A method, system and medium for detecting changes in multi-source heterogeneous remote sensing images

Abstract

This invention relates to a method, system, and medium for detecting changes in multi-source heterogeneous remote sensing images. The change detection method includes: normalizing the multi-source heterogeneous remote sensing images before and after the change, and performing superpixel segmentation on the normalized multi-source heterogeneous remote sensing images to obtain superpixel blocks; inputting each superpixel block into a pre-constructed evolutionary graph network to extract latent evolutionary graph representations; based on the latent evolutionary graph representations of different modalities and after statistical distribution exchange, inputting them into corresponding modality-specific decoders to reconstruct the superpixel blocks; training the evolutionary graph network using the content consistency of the superpixel blocks before and after reconstruction to obtain a converged evolutionary graph network; performing change detection on the multi-source heterogeneous remote sensing images before and after the change based on the converged evolutionary graph network, and obtaining binary change detection results using the Otsu method. This invention achieves change detection of multi-source heterogeneous remote sensing images and further improves detection accuracy.

Description

Technical Field

[0001] This invention belongs to the field of multi-source heterogeneous remote sensing image change detection, specifically relating to a method, system, and medium for multi-source heterogeneous remote sensing image change detection. Background Technology

[0002] Compared to single-modal change detection methods, multi-source heterogeneous remote sensing imagery (i.e., multimodal remote sensing imagery) enables continuous observation regardless of time and weather conditions. By fusing spectral information, texture features, and time-series characteristics, it can depict surface changes in multiple dimensions. However, due to the fundamental differences in imaging mechanisms among different modalities of remote sensing imagery, their characterization features exhibit significant inconsistencies, making change detection difficult through simple pixel-by-pixel or direct comparison methods. Furthermore, multimodal data itself is challenging to acquire, with a limited number of usable samples, and the high cost of manual annotation further exacerbates the scarcity of high-quality labeled data.

[0003] The goal of unsupervised multimodal change detection is to effectively identify surface changes by constructing a unified multimodal image representation space, given the significant differences in imaging mechanisms. Recent research often introduces graph structures to characterize the contextual relationships within multimodal images and measures change through the similarity or structural differences between graphs. These methods are typically based on the assumption that the content structure of multimodal images should remain consistent in unchanged regions. Existing techniques further improve the performance of cross-modal change detection by iteratively comparing graph structure features to identify changed regions; this method constructs graph structures for different modalities, ensuring domain differences in their feature spaces. With the development of graph neural networks, deep learning-based graph change detection methods typically employ graph convolutional networks (GCNs) or graph autoencoders to model graphs constructed from image patches or superpixels, thereby learning cross-modal structural consistency representations. For example, existing techniques use structural relation graph convolutional autoencoders to acquire representative features among multi-source heterogeneous remote sensing images, but their graph structures are usually constructed based on predefined rules and remain unchanged during training. Another example is the combination of scale-target features with GCNs to achieve change detection through multi-scale graph modeling and information propagation. Furthermore, existing techniques also optimize the graph structure learning process based on change alignment graph structure learning methods, which have successfully improved the performance of change detection in multi-source heterogeneous remote sensing images. However, existing methods as a whole still follow a static modeling paradigm, decoupling cross-modal alignment or graph structure modeling from the change detection process, making it difficult to achieve unified end-to-end optimization.

[0004] Although the above methods have made some progress in multimodal change detection tasks, they are still difficult to alleviate the problem of distribution differences between multimodal graph structures due to static structure modeling and phased optimization strategies. Summary of the Invention

[0005] Based on the aforementioned shortcomings and deficiencies in the prior art, one of the objectives of this invention is to at least solve one or more of the aforementioned problems in the prior art. In other words, one of the objectives of this invention is to provide a method, system, and medium for detecting changes in multi-source heterogeneous remote sensing images that meets one or more of the aforementioned requirements.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: A method for detecting changes in multi-source heterogeneous remote sensing images includes the following steps: S1. Normalize the multi-source heterogeneous remote sensing images before and after the change, and perform superpixel segmentation on the normalized multi-source heterogeneous remote sensing images to obtain superpixel blocks. S2. Input each superpixel block into the pre-constructed evolution graph network to extract the latent evolution graph representation. Based on the latent evolution graph representations of different modalities and after statistical distribution exchange, input them into the corresponding modality-specific decoder to reconstruct the superpixel blocks. Use the content consistency of the superpixel blocks before and after reconstruction to train the evolution graph network and obtain the trained and converged evolution graph network. S3. Based on the converged evolution graph network after training, change detection is performed on multi-source heterogeneous remote sensing images before and after the change, and the Otsu method is used to obtain the binary change detection results.

[0007] As a preferred embodiment, the multi-source heterogeneous remote sensing imagery includes optical images of the first mode and synthetic aperture radar (SAR) images of the second mode.

[0008] As a preferred embodiment, step S2, the extraction of the latent evolution map representation includes the following steps: S21, superpixel blocks Input Embedded Projection Layer Projection is performed to obtain the mapped features. ; S22, Based on mapping features The relationship features between pixels in a superpixel block are extracted through three learnable linear mappings: ; in, Indicates the first Based on Learnable mapping operators implemented through convolution. Indicates after the first A learnable mapping operator for dimensionality reduction of relational features It can be 1, 2, or 3; Pixels in a superpixel block With pixels The strength of the association between them is measured by their feature inner product, defined as: ; in, The relational features of pixel i after dimensionality reduction by the third learnable mapping operator are given. The relational features of pixel j after dimensionality reduction by the second learnable mapping operator; Normalizing the association strength yields the relationship matrix: ; Where N is the number of pixels in the superpixel block; S23. Based on the relation matrix, perform weighted aggregation of pixel information to obtain the graph structure features of pixel i: ; in, The relational features of pixel j after dimensionality reduction by the first learnable mapping operator; S24. Graph structure features after aggregating all pixels. A linear transformation is performed to obtain the enhanced graphical representation: ; in, These are learnable parameters; S25. Perform a nonlinear transformation on the enhanced graph representation and add it to the cross-temporal feature representation to obtain the latent evolution graph representation: ; in, This is the graphical representation after nonlinear transformation. For activation functions; The cross-temporal feature representation is extracted by the base feature extractor as the input for mapping features.

[0009] As a preferred embodiment, step S2, the process of reconstructing the superpixel block, includes the following steps: S26. Input the latent evolution graph representation and the corresponding nonlinearly transformed graph representation into a Transformer-based modality-specific decoder for superpixel block reconstruction, resulting in: ; in, and These are the superpixel blocks of the reconstructed first and second modes, respectively. and These are decoders corresponding to the first mode and the second mode, respectively. and These are the graphical representations of the nonlinear transformations corresponding to the first and second modes, respectively. and These are the potential evolution diagrams representing the first and second modes, respectively; S27. Constructing intramodal reconstruction constraints based on content consistency: ; in, and These are superpixel blocks corresponding to the first mode and the second mode, respectively.

[0010] As a preferred embodiment, step S2, the process of reconstructing the superpixel block, further includes the following steps: S28. Exchange the statistical distributions of the graph representations of different modal latent evolution graphs after nonlinear transformation: ; in, and These are the graph structure features corresponding to the exchange of statistical distributions between the first and second modes, respectively. This represents a feature transformation operation based on distribution conditions. and The graphs after nonlinear transformation are shown below. and The statistical distribution information includes the mean and variance. The graph structure features after statistical distribution swapping are input into a modality-specific decoder for superpixel block reconstruction along with the corresponding latent evolution graph representation. ; S28. Constructing cross-modal reconstruction constraints based on content consistency: .

[0011] As a preferred embodiment, in step S2, an optimization objective function is constructed based on intra-modal reconstruction constraints and cross-modal reconstruction constraints: ; Evolutionary graph network is trained using an optimized objective function.

[0012] As a preferred embodiment, step S1 specifically includes: The optical images were preprocessed using the minimum-maximum normalization method. For SAR images, a logarithmic transformation is first performed, followed by preprocessing using the min-max normalization method; Superpixel segmentation is performed on the preprocessed image.

[0013] As a preferred approach, simple linear iterative clustering is used for superpixel segmentation.

[0014] This invention also provides a multi-source heterogeneous remote sensing image change detection system, employing the multi-source heterogeneous remote sensing image change detection method as described in any of the above embodiments. The multi-source heterogeneous remote sensing image change detection system includes: The preprocessing module is used to normalize the multi-source heterogeneous remote sensing images before and after the changes, and to perform superpixel segmentation on the normalized multi-source heterogeneous remote sensing images. The network training module is used to input superpixel blocks of multi-source heterogeneous remote sensing images one by one into a pre-constructed evolution graph network to extract latent evolution graph representations. Based on the latent evolution graph representations of different modalities and after statistical distribution exchange, they are input into the corresponding modality-specific decoders to reconstruct the superpixel blocks. The evolution graph network is trained using the content consistency of the superpixel blocks before and after reconstruction to obtain the converged evolution graph network. The change detection module is used to detect changes in multi-source heterogeneous remote sensing images before and after changes based on the trained converged evolution graph network, and to obtain binary change detection results using the Otsu method.

[0015] The present invention also provides a medium storing instructions that, when executed on a computer, cause the computer to perform the multi-source heterogeneous remote sensing image change detection method as described in any of the preceding embodiments.

[0016] Compared with the prior art, the beneficial effects of this invention are: This invention performs superpixel segmentation on multi-source heterogeneous remote sensing images, inputting superpixel blocks as basic units into an evolutionary graph network to extract latent evolutionary graph representations. Statistical distribution exchange is performed on the graph representations of different modalities, enabling the network to learn modality-independent structural representations, thereby improving feature consistency and robustness. Then, the statistically exchanged graph structural features and latent evolutionary graph representations are input into a modality-specific decoder for superpixel block reconstruction, ensuring consistency between the reconstruction result and the original input, and training the evolutionary graph network. Finally, in the detection phase, the features extracted by the trained evolutionary graph network are used to calculate differences, achieving change detection in multi-source heterogeneous remote sensing images and further improving detection accuracy. Attached Figure Description

[0017] Figure 1 This is a flowchart of the multi-source heterogeneous remote sensing image change detection method of Embodiment 1 of the present invention; Figure 2 This is a module architecture diagram of the multi-source heterogeneous remote sensing image change detection system of Embodiment 1 of the present invention; Figure 3 This is a graph showing the change detection results of Embodiment 1 of the present invention; Detailed Implementation

[0018] To more clearly illustrate the embodiments of the present invention, specific implementation methods will be described below with reference to the accompanying drawings. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings and other implementation methods can be obtained based on these drawings without any creative effort.

[0019] The unsupervised learning and end-to-end evolutionary graph generalization framework for multi-source heterogeneous remote sensing image change detection in this invention consists of two components: an evolutionary graph network framework for extracting latent evolutionary graph representations and a cross-modal graph generalization framework containing a modality-specific decoder for generalizing evolutionary graph representations. First, simple linear iterative clustering is used to perform superpixel segmentation on the multi-source heterogeneous remote sensing images, with superpixel blocks as the basic unit, which are then input into the evolutionary graph network to extract latent evolutionary graph representations. Second, the latent evolutionary graph representations are input into the modality-specific decoder for reconstruction, ensuring the consistency of the internal structural relationships of each modality. Simultaneously, statistical distribution exchange... The strategy involves statistical distribution exchange of the nonlinearly transformed graph representations corresponding to different modalities, allowing the graph representation of one modality to be reconstructed under the statistical distribution of another modality. This gradually eliminates the distribution differences between modalities, making the graph representations gradually approach modality-independent representations. Then, the graph structure features after distribution exchange and the latent evolutionary graph representations are input to the modality-specific decoder for superpixel block reconstruction. The reconstructed superpixel blocks should contain the same content as the superpixel blocks of the input network, so as to train the evolutionary graph network until convergence. Finally, in the detection phase, the features extracted by the trained and converged evolutionary graph network are used for change detection.

[0020] Example 1: like Figure 1 As shown, the multi-source heterogeneous remote sensing image change detection method based on unsupervised learning and end-to-end evolution graph generalization in this embodiment includes the following steps: (1) Normalize the multi-source heterogeneous remote sensing images before and after the change, and use simple linear iterative clustering to perform superpixel segmentation on the multi-source heterogeneous remote sensing images. Use the segmented superpixel blocks as basic units to input into the evolution graph network for operation. Specifically, the multi-source heterogeneous remote sensing images include optical images of the first mode and synthetic aperture radar (SAR) images of the second mode; For optical images, the minimum-maximum normalization method is used for normalization processing; For SAR images, the first step is to perform a logarithmic transformation on the images to suppress speckle noise, and then perform normalization processing on the transformed images using the min-max normalization method. Then, simple linear iterative clustering is used to perform superpixel segmentation on the normalized optical image and SAR image respectively. A co-segmentation strategy is adopted to superimpose the multimodal image pairs in the channel dimension to obtain a unified segmentation result, namely superpixel blocks.

[0021] (2) Input each superpixel block into a pre-constructed evolutionary graph network to extract potential evolutionary graph representations; An evolutionary graph network is constructed to extract latent evolutionary graph representations. The evolutionary graph network is used to construct latent evolutionary graph representations by embedding structural information into basic features. Through learnable embedded projection layers The input is mapped to a unified latent feature space to achieve feature alignment and unified representation across different modalities; for input superpixel blocks ,in, Indicates the number of pixels. The feature dimension is represented, and the mapped feature is obtained after projection. Using a basic feature extractor The mapped features are encoded to extract deep feature representations across time phases. A lightweight Transformer encoding structure is used to model the input features and obtain cross-temporal feature representations: ; To achieve end-to-end generation of latent evolution graph representations, this embodiment directly learns the relational dependencies between pixels in superpixel blocks within the feature space. Specifically, the input features are first projected onto the latent relation embedding space through multiple learnable linear mappings, and relational features between pixels are extracted from different perspectives, resulting in: ; in, Indicates the first Based on Learnable mapping operators implemented through convolution. Indicates after the first A learnable mapping operator for dimensionality reduction of relational features The value is 1, 2, or 3; furthermore, this embodiment does not rely on a pre-constructed adjacency matrix, but implicitly models the relationships between pixels in the latent feature space through a learnable affinity function; pixels With pixels The strength of the correlation between them is determined by the anisotropic kernel operators. The measurement is defined as follows: ; in, The relational features of pixel i after dimensionality reduction by the third learnable mapping operator are given. The relational features of pixel j after dimensionality reduction by the second learnable mapping operator; Based on this, the above relationships are normalized to obtain the relationship matrix: ; Subsequently, based on the constructed relation matrix, the pixel information is weighted and aggregated to achieve graph information propagation, resulting in the graph structure features of pixel i: ; This process adaptively fuses global pixel information based on the strength of the relationships between pixels, thereby obtaining graph structure features that contain contextual information; Furthermore, the graph structure features after aggregating all pixels... A linear transformation is performed to obtain the enhanced graphical representation: ; in, These are learnable parameters; Finally, the enhanced graph representation is subjected to a nonlinear transformation and added to the cross-temporal feature representation to obtain the latent evolution graph representation: ; in, This is the graphical representation after nonlinear transformation. For activation functions; This embodiment This latent evolution graph representation, which integrates pixel relationship information and original feature information, can effectively characterize the structural dependencies between nodes and is updated as the network trains.

[0022] (3) The extracted latent evolution graphs of different modalities are represented and statistically distributed and then input into the corresponding modality-specific decoders to reconstruct the superpixel blocks. The content consistency of the superpixel blocks before and after reconstruction is used to train the evolution graph network and obtain the converged evolution graph network. Specifically, after extracting the latent evolution graph representation, in order to improve the modality independence of the latent evolution graph representation, this embodiment designs cross-modal graph generalization, which simultaneously introduces intramodal content consistency constraints and cross-modal content consistency constraints to achieve unified modeling of graph structure relationships; To maintain modality specificity while achieving cross-modality compatibility, modality-specific decoders are first constructed for different modalities to reconstruct superpixel blocks. The latent evolution graph representation is then fused with the corresponding nonlinearly transformed graph representation and input into the Transformer-based modality-specific decoder to achieve superpixel reconstruction. This process enables the network model to recover the structural information of the original image from the graph structure representation. ; in, and These are the superpixel blocks of the reconstructed first and second modes, respectively. and These are decoders corresponding to the first mode and the second mode, respectively. and These are the graphical representations of the nonlinear transformations corresponding to the first and second modes, respectively. and These are the potential evolution diagrams representing the first and second modes, respectively; To ensure that the learned graph representations fully retain the intrinsic structural information of each modality, this embodiment introduces intra-modal reconstruction constraints. By constraining the difference between the reconstruction results and the original input, the network model maintains the consistency of the internal structural relationships of each modality during the optimization process. The specific intra-modal reconstruction constraints are as follows: ; in, and These are superpixel blocks corresponding to the first mode and the second mode, respectively; Although the aforementioned reconstruction constraints effectively preserve modality-specific structural information, the learned graph features may still exhibit modality-related distribution differences. Therefore, this embodiment further proposes a cross-modal distribution generalization mechanism. By exchanging the statistical distributions of different modalities at the feature level, the graph structure features of one modality are reconstructed under the statistical distribution conditions of another modality, thereby guiding the network model to gradually learn modality-independent graph structure representations. Specifically, the statistical distributions of the nonlinearly transformed graph representations corresponding to the latent evolution graph representations of different modalities are exchanged to obtain the graph structure features after statistical distribution exchange: ; in, and These are the graph structure features corresponding to the exchange of statistical distributions between the first and second modes, respectively. and The graphs after nonlinear transformation are shown below. and The statistical distribution information includes the mean and variance. This represents a feature transformation operation based on distribution conditions, used to inject the statistical properties of one mode into the features of another mode, thereby achieving cross-modal distribution exchange; Subsequently, the graph structure features after distribution exchange and the corresponding latent evolution graph representation are input together into the corresponding modality-specific decoder for superpixel block reconstruction: ; To ensure that the reconstructed features retain original content information while remaining invariant to changes in modality distribution, this embodiment further introduces content consistency constraints to constrain the content consistency between the reconstructed features and the original features: ; Finally, by combining all the above constraints, an overall optimization objective function is constructed. This is used to uniformly train the network model until the evolution graph network converges; the overall optimization objective function is: ; This embodiment utilizes content consistency training to improve the model's ability to extract representations of modality-independent graphs.

[0023] (4) Based on the converged evolution graph network, change detection is performed on the multi-source heterogeneous remote sensing images before and after the change, and the Otsu method is used to obtain the binary change detection results.

[0024] In the change detection stage, the features extracted by the evolution graph network are subtracted to obtain the feature difference value, and the Otsu method is used to obtain the change binary map.

[0025] Based on the above-mentioned multi-source heterogeneous remote sensing image change detection method, such as Figure 2 As shown, this embodiment provides a corresponding multi-source heterogeneous remote sensing image change detection system, including the following functional modules: preprocessing module, network training module, and change detection module; The aforementioned preprocessing module is used to normalize the multi-source heterogeneous remote sensing images before and after the changes, and to perform superpixel segmentation on the normalized multi-source heterogeneous remote sensing images. The aforementioned network training module is used to input superpixel blocks of multi-source heterogeneous remote sensing images one by one into a pre-constructed evolutionary graph network to extract latent evolutionary graph representations. Based on the latent evolutionary graph representations of different modalities and after statistical distribution exchange, they are input into the corresponding modality-specific decoders to reconstruct the superpixel blocks. The evolutionary graph network is trained using the content consistency of the superpixel blocks before and after reconstruction, and the converged evolutionary graph network is obtained. The aforementioned change detection module is used to perform change detection on multi-source heterogeneous remote sensing images before and after changes based on the trained converged evolution graph network, and to obtain binary change detection results using the Otsu method. The specific processing procedures of the above functional modules can be found in the detailed description of the multi-source heterogeneous remote sensing image change detection method, and will not be repeated here.

[0026] In addition, the readable storage medium of this embodiment stores instructions that, when executed on a computer, cause the computer to perform the above-described method for detecting changes in multi-source heterogeneous remote sensing images, thereby achieving intelligent detection of changes in multi-source heterogeneous remote sensing images.

[0027] Example 2: The difference between the multi-source heterogeneous remote sensing image change detection method in this embodiment and that in embodiment 1 is that: statistical distribution exchange is not performed and cross-modal reconstruction constraints are not performed. The optimization objective function in this embodiment is: ; The remaining steps are the same as in Example 1; Accordingly, the functional modules of the multi-source heterogeneous remote sensing image change detection system are adaptively adjusted; The readable storage medium of this embodiment stores instructions. When the instructions are executed on a computer, the computer performs the above-described method for detecting changes in multi-source heterogeneous remote sensing images, thereby achieving intelligent detection of changes in multi-source heterogeneous remote sensing images.

[0028] Comparative Example 1: The difference between the multi-source heterogeneous remote sensing image change detection method in this comparative example and Example 1 is that it does not perform intra-modal reconstruction constraints and cross-modal reconstruction constraints. The remaining steps are the same as in Example 1.

[0029] The effectiveness of Example 1, Example 2, Comparative Example 1, and existing multi-source heterogeneous remote sensing image change detection methods AGSCC, IST_CRF, SCACS, and AEKAN was compared and verified: Using the publicly available multimodal change detection dataset Gloucester, the Gloucester training dataset consists of SAR and optical aerial images. The preprocessed SAR and optical images are 4135×2325 pixels in size. The detection results are shown in Table 1. The multi-source heterogeneous remote sensing image change detection method in Example 1 outperforms existing methods in terms of average intersection-over-union (IoU) ratio, especially compared to the IST_CRF algorithm, where Example 1 achieves a 24.0% higher IoU ratio. Compared to AEKAN, which is also based on deep learning algorithms, it achieves a 3.1% higher IoU ratio. Furthermore, Example 1 demonstrates the best performance in both foreground and background IoU ratios, outperforming the IST_CRF algorithm by 41.7% in foreground IoU ratio and surpassing the second-best IRG_McS algorithm by 2.9%. Additionally, Example 1 outperforms the AEKAN algorithm by 2.2% and 3.7% in F1 score and Kappa coefficient, respectively. The results of Example 1 are visualized as follows: Figure 3The images shown are the optical image before the change, the SAR image after the change, and the binary change detection results. In the binary change detection results, black represents the unchanged area, white represents the changed area, green represents the area where the change was not detected, red represents the area where the unchanged area was detected as the changed area, and white represents the changed part.

[0030] Table 1 Comparison of test results .

[0031] In addition, to verify the effectiveness of the end-to-end evolution graph modeling mechanism and cross-modal graph distribution generalization mechanism proposed in this invention, the reconstruction constraints were compared and verified respectively. The experimental results are shown in Table 2: Table 2 Comparison of test results ; Comparative Example 1, without the aforementioned mechanism, achieved an average cross-union ratio (CUN) of 85.3%, a foreground CUN of 72.5%, an F1 score of 84.0%, and a Kappa coefficient of 83.0% on the Gloucester dataset. Example 2, after introducing the evolutionary graph modeling mechanism, improved the average CUN to 85.9%, the foreground CUN to 73.9%, and the F1 score to 85.0%. Example 1, further combined with the cross-modal graph distribution generalization mechanism, achieved an average CUN of 86.4%, a foreground CUN of 74.7%, a background CUN of 98.2%, an overall accuracy of 98.3%, an F1 score of 85.5%, and a Kappa coefficient of 84.6%, all achieving optimal results. Compared with the base model that does not employ the above mechanisms, the average crossover ratio (CRO) is improved by 1.1 percentage points, the foreground CRO is improved by 2.2 percentage points, the F1 score is improved by 1.5 percentage points, and the Kappa coefficient is improved by 1.6 percentage points. These results indicate that the synergistic effect of the two mechanisms can simultaneously improve the completeness, accuracy, and robustness of change detection results.

[0032] The above description is merely a detailed explanation of preferred embodiments and principles of the present invention. For those skilled in the art, there may be changes in specific implementation methods based on the ideas provided by the present invention, and these changes should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting changes in multi-source heterogeneous remote sensing images, characterized in that, Includes the following steps: S1. Normalize the multi-source heterogeneous remote sensing images before and after the change, and perform superpixel segmentation on the normalized multi-source heterogeneous remote sensing images to obtain superpixel blocks. S2. Input each superpixel block into the pre-constructed evolution graph network to extract the latent evolution graph representation. Based on the latent evolution graph representations of different modalities and after statistical distribution exchange, input them into the corresponding modality-specific decoder to reconstruct the superpixel blocks. Use the content consistency of the superpixel blocks before and after reconstruction to train the evolution graph network and obtain the trained and converged evolution graph network. S3. Based on the converged evolution graph network after training, change detection is performed on multi-source heterogeneous remote sensing images before and after the change, and the Otsu method is used to obtain the binary change detection results.

2. The method for detecting changes in multi-source heterogeneous remote sensing images according to claim 1, characterized in that, The multi-source heterogeneous remote sensing images include optical images of the first mode and synthetic aperture radar (SAR) images of the second mode.

3. The method for detecting changes in multi-source heterogeneous remote sensing images according to claim 2, characterized in that, In step S2, the extraction of the latent evolution map representation includes the following steps: S21, superpixel blocks Input Embedded Projection Layer Projection is performed to obtain the mapped features. ; S22, Based on mapping features The relationship features between pixels in a superpixel block are extracted through three learnable linear mappings: ； in, Indicates the first Based on Learnable mapping operators implemented through convolution. Indicates after the first A learnable mapping operator for dimensionality reduction of relational features It can be 1, 2, or 3; Pixels in a superpixel block With pixels The strength of the association between them is measured by their feature inner product, defined as: ； in, The relational features of pixel i after dimensionality reduction by the third learnable mapping operator are given. The relational features of pixel j after dimensionality reduction by the second learnable mapping operator; Normalizing the association strength yields the relationship matrix: ； Where N is the number of pixels in the superpixel block; S23. Based on the relation matrix, perform weighted aggregation of pixel information to obtain the graph structure features of pixel i: ； in, The relational features of pixel j after dimensionality reduction by the first learnable mapping operator; S24. Graph structure features after aggregating all pixels. A linear transformation is performed to obtain the enhanced graphical representation: ； in, These are learnable parameters; S25. Perform a nonlinear transformation on the enhanced graph representation and add it to the cross-temporal feature representation to obtain the latent evolution graph representation: ； in, This is the graphical representation after nonlinear transformation. For activation functions; The cross-temporal feature representation is extracted by the base feature extractor as the input for mapping features.

4. The method for detecting changes in multi-source heterogeneous remote sensing images according to claim 3, characterized in that, In step S2, the process of reconstructing the superpixel block includes the following steps: S26. Input the latent evolution graph representation and the corresponding nonlinearly transformed graph representation into a Transformer-based modality-specific decoder for superpixel block reconstruction, resulting in: ； in, and These are the superpixel blocks of the reconstructed first and second modes, respectively. and These are decoders corresponding to the first mode and the second mode, respectively. and These are the graphical representations of the nonlinear transformations corresponding to the first and second modes, respectively. and These are the potential evolution diagrams representing the first and second modes, respectively; S27. Constructing intramodal reconstruction constraints based on content consistency: ； in, and These are superpixel blocks corresponding to the first mode and the second mode, respectively.

5. The method for detecting changes in multi-source heterogeneous remote sensing images according to claim 4, characterized in that, In step S2, the process of reconstructing the superpixel block further includes the following steps: S28. Exchange the statistical distributions of the graph representations of different modal latent evolution graphs after nonlinear transformation: ； in, and These are the graph structure features corresponding to the exchange of statistical distributions between the first and second modes, respectively. This represents a feature transformation operation based on distribution conditions. and The graphs after nonlinear transformation are shown below. and The statistical distribution information includes the mean and variance. The graph structure features after statistical distribution swapping are input into a modality-specific decoder for superpixel block reconstruction along with the corresponding latent evolution graph representation. ； S28. Constructing cross-modal reconstruction constraints based on content consistency: 。 6. The method for detecting changes in multi-source heterogeneous remote sensing images according to claim 5, characterized in that, In step S2, an optimization objective function is constructed based on intra-modal reconstruction constraints and cross-modal reconstruction constraints: ； Evolutionary graph network is trained using an optimized objective function.

7. The method for detecting changes in multi-source heterogeneous remote sensing images according to any one of claims 2-6, characterized in that, Step S1 specifically includes: The optical images were preprocessed using the minimum-maximum normalization method. For SAR images, a logarithmic transformation is first performed, followed by preprocessing using the min-max normalization method; Superpixel segmentation is performed on the preprocessed image.

8. The method for detecting changes in multi-source heterogeneous remote sensing images according to claim 7, characterized in that, Superpixel segmentation is performed using simple linear iterative clustering.

9. A multi-source heterogeneous remote sensing image change detection system, employing the multi-source heterogeneous remote sensing image change detection method as described in any one of claims 1-8, characterized in that, The multi-source heterogeneous remote sensing image change detection system includes: The preprocessing module is used to normalize the multi-source heterogeneous remote sensing images before and after the changes, and to perform superpixel segmentation on the normalized multi-source heterogeneous remote sensing images to obtain superpixel blocks. The network training module is used to input superpixel blocks one by one into a pre-constructed evolution graph network to extract latent evolution graph representations. Based on the latent evolution graph representations of different modalities and after statistical distribution exchange, they are input into the corresponding modality-specific decoders to reconstruct the superpixel blocks. The evolution graph network is trained using the content consistency of the superpixel blocks before and after reconstruction, and the converged evolution graph network is obtained. The change detection module is used to detect changes in multi-source heterogeneous remote sensing images before and after changes based on the trained converged evolution graph network, and to obtain binary change detection results using the Otsu method.

10. A medium storing instructions, characterized in that, When the instructions are executed on a computer, the computer performs the method for detecting changes in multi-source heterogeneous remote sensing images as described in any one of claims 1-8.

Images