Multimodal remote sensing image change detection method, system and readable storage medium

Abstract

The present invention relates to a multimodal remote sensing image change detection method, system and readable storage medium. The multimodal remote sensing image is first normalized, superpixel segmentation is performed using simple linear iterative clustering, and the pixels of the superpixels are used as nodes to construct the original graph structure; the original graph structure is input into a graph attention network to extract the graph structure features of the superpixels; the graph structure features are style regularized and style randomized with adjacent superpixels to obtain new style features of the superpixels; the new style features are input into a graph structure decoder to obtain a reconstructed graph structure of the superpixels; the graph attention network is trained using the content consistency between the reconstructed graph structure and the original graph structure to obtain a target graph attention network; the original graph structure obtained by processing the multimodal remote sensing images before and after the change to be detected is input into the target graph attention network to obtain the graph structure features before and after the change, and the change binary map is obtained using the Otsu method. The present invention effectively improves the accuracy of change detection.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing image change detection, and in particular relates to a multimodal remote sensing image change detection method, system and readable storage medium. Background Art

[0002] With the rapid development of remote sensing technology, long-term observational data series can be acquired for remote sensing image change detection, thereby exploring the evolutionary patterns of the Earth. Compared with single-modal remote sensing image change detection, multimodal remote sensing image change detection enables all-weather and all-day observations. By integrating multi-dimensional features such as spectrum, texture, and time series, it provides a richer observational perspective for surface change analysis. This data advantage not only brings new research opportunities but also poses a severe challenge to traditional change detection methods. Unsupervised learning technology is now being applied to remote sensing image target detection tasks. It not only overcomes the reliance on large amounts of annotated data and addresses the problem of label scarcity, but also fully explores the potential correlations and complementarities of multimodal data, improving the accuracy and robustness of change detection. Unsupervised multimodal remote sensing image change detection has practical application value in the following two aspects: 1) Disaster assessment: Multi-time series remote sensing change detection can effectively monitor the evolution of geological disasters. By fusing multimodal data of SAR and optical images, it can break through the interference of clouds and fog to achieve millimeter-level monitoring of landslide displacement; 2) Environmental monitoring: As global climate change intensifies, timely understanding of surface cover evolution is crucial for ecological and environmental protection. For example, by detecting changes in ecologically sensitive areas such as deforestation and wetland degradation, it can provide data support for carbon sink assessment and biodiversity protection.

[0003] Due to the different imaging mechanisms of multimodal images, their features differ, making change detection impossible through direct comparison. Furthermore, multimodal image samples are inherently limited, resulting in a shortage of available labels and high labeling costs. Existing unsupervised change detection methods for multimodal remote sensing images primarily exploit cross-modal image correlations through feature alignment and modality conversion to detect regions of change. However, there are still issues with missed detections or false alarms of changed regions due to global-scale imaging differences between multimodal sensors, as well as discontinuities in changed regions due to local-scale imaging disturbances in multimodal remote sensing images.

[0004] To address the issue of imaging discrepancies between multimodal sensors, existing image regression methods based on structural periodic consistency, for example, convert one image into the domain of another, compare the images, and then directly extract changes. This method spatially aligns the image domains, but the quality of the regression image can affect detection accuracy. Another example is that existing technologies fuse heterogeneous images and generate pre- and post-event phase images that are consistent in both the spectral and spatial domains based on the fusion results, thereby directly measuring the magnitude of change. This method can align the features of optical images and synthetic aperture radar (SAR) images through data fusion strategies, but due to the modal differences between heterogeneous images, data anomalies are prone to occur after image fusion. Furthermore, existing technologies also use affinity matrices to guide images into a common space and then compare features within this space. This method does not consider the modal differences between heterogeneous images, but detection performance depends on the feature extraction effectiveness of the feature learning network or algorithm.

[0005] To address the problem of establishing contextual relationships in multimodal images, many existing methods use graph structures to mine shared features between images, and can also evaluate the degree of change between heterogeneous images through graph similarity metrics. For example, a graph structure method is constructed for multimodal images to effectively represent complex data, and a structural relationship graph convolutional autoencoder is proposed to obtain representative features between multimodal remote sensing images. For another example, a graph structure learning method based on change alignment is used to optimize the learning process of the graph structure. This method successfully improves the performance of change detection in multimodal remote sensing images. However, since the graph structures constructed by different modal images still have heterogeneity and there are domain differences between the graph structures, the problem of discontinuity in the change area caused by local scale imaging perturbations still exists.

[0006] Although the above existing methods have improved the performance of change detection in multimodal remote sensing images, it is still difficult to effectively extract modal differences in unsupervised situations and multimodal graph structures. Summary of the Invention

[0007] Based on the above-mentioned shortcomings and deficiencies in the prior art, one of the objects of the present invention is to at least solve one or more of the above-mentioned problems in the prior art. In other words, one of the objects of the present invention is to provide a multimodal remote sensing image change detection method, system and readable storage medium based on unsupervised learning and graph structure style transfer that meet one or more of the above-mentioned needs.

[0008] In order to achieve the above-mentioned object of the invention, the present invention adopts the following technical solutions:

[0009] A multimodal remote sensing image change detection method comprises the following steps:

[0010] S1. Normalizing the multimodal remote sensing images in the multimodal remote sensing image training set, performing superpixel segmentation on the normalized multimodal remote sensing images using simple linear iterative clustering, and constructing the original graph structure of each superpixel using the pixels of the superpixels as graph structure nodes; wherein the multimodal remote sensing image training set includes the multimodal remote sensing images before the change and the multimodal remote sensing images after the change;

[0011] S2. Input the original graph structure of the superpixel into the graph attention network to extract features and obtain the graph structure features of the superpixel;

[0012] S3. Perform style regularization on the graph structure features of the superpixel and perform style randomization with neighboring superpixels to obtain new style features of the superpixel.

[0013] S4. Input the new style features of the superpixel into the graph structure decoder to reconstruct the graph structure and obtain the reconstructed graph structure of the superpixel. Use the content consistency between the reconstructed graph structure of the superpixel and the original graph structure to train the graph attention network and obtain the converged graph attention network.

[0014] S5. The multimodal remote sensing images before and after the change to be measured are processed by the above step S1, and the obtained original graph structure is input into the converged graph attention network to obtain the graph structure features before and after the change, and the change binary map is obtained using the Otsu method.

[0015] As a preferred solution, step S3 specifically includes the following steps:

[0016] Step S31: performing style regularization on the graph structure features of the current superpixel, and exchanging the style information of the graph structure before and after the change;

[0017] Step S32: Select neighboring superpixels and perform style randomization with the style information of the current superpixel to obtain a new style feature of the current superpixel;

[0018] The process of style randomization is:

[0019] ;

[0020] ;

[0021] in, and are the mean and variance of the graph structure features of the current superpixel, and are the mean and variance of the graph structure features of neighboring superpixels, is a random number, and are the randomly generated mean and variance, i.e., the new style features of the current superpixel.

[0022] As a preferred solution, in step S31, adaptive instance normalization is used to perform style exchange on the graph structure before and after the change.

[0023] As a preferred solution, the constraints for the content consistency between the reconstructed graph structure of superpixels and the original graph structure constructed using the L2 norm loss function are as follows:

[0024] ;

[0025] Where N is the number of superpixels, and Represent the original graph structure of the i-th superpixel of the multimodal remote sensing image before and after the change, and They represent the reconstructed graph structures of the i-th superpixel of the multimodal remote sensing image before and after the style randomization.

[0026] As a preferred solution, in step S4, unsupervised training of the graph attention network is performed using the constraint of style change but content consistency.

[0027] As a preferred solution, the multimodal remote sensing image includes an optical image and a synthetic aperture radar (SAR) image.

[0028] As a preferred solution, step S1 specifically includes the following steps:

[0029] S11, performing minimum-maximum normalization on the optical image to obtain a normalized optical image;

[0030] Perform logarithmic transformation on the SAR image, and then perform minimum-maximum normalization on the transformed image to obtain a normalized SAR image;

[0031] S12. Performing superpixel segmentation on the normalized optical image using simple linear iterative clustering, and applying the segmentation result to the normalized SAR image. Then, the normalized optical image and the normalized SAR image are spliced in the channel dimension to generate a unified superpixel segmentation result.

[0032] S13. Build a graph structure by using each pixel in the superpixel as a graph structure node, which serves as the original graph structure of the superpixel.

[0033] As a preferred solution, in step S5, the structural features of the graph before and after the change are subtracted to obtain a feature difference, and the change binary graph is obtained using the Otsu method.

[0034] The present invention further provides a multimodal remote sensing image change detection system, which applies the multimodal remote sensing image change detection method described in any of the above solutions, and the multimodal remote sensing image change detection system includes:

[0035] A normalization module is used to normalize the multimodal remote sensing images in the multimodal remote sensing image training set; it is also used to normalize the multimodal remote sensing images before and after the change to be measured;

[0036] The superpixel segmentation module is used to perform superpixel segmentation on the normalized multimodal remote sensing image using simple linear iterative clustering;

[0037] A graph structure construction module is used to construct the original graph structure of each superpixel by using the pixels of the superpixel as graph structure nodes;

[0038] The graph attention network module is used to input the original graph structure of the superpixel into the graph attention network to extract features and obtain the graph structure features of the superpixel;

[0039] The graph style transfer module is used to regularize the graph structure features of superpixels and randomize the styles with neighboring superpixels to obtain new style features of superpixels;

[0040] The graph structure decoder module is used to input the new style features of superpixels into the graph structure decoder to reconstruct the graph structure and obtain the reconstructed graph structure of superpixels;

[0041] The training module is used to train the graph attention network using the content consistency between the reconstructed graph structure of superpixels and the original graph structure to obtain a converged graph attention network;

[0042] The detection module is used to input the graph attention network after the original graph structure converges, obtain the graph structure features before and after the change, and use the Otsu method to obtain the change binary map.

[0043] The present invention also provides a readable storage medium, in which instructions are stored. When the instructions are executed on a computer, the computer executes the multimodal remote sensing image change detection method as described in any of the above solutions.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] (1) The present invention selects the style information of neighboring superpixels and the current superpixel to perform style randomization, so that the graph attention network model can learn rich style information, making the model more robust to style information and improving the accuracy of change detection;

[0046] (2) The present invention uses graph structure style transfer to exchange the styles of the multimodal remote sensing image graph structures before and after the change, and uses the graph structure content consistency constraint to train the graph attention network so that the features it can extract contain more content-independent information, effectively improving the accuracy of multimodal remote sensing image change detection;

[0047] (3) The present invention uses neighboring superpixels to randomly generate styles and randomly generates and exchanges style features within a local neighborhood, so that adjacent superpixels can absorb the style information of the neighboring area while maintaining their own characteristics. This can not only effectively alleviate the style mutation problem in the boundary area, but also enrich the feature representation of the superpixel and improve the continuity of the feature space. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 is a flow chart of a multimodal remote sensing image change detection method according to embodiment 1 of the present invention;

[0049] Figure 2 1 is a module architecture diagram of a multimodal remote sensing image change detection system according to embodiment 1 of the present invention;

[0050] Figure 3 This is a diagram of the change detection results of Example 1 of the present invention. DETAILED DESCRIPTION

[0051] To more clearly illustrate the embodiments of the present invention, specific embodiments of the present invention will be described below with reference to the accompanying drawings. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings and other embodiments can be obtained based on these drawings without inventive efforts.

[0052] Example 1:

[0053] The multimodal remote sensing image change detection framework based on unsupervised learning and graph structure style transfer in this embodiment consists of three components: a graph attention network, a graph style transfer framework, and a decoder for reconstructing the graph structure. The specific multimodal remote sensing image change detection process includes: first, using simple linear iterative clustering to perform superpixel segmentation on the multimodal remote sensing image, using each pixel of the superpixel as a node to construct a graph structure, and inputting it into the graph attention network for feature extraction; second, the graph style transfer framework performs style regularization on the above-extracted features, and randomly generates a new style with neighboring superpixels and replaces the original style of the feature, so that the network can learn more styles, so that the extracted features contain more style-independent knowledge; third, the features of the new style are input into the graph structure decoder to reconstruct the graph structure. The reconstructed graph structure should contain the same content as the graph structure input to the graph attention network, and the graph attention network is trained using the consistency of the graph structure content; finally, in the detection stage, the features extracted by the converged graph attention network are used for change detection.

[0054] like Figure 1 As shown, the multimodal remote sensing image change detection method based on unsupervised learning and graph structure style transfer in this embodiment includes the following steps:

[0055] (1) Collect multimodal remote sensing images before and after the change to construct a multimodal remote sensing image training set; the multimodal remote sensing image before the change is referred to as the pre-change image X, and the multimodal remote sensing image after the change is referred to as the post-change image Y;

[0056] The multimodal remote sensing images of this embodiment include optical images and synthetic aperture radar (SAR) images.

[0057] (2) Normalize the multimodal remote sensing images in the multimodal remote sensing image training set, perform superpixel segmentation on the normalized multimodal remote sensing images using simple linear iterative clustering, and use the pixels of the superpixels as graph structure nodes to construct the original graph structure of each superpixel;

[0058] For optical images, the minimum-maximum normalization method is used for normalization to obtain normalized optical images. For SAR images, the SAR images need to be logarithmically transformed to suppress speckle noise, and then the transformed images are subjected to minimum-maximum normalization to obtain normalized SAR images.

[0059] Next, simple linear iterative clustering was used for superpixel segmentation, and a co-segmentation strategy was adopted to superimpose the multimodal image pairs in the channel dimension to obtain a unified segmentation result. Specifically, simple linear iterative clustering was used to perform superpixel segmentation on the normalized optical image, and the segmentation result was applied to the normalized SAR image. The normalized optical image and the normalized SAR image were then spliced in the channel dimension to generate a unified superpixel segmentation result.

[0060] Finally, based on the superpixels obtained by segmentation, each pixel in the superpixel is used as a graph structure node to construct a graph structure, and the structural information of the multimodal remote sensing image is obtained as the original graph structure of the superpixel.

[0061] (3) Build a graph attention network GAT, and input the original graph structure of the superpixel into the graph attention network to extract features, and obtain the graph structure features of the superpixel. That is, in this embodiment, the image feature F before the change can be obtained based on the image before the change X and the image after the change Y. X and the image feature F after the change Y ,Then the following step (4) of graph style knowledge transfer is carried out;

[0062] Among them, the graph attention network can efficiently extract the features of the graph structure. In addition, in order to improve the accuracy of establishing the connection relationship between graph vertices and improve the ability of the graph neural network to model the image structure features, the attention mechanism GAM is introduced. The specific details of the attention mechanism can be referred to the existing technology and will not be elaborated here.

[0063] (4) Perform style regularization on the graph structure features of the superpixel, then select neighboring superpixels and the current superpixel for style randomization, and use the consistency of graph structure content to train the graph attention network to improve the network model's ability to extract style-independent features; the style regularization process can refer to the existing technology and will not be described here;

[0064] After extracting the graph structure features, this embodiment applies style regularization to the features and exchanges the style information of the graph structure before and after the change, using the constraint of style change but content consistency to perform unsupervised training.

[0065] Although constructing a graph structure for multimodal remote sensing imagery and detecting changes in structural information can reduce modal differences, for images with high change detection accuracy, the graph structures constructed from remote sensing images of different modalities still face modal differences, resulting in false alarms for objects with significant imaging differences between different modalities. Based on this, this embodiment introduces graph structure style transfer, which exchanges the style of the graph structure before and after the change. The graph attention network is trained using content consistency constraints, so that the features it can extract contain more content-independent information.

[0066] Specifically, adaptive instance normalization is used to exchange the styles of the graph structures before and after the change;

[0067] Then, neighboring superpixels are selected and style randomized with the style information of the current superpixel. With this design, the network model can learn rich style information, making the model more robust to style information.

[0068] The complexity of objects in remote sensing images also poses challenges to change detection. The sudden change of spectral characteristics at the boundaries of objects often produces obvious dividing lines when superpixel segmentation is used to process complex scenes such as the water-land boundary, resulting in significant differences in style features between adjacent superpixels. This may introduce false alarms in the boundary area, thereby affecting the accuracy of subsequent change detection. Based on this, this embodiment introduces random generation of adjacent superpixels and performs random generation and exchange of style features within the local neighborhood, so that adjacent superpixels can absorb the style information of the adjacent area while maintaining their own characteristics. This strategy can not only effectively alleviate the style mutation problem in the boundary area, but also enrich the feature representation of superpixels and improve the continuity of the feature space.

[0069] In this embodiment, the image feature F before the change X For example, the process of style randomization is described in detail as follows:

[0070] ;

[0071] ;

[0072] in, and are the mean and variance of the graph structure features of the current superpixel, and are the mean and variance of the graph structure features of neighboring superpixels, is a random number, and are the randomly generated mean and variance, i.e., the new style features of the current superpixel;

[0073] The feature content after style randomization should be consistent with that before transformation, so the graph structure decoder before and after transformation are used to reconstruct the graph structure respectively. The reconstructed graph structure should maintain content consistency with the graph structure of the input graph attention network. The constraint condition for the content consistency between the reconstructed graph structure and the original graph structure of the superpixel constructed using the L2 norm loss function is:

[0074] ;

[0075] Where N is the number of superpixels, and Represent the original graph structure of the i-th superpixel of the multimodal remote sensing image before and after the change, and Respectively represent the reconstructed graph structures of the i-th superpixel of the multimodal remote sensing image before and after the style randomization;

[0076] The graph attention network is trained through the above steps (1) to (4). After the training converges, a converged graph attention network is obtained. In this embodiment, the graph attention network is trained unsupervisedly using the constraint of style change but content consistency.

[0077] (5) Based on the graph attention network after training convergence, change detection is performed on the test samples, and the binary change detection results are obtained using the Otsu method;

[0078] Specifically, the multimodal remote sensing images before and after the change to be measured are processed by the above step (2), and the original graph structures before and after the change are input into the converged graph attention network to obtain the graph structure features before and after the change. The graph structure features before and after the change are subtracted to obtain the feature difference, and the change binary map is obtained using the Otsu method.

[0079] Based on the above multimodal remote sensing image change detection method, such as Figure 2 As shown, the multimodal remote sensing image change detection system provided in this embodiment includes the following functional modules: normalization module, superpixel segmentation module, graph structure construction module, graph attention network module, graph style transfer module, graph structure decoder module, training module and detection module;

[0080] The normalization module of this embodiment is used to normalize the multimodal remote sensing images of the multimodal remote sensing image training set, and is also used to normalize the multimodal remote sensing images before and after the change to be measured;

[0081] The superpixel segmentation module of this embodiment is used to perform superpixel segmentation on the normalized multimodal remote sensing image using simple linear iterative clustering;

[0082] The graph structure construction module of this embodiment is used to construct the original graph structure of each superpixel using the pixels of the superpixel as graph structure nodes;

[0083] The graph attention network module of this embodiment is used to input the original graph structure of the superpixel into the graph attention network to extract features and obtain the graph structure features of the superpixel;

[0084] The graph style transfer module of this embodiment is used to perform style regularization on the graph structure features of the superpixel and perform style randomization with neighboring superpixels to obtain new style features of the superpixel;

[0085] The graph structure decoder module of this embodiment is used to input the new style features of the superpixel into the graph structure decoder, reconstruct the graph structure, and obtain the reconstructed graph structure of the superpixel;

[0086] The training module of this embodiment is used to train the graph attention network using the content consistency between the reconstructed graph structure of superpixels and the original graph structure to obtain a converged graph attention network;

[0087] The detection module of this embodiment is used to input the original graph structure into the converged graph attention network to obtain the graph structure features before and after the change, and use the Otsu method to obtain the change binary map;

[0088] The specific processing process of the above functional modules can refer to the specific description in the above multimodal remote sensing image change detection method, which will not be repeated here.

[0089] The readable storage medium of this embodiment stores instructions. When the instructions are executed on a computer, the computer executes the above-mentioned multimodal remote sensing image change detection method to realize intelligent multimodal remote sensing image change detection.

[0090] The multimodal remote sensing image change detection method of this embodiment is used in a specific practical application, and the specific process is as follows:

[0091] (1) We use the Shuguang dataset, a public multimodal change detection dataset. The Shuguang training dataset consists of SAR images and optical images. The preprocessed SAR and optical images are 921 × 593 pixels in size.

[0092] (2) Build a graph attention network and input the graph structure of multimodal images into the network to extract features;

[0093] Build a graph attention network to extract features. Use the existing commonly used three-layer graph convolution layer to build a graph convolutional neural network with a network structure of GCN-GAT-GCN to extract deep information from the graph structure and obtain graph structure features.

[0094] (3) Perform style randomization on the extracted features to train the network model, and use the network model for change detection, such as Figure 3 As shown, the white part is the changed part, and the black part is the unchanged part; the green area is the changed but not detected, which is the missed detection part; the red area detects the unchanged part as a change, which is the false alarm part.

[0095] The detection method of the present invention is compared with the existing commonly used detection methods GIR_MRF, SCACS, IRG_McS, and SRGCAE. As shown in Table 1, the detection method of the present invention is superior to the existing detection methods in the average intersection-over-union (mIoU) indicator, especially compared with the existing SCACS algorithm, the average intersection-over-union (mIoU) indicator is 17.3% higher; compared with SRGCAE, which is also based on a deep learning algorithm, the average intersection-over-union indicator is 4.2% higher. In addition, the present invention shows the best performance in foreground intersection-over-union (IoU_1) and background intersection-over-union (IoU_2), which is 32.7% higher than the SCACS algorithm in IoU_1 and 8.3% higher than the second best result SRGCAE algorithm. Although the present invention is slightly lower than the GIR_MRF algorithm and IRG_McS algorithm in F1 score and Kappa coefficient Kappa, it is 18.2% higher than the GIR_MRF algorithm and 14.6% higher than the IRG_MCS algorithm in IoU_1.

[0096] Table 1 Comparative analysis indicators of different detection methods

[0097] .

[0098] The above description is only a detailed description of the preferred embodiments and principles of the present invention. For ordinary technicians in this field, based on the ideas provided by the present invention, there may be changes in the specific implementation methods, and these changes should also be considered as the scope of protection of the present invention.

Claims

1. A multimodal remote sensing image change detection method, characterized in that: The following steps are involved: S1. Normalizing the multimodal remote sensing images in the multimodal remote sensing image training set, performing superpixel segmentation on the normalized multimodal remote sensing images using simple linear iterative clustering, and constructing the original graph structure of each superpixel using the pixels of the superpixels as graph structure nodes; wherein the multimodal remote sensing image training set includes the multimodal remote sensing images before the change and the multimodal remote sensing images after the change; S2. Input the original graph structure of the superpixel into the graph attention network to extract features and obtain the graph structure features of the superpixel; S3. Perform style regularization on the graph structure features of the superpixel and perform style randomization with neighboring superpixels to obtain new style features of the superpixel. S4. Input the new style features of the superpixel into the graph structure decoder to reconstruct the graph structure and obtain the reconstructed graph structure of the superpixel. Use the content consistency between the reconstructed graph structure of the superpixel and the original graph structure to train the graph attention network and obtain the converged graph attention network. S5. The multimodal remote sensing images before and after the change to be measured are processed by the above step S1, and the obtained original graph structure is input into the converged graph attention network to obtain the graph structure features before and after the change, and the change binary map is obtained using the Otsu method.

2. The multimodal remote sensing image change detection method according to claim 1, characterized in that: The step S3 specifically includes the following steps: Step S31: performing style regularization on the graph structure features of the current superpixel, and exchanging the style information of the graph structure before and after the change; Step S32: Select neighboring superpixels and perform style randomization with the style information of the current superpixel to obtain a new style feature of the current superpixel; The process of style randomization is: ； ； in, and are the mean and variance of the graph structure features of the current superpixel, and are the mean and variance of the graph structure features of neighboring superpixels, is a random number, and are the randomly generated mean and variance, i.e., the new style features of the current superpixel.

3. The multimodal remote sensing image change detection method according to claim 2, characterized in that: In step S31, adaptive instance normalization is used to perform style exchange on the graph structure before and after the change.

4. The multimodal remote sensing image change detection method according to any one of claims 1 to 3, characterized in that: The constraints for the content consistency between the reconstructed graph structure of superpixels and the original graph structure constructed using the L2 norm loss function are: ； Where N is the number of superpixels, and Represent the original graph structure of the i-th superpixel of the multimodal remote sensing image before and after the change, and They represent the reconstructed graph structures of the i-th superpixel of the multimodal remote sensing image before and after the style randomization.

5. The multimodal remote sensing image change detection method according to claim 4, characterized in that: In step S4, unsupervised training of the graph attention network is performed using the constraint of style change but content consistency.

6. The multimodal remote sensing image change detection method according to any one of claims 1 to 3, characterized in that: The multimodal remote sensing images include optical images and synthetic aperture radar (SAR) images.

7. The multimodal remote sensing image change detection method according to claim 6, characterized in that: The step S1 specifically includes the following steps: S11, performing minimum-maximum normalization on the optical image to obtain a normalized optical image; Perform logarithmic transformation on the SAR image, and then perform minimum-maximum normalization on the transformed image to obtain a normalized SAR image; S12. Performing superpixel segmentation on the normalized optical image using simple linear iterative clustering, and applying the segmentation result to the normalized SAR image. Then, the normalized optical image and the normalized SAR image are spliced in the channel dimension to generate a unified superpixel segmentation result. S13. Build a graph structure by using each pixel in the superpixel as a graph structure node, which serves as the original graph structure of the superpixel.

8. The multimodal remote sensing image change detection method according to any one of claims 1 to 3, characterized in that: In step S5, the structural features of the graph before and after the change are subtracted to obtain a feature difference, and the change binary graph is obtained using the Otsu method.

9. A multimodal remote sensing image change detection system, applying the multimodal remote sensing image change detection method according to any one of claims 1 to 8, characterized in that: The multimodal remote sensing image change detection system includes: A normalization module is used to normalize the multimodal remote sensing images in the multimodal remote sensing image training set; it is also used to normalize the multimodal remote sensing images before and after the change to be measured; The superpixel segmentation module is used to perform superpixel segmentation on the normalized multimodal remote sensing image using simple linear iterative clustering; A graph structure construction module is used to construct the original graph structure of each superpixel by using the pixels of the superpixel as graph structure nodes; The graph attention network module is used to input the original graph structure of the superpixel into the graph attention network to extract features and obtain the graph structure features of the superpixel; The graph style transfer module is used to regularize the graph structure features of superpixels and randomize the styles with neighboring superpixels to obtain new style features of superpixels; The graph structure decoder module is used to input the new style features of superpixels into the graph structure decoder to reconstruct the graph structure and obtain the reconstructed graph structure of superpixels; The training module is used to train the graph attention network using the content consistency between the reconstructed graph structure of superpixels and the original graph structure to obtain a converged graph attention network; The detection module is used to input the graph attention network after the original graph structure converges, obtain the graph structure features before and after the change, and use the Otsu method to obtain the change binary map.

10. A readable storage medium having instructions stored therein, characterized in that: When the instructions are executed on a computer, the computer is caused to execute the multimodal remote sensing image change detection method according to any one of claims 1 to 8.

Images