature maps in the height and width directions respectively, represents function; Obtain the channel-height attention feature map and the channel-width attention feature map After that, generate the final global attention map , and the calculation method is as follows: ； Among them, , , are adjustable weight hyperparameters used to balance the contributions of attention in different dimensions; Obtain the final global attention map After that, apply it to the input features to generate an updated feature representation : ； Among them, are deep features.

5. The method for remote sensing image change detection based on lightweight structure perception according to claim 4, characterized in that, Deep feature extraction is performed on the first preliminary feature to obtain the first deep semantic feature denoted as , deep feature extraction is performed on the first preliminary feature to obtain the first deep semantic feature denoted as ; and The generation processes of are both the same as the deep feature ​ 6. The method for remote sensing image change detection based on lightweight structure perception according to claim 2, wherein The structure-aware enhancement module includes: a parallel detail aggregation branch and a difference extraction branch; The detail aggregation branch aggregates the common information of the first deep-level semantic features and the second deep-level semantic features; The detail aggregation branch includes two parallel paths: the feature splicing path and the feature summation path, obtaining the spliced feature and the summation feature ; Feature concatenation path: Concatenate the first deep semantic feature and the second deep semantic feature in the channel dimension, and use a multi-layer perceptron to extract richer detailed features. The calculation method is as follows: ； Among them, represents a 3×3 convolution, represents a multi-layer perceptron, represents a concatenation operation along the channel dimension; Feature summation path: Perform element-wise summation on the dual-temporal features, and enhance the details through processing by a multi-layer perceptron. The calculation method is as follows: ； Among them, represents an element-wise summation operation; Finally, for the splicing feature and the summation feature perform aggregation, and the calculation method is as follows: ； Among them, represents the output of the detail aggregation branch.

7. The method for remote sensing image change detection based on lightweight structure perception according to claim 6, wherein Difference extraction branch: Aims to generate complete and detailed feature changes by refining the differences between the dual-temporal features. The calculation algorithm is: ； Among them, represents a 3×3 convolution, represents a multi-layer perceptron, represents an absolute value operation to ensure the stability of the differential features, represents the output value of the differential extraction branch; Finally, fuse the outputs of the difference extraction branch and the detail aggregation branch to generate an enhanced feature representation : 。 8. The method for remote sensing image change detection based on lightweight structure perception according to claim 1, characterized in that Perform feature concatenation on the first deep semantic feature, the second deep semantic feature, and the enhanced feature, perform preliminary feature fusion on the concatenated feature, and perform deep fusion on the preliminary fusion feature, including: ； Among them, represents a splicing operation along the channel dimension, represents the third depthwise separable convolution operation, represents the third cross-dimensional interactive self-attention module, is the depth feature finally used for change detection.

9. The method for remote sensing image change detection based on lightweight structure perception according to claim 1, characterized in that, Classify the deep fusion feature to obtain the change detection result between the dual-temporal remote sensing images, which is achieved through a classification head; the working process of the classification head is: Let the number of change categories be , the classification head performs channel mapping on the final depth feature through a 1×1 convolution to obtain the feature , and applies the mapped feature to the classification of change categories, and generates change detection results through per-pixel classification operations: ； ； ； Among them, represents a 1×1 convolution, The function normalizes the classification result of each pixel into a probability distribution, ensuring that the sum is 1, The function represents taking the category corresponding to the maximum probability; combining the predicted categories of all pixels into a complete change detection map , in , the value of each pixel represents its stated change category.

10. A remote sensing image change detection system based on lightweight structure perception, characterized in that, Including: An acquisition module, which is configured to: acquire a pair of dual-temporal remote sensing images, and the dual-temporal remote sensing images include: a first remote sensing image and a second remote sensing image; A detection module, which is configured to: input the dual-temporal remote sensing images into a trained remote sensing image change detection model to obtain the change detection result of the dual-temporal remote sensing images; Among them, the trained remote sensing image change detection model is used to perform preliminary feature extraction on the first remote sensing image to obtain a first preliminary feature; perform deep feature extraction on the first preliminary feature to obtain a first deep semantic feature; perform preliminary feature extraction on the second remote sensing image to obtain a second preliminary feature; perform deep feature extraction on the second preliminary feature to obtain a second deep semantic feature; perform enhancement processing on the first deep semantic feature and the second deep semantic feature to obtain an enhanced feature; perform feature concatenation on the first deep semantic feature, the second deep semantic feature, and the enhanced feature, perform preliminary feature fusion on the concatenated feature, perform deep fusion on the preliminary fusion feature, classify the deep fusion feature, and obtain the change detection result between the dual-temporal remote sensing images.

Images