Model and data jointly driven unsupervised hyperspectral change detection method and apparatus

Abstract

The application relates to a model and data joint driving unsupervised hyperspectral change detection method and device. The method comprises the following steps: acquiring a plurality of double-time-phase hyperspectral remote sensing images, and extracting a plurality of super-pixel features from the double-time-phase hyperspectral remote sensing images. With the feature distance of the same type of super-pixel between each super-pixel feature and the feature similarity of the different types of super-pixel between each super-pixel feature as the reference, an optimization model for the discriminative change component in the initial change feature matrix is constructed. The discriminative change component is optimized. On the one hand, a flexible self-training network is introduced into the model optimization, the unsupervised hyperspectral change detection without manual annotation data is realized by combining prior knowledge and deep information mining; on the other hand, the key component that can best reflect the real change of the ground surface can be automatically selected from the redundant initial change feature, the pseudo change caused by the background noise and seasonal fluctuation is inhibited, and the detection accuracy of the model is improved.

Description

Technical Field

[0001] This application relates to the field of remote sensing image processing technology, and in particular to an unsupervised hyperspectral change detection method and apparatus driven by a combination of model and data. Background Technology

[0002] Change detection technology utilizes multi-temporal remote sensing image data obtained from remote sensing sensors. Based on image processing and mathematical models, it identifies the state differences of surface entities or phenomena at different times. This technology provides crucial support for scientific decision-making in human activities and is of great significance for sustainable land management and ecological environmental protection. Among various data sources, hyperspectral images, with hundreds of spectral bands, can provide near-continuous spectral curves of ground features. Change detection technology based on hyperspectral images can fully utilize the rich spectral information in hyperspectral images to achieve fine-grained identification of surface changes.

[0003] In conventional approaches, some methods achieve change detection by integrating manually designed prior knowledge and physical constraints into specialized empirical mathematical models. However, these methods suffer from poor accuracy when faced with complex changes or imaging noise interference. Other methods achieve change detection by training deep learning networks, but the training process relies on a large amount of costly labeled sample data, and the networks lack interpretability. Summary of the Invention

[0004] Therefore, it is necessary to provide a model- and data-driven unsupervised hyperspectral change detection method, device, computer equipment, computer-readable storage medium, and computer program product that can solve the above-mentioned technical problems.

[0005] Firstly, this application provides an unsupervised hyperspectral change detection method jointly driven by a model and data. The method includes: Multiple dual-temporal hyperspectral remote sensing images are acquired, and an initial variation feature matrix is ​​extracted from the dual-temporal hyperspectral remote sensing images; wherein, the initial variation feature matrix includes multiple superpixel features; Using the feature distance between similar superpixels among the superpixel features and the feature similarity between dissimilar superpixels among the superpixel features as references, an optimization model is constructed for the discriminative change components in the initial change feature matrix. Based on the auxiliary variables, the optimization model is decoupled to obtain the expressions for the auxiliary variables and the changing components. During the preset iteration process, the deep unfolded network containing the expressions of auxiliary variables and the expressions of change components is trained in an unsupervised manner. When the iteration termination condition is met, the final discriminative change component is output. Based on the final discriminative change components, a clustering method is used to output the change detection results map of the dual-temporal hyperspectral remote sensing image.

[0006] Secondly, this application also provides an unsupervised hyperspectral change detection device jointly driven by a model and data. The device includes: The feature acquisition module is used to acquire multiple dual-temporal hyperspectral remote sensing images and extract an initial change feature matrix from the dual-temporal hyperspectral remote sensing images; wherein, the initial change feature matrix includes multiple superpixel features; The model building module is used to build an optimization model for the discriminative change components in the initial change feature matrix, taking the feature distance between similar superpixels and the feature similarity between dissimilar superpixels as references. The model decoupling module decouples the optimization model based on auxiliary variables, obtaining expressions for the auxiliary variables and the changing components. The processing module is used to perform unsupervised training on a deep unfolded network containing expressions for auxiliary variables and expressions for change components during a preset iteration process, and output the final discriminative change component when the iteration termination condition is met. The output module is used to output the change detection result map of the dual-temporal hyperspectral remote sensing image based on the final discriminative change component through a clustering method.

[0007] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the above method steps.

[0008] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the above method steps.

[0009] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, implements the above method steps.

[0010] The above-mentioned unsupervised hyperspectral change detection methods, devices, computer equipment, storage media, and computer program products driven by the combined models and data; Multiple dual-temporal hyperspectral remote sensing images are acquired, and an initial variation feature matrix is ​​extracted from the dual-temporal hyperspectral remote sensing images; wherein, the initial variation feature matrix includes multiple superpixel features.

[0011] By acquiring dual-temporal images, a data foundation is provided for subsequent time-series comparisons; superpixel features effectively suppress common noise in hyperspectral images and preserve the edge contours of ground objects, reducing the complexity of subsequent calculations.

[0012] Using the feature distance between similar superpixels and the feature similarity between dissimilar superpixels as references, an optimization model is constructed for the discriminative change components in the initial change feature matrix.

[0013] The optimized model simultaneously considers minimizing the distance between similar regions and minimizing the similarity between dissimilar regions. It forces the model to make the feature distribution of similar regions more compact and the feature distribution of dissimilar regions more loose in the transformed feature space, thereby significantly improving the distinguishability between similar and dissimilar regions. Optimization of discriminative change components can automatically select the key components that best reflect the true changes on the land surface from redundant initial change features and suppress spurious changes caused by background noise.

[0014] Based on auxiliary variables, the optimization model is decoupled to obtain the auxiliary variable expression and the change component expression. Introducing auxiliary variables for decoupling decomposes the complex original problem into two relatively simple sub-problems, which can improve the efficiency of network training.

[0015] During the preset iteration process, the deep unfolded network containing the auxiliary variable expression and the change component expression is trained unsupervised. When the iteration termination condition is met, the final discriminative change component is output. Deep information mining is performed through the self-trained PnP network corresponding to the auxiliary variable expression. No manual sample labeling is required, which reduces the cost of network training.

[0016] Based on the final discriminative change components, a clustering method is used to output the change detection result map of the dual-temporal hyperspectral remote sensing image.

[0017] This method has two advantages. First, superpixel features effectively suppress common noise in hyperspectral images. By optimizing the model for discriminative change components, it can automatically select the key components that best reflect the true changes on the land surface from redundant initial change features, suppressing spurious changes caused by background noise. Based on the joint driving mode of the optimized model and data, it can still output highly accurate change detection results even in the face of complex changes or imaging noise interference. Second, the unsupervised training process of the deep unfolded network does not require costly labeled samples, greatly reducing the network training cost. The deep unfolded network contains specific variable expressions, and the output discriminative change components have corresponding physical meanings, improving the interpretability of the network. Attached Figure Description

[0018] Figure 1This is a flowchart illustrating an unsupervised hyperspectral change detection method driven by both model and data in one embodiment. Figure 2 This is a flowchart illustrating a method for extracting an initial change feature matrix in one embodiment; Figure 3 This is a flowchart illustrating a method for determining an optimization model in one embodiment; Figure 4 This is a schematic diagram of the optimization solution process in one embodiment; Figure 5 This is a block diagram of an unsupervised hyperspectral change detection device jointly driven by model and data in one embodiment. Figure 6 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0020] In one embodiment, such as Figure 1 As shown, a model- and data-driven unsupervised hyperspectral change detection method is provided, including: S101: Acquire multiple dual-temporal hyperspectral remote sensing images and extract the initial change feature matrix from the dual-temporal hyperspectral remote sensing images.

[0021] The initial change feature matrix includes multiple superpixel features.

[0022] Among them, dual-temporal hyperspectral remote sensing images can represent remote sensing satellite or aerial images of the same geographic area at two different points in time. For example, hyperspectral satellite remote sensing images of a certain area before and after a fire. The hyperspectral representation in dual-temporal hyperspectral remote sensing images divides the spectrum into extremely fine segments, and can simultaneously record tens or even hundreds of continuous and very narrow (generally less than 10 nanometers) bands of information.

[0023] The initial variation feature matrix can contain information such as differences, ratios, and correlations of each pixel or region across different bands. This initial variation feature matrix transforms the original, massive dual-temporal hyperspectral image (potentially containing millions of pixels × 200 bands) into a lower-dimensional, more condensed initial feature representation, providing input data for subsequent optimization models.

[0024] Superpixel features can be numerical vectors generated to describe each pixel block. It's important to understand that superpixel features don't describe a single pixel, but rather an entire region. For example, for a superpixel block, superpixel features could be the mean, variance, maximum / minimum values ​​of all pixels within that block, as well as geometric features such as the block's area and compactness.

[0025] Superpixel features can be obtained by processing dual-temporal hyperspectral remote sensing images using the superpixel algorithm. The superpixel algorithm divides the image into many small patches with consistent internal features and clear boundaries between them. The boundaries of these patches are usually attached to the edges of real-world features in the image, such as road edges, riverbanks, and building boundaries.

[0026] Specifically, extracting an initial variation feature matrix from dual-temporal hyperspectral remote sensing images can include: extracting features from the first-temporal and second-temporal remote sensing images respectively using a dual-branch convolutional network to obtain first-temporal features and second-temporal features; performing difference operations on the first-temporal and second-temporal features to obtain difference features; filtering the bands corresponding to the difference features from multiple difference features to obtain depth difference features corresponding to target bands that meet the band filtering conditions; performing band enhancement processing on the depth difference features based on band attention weights to obtain band-enhanced difference features; and performing superpixel segmentation on the band-enhanced difference features to obtain an initial variation feature matrix containing multiple superpixel features.

[0027] First, the two branches process their respective time phases independently, avoiding information confusion and making the learned features more interpretable. Second, hyperspectral difference features may have more than 200 bands, and direct processing is computationally intensive and prone to overfitting. Selecting and retaining key bands significantly improves computational efficiency. Third, the SLIC algorithm's specially optimized boundary-fitting capability ensures that superpixels do not cross the edges of real-world features. The final change detection map maintains the complete outlines of buildings and clear boundaries of forests, rather than fragmented scattered points, thus improving the model's detection accuracy.

[0028] Specifically, extracting the initial change feature matrix from the dual-temporal hyperspectral remote sensing image may further include: performing superpixel segmentation on the dual-temporal hyperspectral remote sensing image to obtain an initial change feature matrix containing multiple superpixel features.

[0029] By acquiring dual-temporal images, a data foundation is provided for subsequent time-series comparisons; superpixel features effectively suppress common noise in hyperspectral images and preserve the edge contours of ground objects, reducing the complexity of subsequent calculations.

[0030] S102, using the feature distance between similar superpixels among superpixel features and the feature similarity between dissimilar superpixels among superpixel features as references, construct an optimization model for the discriminative change components in the initial change feature matrix.

[0031] In this context, "same-class superpixels" can refer to two superpixels that belong to either the changing class or the invariant class. For example, if the regions corresponding to two superpixels represent the same feature in both the first and second time phases, then these two superpixels belong to the invariant class and are considered "same-class superpixels." Conversely, if the regions corresponding to two superpixels represent different features in both the first and second time phases, then these two superpixels belong to the changing class and are considered "same-class superpixels."

[0032] Feature distance can represent the numerical difference between the features of two superpixels. Specifically, it can be the distance between superpixels of the same type.

[0033] Outlier superpixels can be defined as two superpixels belonging to the changing and invariant classes, respectively. For example, if one superpixel corresponds to the same feature in the first and second time phases, and another superpixel corresponds to a different feature in the first and second time phases, then these two superpixels belong to the invariant and changing classes, respectively, and are thus called outlier superpixels.

[0034] Feature similarity is an indicator of how close two superpixel features are. Specifically, it can be the similarity between two dissimilar superpixels.

[0035] Discriminative variation components can represent spurious changes caused by noise removal while retaining signals generated by actual changes in surface objects.

[0036] The optimization model can be a mathematical expression for the discriminative variation component. This optimization model can be a mathematical equation with specific constraints or regularization objectives, including but not limited to: minimizing the feature distance between similar superpixels and minimizing the feature similarity between dissimilar superpixels.

[0037] Specifically, after obtaining the features of each superpixel, an optimization model is constructed for the discriminative change components in the initial change feature matrix based on the constraints of minimizing the distance between similar classes and minimizing the similarity between dissimilar classes.

[0038] Specifically, after obtaining the features of each superpixel, the intra-class aggregation constraint expression for the discriminative change components in the initial change feature matrix can be constructed using the feature distance between superpixels of the same class among the features as a reference; the inter-class separation constraint expression for the discriminative change components can be constructed using the feature similarity between dissimilar superpixels among the features as a reference; penalties are applied to the non-zero change-related elements in each superpixel feature to obtain the sparsity constraint expression for the discriminative change feature components; and an optimization model for the discriminative change components is constructed based on the intra-class aggregation constraint expression, the inter-class separation constraint expression, and the sparsity constraint expression.

[0039] First, after optimization, superpixel features of the same class tend to cluster in a small neighborhood in the feature space, causing superpixel features of the variable class and the invariant class to move closer to their respective class centers, enhancing intra-class compactness. Second, when features of different change types (variable regions and invariant regions) overlap in space, the model has difficulty distinguishing them. Inter-class separation constraints force the features of the variable class and the invariant class to maintain distance, reducing confusion between classes. Third, many bands in hyperspectral images contain noise or redundant information. Sparse constraints automatically compress the weights of these useless features to zero, achieving implicit feature selection and improving the model's generalization ability.

[0040] More specifically, for constructing intra-class aggregation constraint expressions for discriminative change components in the initial changed feature matrix, with reference to the feature distance between similar superpixel features, the process includes: constructing an intra-class graph by using each superpixel feature as a vertex; filtering similar vertices in the intra-class graph and determining the intra-class edge weights for each type of inner edge in the intra-class graph; wherein the expression for the intra-class edge weights includes the average feature distance between similar vertices; and constructing intra-class aggregation constraint expressions for discriminative change components in the initial changed feature matrix based on the various inner edge weights.

[0041] First, the edges of the intra-class graph can connect any number of superpixel vertices, allowing the model to simultaneously capture the global clustering structure among multiple superpixels of the same type, thus more completely preserving the inherent distribution patterns of both changed and invariant regions. Second, for clusters of the same type with highly consistent features (high confidence), the model applies strong aggregation pressure; for clusters whose features are relatively dispersed (potentially containing noise or boundary regions), the model applies weaker aggregation pressure, avoiding forced convergence that could lead to information distortion. This achieves adaptive weighted clustering and enhances the model's robustness. Third, the set-weighted aggregation constraint forces superpixels belonging to the same side to maintain feature similarity, reducing noise and making the edges of ground features in the change detection results clearer and the regions more complete.

[0042] More specifically, for constructing inter-class separation constraint expressions for discriminative change components with reference to the feature similarity of dissimilar superpixels among each superpixel feature, the process includes: constructing an inter-class graph by using each superpixel feature as a vertex; filtering dissimilar vertices in the inter-class graph and determining the inter-class edge weights of each edge in the inter-class graph; wherein the expression for the inter-class edge weights includes the average feature similarity of dissimilar vertices; and constructing inter-class separation constraint expressions for discriminative change components in the initial change feature matrix based on the inter-class edge weights.

[0043] First, inter-class graphs explicitly characterize the proximity relationships and topological structures between vertices of different classes by connecting vertices of different classes. This allows the optimization model to constrain the similarity of vertices of different classes, thereby enhancing the separability between different classes. Second, the lower the similarity (the less similar) between two vertices of different classes, the larger the edge weight, and the stronger the separation constraint imposed by the model on the vertices within that edge. Conversely, the edge weight of vertices of different classes with high similarity is small, and the separation constraint is weak. This adaptive weighting mechanism makes the separation constraint more intelligent and efficient. Third, for hard-to-distinguish samples that are located at class boundaries and have similar features, graph inter-class constraints consider the complex associations formed by multiple intermediate vertices and mine the higher-order topological relationships between vertices in the graph, thereby enhancing the ability to distinguish hard-to-distinguish samples.

[0044] Specifically, an optimization model is constructed for the discriminative change components in the initial change feature matrix, taking into account the feature distance between similar superpixels and the feature similarity between dissimilar superpixels.

[0045] The optimized model simultaneously considers minimizing the distance between similar categories and minimizing the similarity between dissimilar categories. It forces the model to make the feature distribution of regions belonging to both change and invariance more compact in the transformed feature space, and to make the features of regions belonging to change and invariance more dispersed, thereby significantly improving the distinguishability between changed and unchanged regions. The optimization of discriminative change components can automatically select the key components that best reflect the true changes on the land surface from redundant initial change features, and suppress spurious changes caused by background noise and seasonal fluctuations.

[0046] S103. Based on the auxiliary variables, the optimization model is decoupled to obtain the expressions for the auxiliary variables and the changing components.

[0047] Auxiliary variables are intermediate variables introduced to simplify the solution process for the changing components.

[0048] An auxiliary variable expression can be a mathematical formula used to describe the relationship between the auxiliary variable and the variable component.

[0049] The expression for the variation component can be a mathematical formula used to describe the relationship between the variation component and the original variable (image features).

[0050] By using auxiliary variables, the coupling terms are broken down into different subproblems. Each subproblem becomes a standard problem with mature and fast solutions (such as soft thresholding, Fourier transform, and least squares). This significantly improves the overall solution efficiency and enhances model training efficiency.

[0051] S104. During the preset iteration process, the deep unfolded network containing the expressions of auxiliary variables and the expressions of change components is trained in an unsupervised manner. When the iteration termination condition is met, the final discriminative change component is output.

[0052] Here, "deep unfolded network" refers to a trainable deep network obtained by mapping the iterative algorithm using an end-to-end mapping method. A deep unfolded network includes expressions for auxiliary variables and expressions for changing components.

[0053] Specifically, the deep unfolded network includes a PnP network corresponding to the expression of the auxiliary variables. The PnP network (Plug-and-Play Network) is used to solve the subproblem of the auxiliary variables. Based on the mathematical expression of the auxiliary variables, their solution can be obtained using the PnP network, thereby leveraging the deep learning model to deeply mine information about changes. In one optimization iteration, the PnP network is trained multiple times, and the trained PnP network is used to update the auxiliary variables.

[0054] The iteration termination condition can be the difference in overall accuracy between adjacent iterations. If more than three times in a row If the value is no longer greater than 0, then stop iterating.

[0055] Specifically, during the preset iteration process, a deep unfolded network containing expressions for auxiliary variables and changing components is trained. When the iteration termination condition is met, the final discriminative changing component is output. This includes: in the first iteration, pseudo-labels generated using traditional unsupervised methods on the training set are used, and the top N percent of pseudo-labels with confidence rankings are selected as the sample set. A deep unfolded network containing expressions for auxiliary variables and changing components is trained based on this sample set. In the next iteration (target number of iterations), pseudo-labels are obtained by passing the PnP network through a normalization layer, and the top A percent of pseudo-labels with confidence rankings are selected as the target sample set. The fusion result of the target sample set and the sample set is used as the training sample for the next iteration. This process continues until the deep unfolded network meets the iteration termination condition, at which point the final discriminative changing component is output. Where the target number of iterations is greater than 1, A = N × the current iteration number.

[0056] First, selecting the top N percentile high-confidence pseudo-labels as the sample set effectively filters out noise and errors in the output of traditional methods, preventing the model from learning incorrect patterns in the initial stage and reducing the risk of training divergence. Second, using the currently trained PnP network to generate new pseudo-labels, the quality of which is usually better than the previous round (because the model is improving), forming a positive loop of model evolution to pseudo-label evolution. Third, fusing the old sample set (from traditional methods or earlier iterations) with the new sample set (from the current model) preserves historical high-confidence knowledge. The model does not lose the correct patterns it has already learned while learning new knowledge.

[0057] S105, based on the final discriminative change component, the change detection result map of the dual-temporal hyperspectral remote sensing image is output through a clustering method.

[0058] The final discriminative change component can be the discriminative change component output when the deep unfolded network completes its iteration.

[0059] The change detection map can be the clustering result of the final discriminative change components. The change detection result map can be a binary map or a probability map with the same image size as the input dual-temporal hyperspectral image.

[0060] Specifically, the discriminant change component These are superpixel-level features, which are converted into pixel-level features. Then, the PCA-Kmeans method is applied to obtain the change detection result map through clustering.

[0061] In the unsupervised hyperspectral change detection method driven by the above model and data, multiple dual-temporal hyperspectral remote sensing images are acquired, and an initial change feature matrix is ​​extracted from the dual-temporal hyperspectral remote sensing images; wherein, the initial change feature matrix includes multiple superpixel features.

[0062] By acquiring dual-temporal images, a data foundation is provided for subsequent time-series comparisons; superpixel features effectively suppress common noise in hyperspectral images and preserve the edge contours of ground objects, reducing the complexity of subsequent calculations.

[0063] Using the feature distance between similar superpixels and the feature similarity between dissimilar superpixels as references, an optimization model is constructed for the discriminative change components in the initial change feature matrix.

[0064] The optimized model simultaneously considers minimizing the distance between similar regions and minimizing the similarity between dissimilar regions. It forces the model to make the feature distribution of similar regions more compact and the feature distribution of dissimilar regions more loose in the transformed feature space, thereby significantly improving the distinguishability between similar and dissimilar regions. Optimization of discriminative change components can automatically select the key components that best reflect the true changes on the land surface from redundant initial change features and suppress spurious changes caused by background noise.

[0065] Based on auxiliary variables, the optimization model is decoupled to obtain the auxiliary variable expression and the change component expression. Introducing auxiliary variables for decoupling decomposes the complex original problem into two relatively simple sub-problems, which can improve the efficiency of network training.

[0066] During the preset iteration process, the deep unfolded network containing the auxiliary variable expression and the change component expression is trained unsupervised. When the iteration termination condition is met, the final discriminative change component is output. Deep information mining is performed through the self-trained PnP network corresponding to the auxiliary variable expression. No manual sample labeling is required, which reduces the cost of network training.

[0067] Based on the final discriminative change components, a clustering method is used to output the change detection result map of the dual-temporal hyperspectral remote sensing image.

[0068] This method has two advantages. First, superpixel features effectively suppress common noise in hyperspectral images. By optimizing the model for discriminative change components, it can automatically select the key components that best reflect the true changes on the land surface from redundant initial change features, suppressing spurious changes caused by background noise. Based on the joint driving mode of the optimized model and data, it can still output highly accurate change detection results even in the face of complex changes or imaging noise interference. Second, the unsupervised training process of the deep unfolded network does not require costly labeled samples, greatly reducing the network training cost. The deep unfolded network contains specific variable expressions, and the output discriminative change components have corresponding physical meanings, improving the interpretability of the network.

[0069] In one embodiment, such as Figure 2 The flowchart shown illustrates the method for extracting the initial change feature matrix. Specifically, step S101 above, extracting the initial change feature matrix from a dual-temporal hyperspectral remote sensing image, includes: S201, a dual-branch convolutional network is used to extract features from the first and second phase remote sensing images in the dual-temporal hyperspectral remote sensing images, respectively, to obtain the first phase features and the second phase features.

[0070] Among them, a dual-branch convolutional network can be an architecture that includes two parallel convolutional neural network branches with the same structure.

[0071] The first phase remote sensing image can be an image acquired at an earlier time, while the second phase remote sensing image can be an image acquired at a later time.

[0072] The first and second temporal features can be high-dimensional abstract representations extracted from two temporal remote sensing images respectively through a bi-branch convolutional network. They are the results of compression, dimensionality upscaling, and nonlinear transformation of the original image information. The first temporal features are obtained by feature extraction from the first temporal remote sensing image, and the second temporal features are obtained by feature extraction from the second temporal remote sensing image.

[0073] Specifically, a dual-branch convolutional network is used to extract features from the first-phase and second-phase remote sensing images in the dual-temporal hyperspectral remote sensing images, respectively, to obtain the first-phase features. Second phase characteristics .

[0074] S202, perform a difference operation on the first phase feature and the second phase feature to obtain the difference feature.

[0075] Specifically, regarding the characteristics of the first time phase Second phase characteristics Differential operations are performed to obtain the difference features. The expression is as follows: .

[0076] S203, among multiple difference features, the bands corresponding to the difference features are screened to obtain the depth difference features corresponding to the target bands that meet the band screening conditions.

[0077] The target band can be a band with a larger variance. A larger band variance contains more useful information about changes. By calculating the variance of each band and sorting them, the band with the largest variance is selected as the target band.

[0078] Specifically, from multiple differences In this process, the bands corresponding to the difference features are filtered to obtain the depth difference features corresponding to the target bands that meet the band filtering conditions. and the corresponding band-selected dual-temporal image and .

[0079] S204, based on band attention weights, performs band enhancement processing on depth difference features to obtain band-enhanced difference features.

[0080] Among them, the band attention weight can be a weight set according to the importance of different bands.

[0081] Specifically, a similarity-guided band enhancement strategy can be designed by leveraging the differences in contribution across different bands. This strategy calculates features using cosine similarity. and Similarity between corresponding bands: ; in, Representing the Each band, This indicates the L2-norm operation.

[0082] The smaller the similarity (i.e., the greater the difference) of the bands Including more reliable change information contributes more to change detection performance; therefore, similarity-based band attention weights are defined. as follows: ; Based on band attention weights For depth difference features Perform band enhancement processing (element-by-element multiplication and residual calculation) to obtain the differential characteristics after band enhancement. The expression is as follows: ; in, This indicates element-wise multiplication.

[0083] S205, perform superpixel segmentation on the differential features after band enhancement to obtain an initial variation feature matrix containing multiple superpixel features.

[0084] Superpixel segmentation (SLIC) is an algorithm that merges spatially adjacent pixels with similar color, texture, and other features into irregular small blocks (i.e., superpixels).

[0085] Specifically, the superpixel segmentation SLIC method can be used to extract the difference features after band enhancement. The image is divided into M superpixels. By averaging the pixels within each superpixel in each band, superpixel-level features are obtained. These superpixel-level features are then integrated to construct a 2D difference feature matrix. This serves as the initial feature matrix for optimization modeling. Wherein, It is the difference feature vector of the m-th superpixel.

[0086] In this embodiment, firstly, the two branches process their respective temporal phases independently, avoiding information confusion and making the learned features more interpretable. Secondly, hyperspectral difference features may have more than 200 bands, and direct processing would be computationally intensive and prone to overfitting. Selecting and retaining key bands significantly improves computational efficiency. Thirdly, the SLIC algorithm's specially optimized boundary-fitting capability ensures that superpixels do not cross the edges of real-world features. The final change detection result map maintains the complete outlines of buildings and clear boundaries of forests, rather than fragmented scattered points, thereby improving the model's prediction accuracy.

[0087] In one embodiment, an optimization model for the discriminative change components in the initial changed feature matrix is ​​constructed, taking the feature distance between similar superpixels and the feature similarity between dissimilar superpixels as references. This includes: constructing intra-class aggregation constraint expressions for the discriminative change components in the initial changed feature matrix, taking the feature distance between similar superpixels as references; constructing inter-class separation constraint expressions for the discriminative change components, taking the feature similarity between dissimilar superpixels as references; penalizing non-zero change-related elements in each superpixel feature to obtain sparsity constraint expressions for the discriminative change components; and constructing an optimization model for the discriminative change components based on the intra-class aggregation constraint expressions, inter-class separation constraint expressions, and sparsity constraint expressions.

[0088] Wherein, the initial change feature matrix Interference from mid-class features can affect the discriminative power of intra-class and inter-class features, making accurate change recognition difficult. From the perspective of image decomposition, It can be viewed as a component of discriminative change. and interference components Combining, that is: ; To suppress feature interference and achieve accurate change detection, the core is to... Extract The corresponding optimization objective is: ; The first item is the data fidelity item, used to ensure the accuracy of the variable. Fitting the original data The second term is a regularization constraint term, used to constrain... Its characteristics.

[0089] A well-designed set of constraints is crucial for extracting discriminative change features. Crucial. Through analysis and its To address the relationship between the variable feature subspace and the classical global sparsity change prior, this embodiment proposes a priori relationships and constructs three constraints based on these constraints. The resulting optimization model with inherent change prior constraints is as follows: ; in, It is an L1-norm operation. and The regularization parameter used to balance the constraint strength, This is an intra-class aggregation constraint expression. This is a separation constraint expression between classes. This is a sparse constraint expression.

[0090] Specifically, for sparse constraint expressions, based on the L1 norm... This is a classic sparsity constraint derived from the widely used global sparse change prior. In change detection tasks, changes are typically low-probability events, therefore the changed region is much smaller than the unchanged region. Because... Each column represents the intensity of change in a superpixel, while the values ​​in non-change regions approach zero. It exhibits global sparsity, containing a large number of near-zero elements. Therefore, a penalty is imposed on the non-zero change-related elements in each superpixel feature to constrain... The sparsity is used to obtain a sparsity constraint expression for the discriminative change feature components. .

[0091] For example, such as Figure 3 The flowchart illustrates the method for determining the optimization model. It involves extracting features from the initial change features, obtaining sparsity constraints based on the optimization objective using global sparsity priors, obtaining intra-class aggregation constraints and inter-class separation constraints based on the optimization objective using the subspace relationship priors of the change features (construction of intra-class graph structure and inter-class graph structure), and then constructing an optimization model for the discriminative change components (an optimization model with change prior constraints) using the intra-class aggregation constraint expressions, inter-class separation constraint expressions, and sparsity constraint expressions.

[0092] In this embodiment, firstly, after optimization, superpixel features of the same type tend to cluster in a small neighborhood in the feature space, causing the superpixel features of the variable class and the invariant class to move closer to their respective class centers, enhancing intra-class compactness. Secondly, when features of different change types (variable regions and invariant regions) overlap in space, the model has difficulty distinguishing them. Inter-class separation constraints force the features of the variable class to maintain distance, reducing confusion between classes. Thirdly, many bands in hyperspectral images contain noise or redundant information. Sparse constraints automatically compress the weights of these useless features to zero, achieving implicit feature selection and improving the model's generalization ability.

[0093] In one embodiment, using the feature distance between similar superpixels among superpixel features as a reference, an intra-class aggregation constraint expression is constructed for the discriminative change components in the initial changed feature matrix, including: Each superpixel feature is used as a vertex in the intra-class graph to construct the intra-class graph; vertices of the same type are filtered in the intra-class graph, and the intra-class edge weights of each type of intra-class edge are determined; the expression of the intra-class edge weights includes the average feature distance of vertices of the same type; based on the intra-class edge weights, an intra-class aggregation constraint expression is constructed for the discriminative change components in the initial change feature matrix.

[0094] In this context, each vertex in the intra-class graph can represent a superpixel feature.

[0095] An intra-class graph can represent a special graph structure (such as a hypergraph) in which an edge can connect any number (≥2) of vertices, not just two vertices.

[0096] The intra-class edge weight is assigned a numerical value to each edge, representing the overall aggregation or importance among all vertices (superpixels) within that edge.

[0097] Specifically, constraint terms Utilizing the prior knowledge of the changing feature subspace relations, let the matrix The features of superpixels of the same class should be as similar as possible (achieving intra-class aggregation). The core logic is: using an initial feature matrix... Using the feature distance of the same type of superpixel as a reference, the feature distance of the corresponding superpixel in B is constrained by graph regularization.

[0098] Intraclass graph construction process: (1) Graph structure definition: Initial feature matrix Feature vectors of each superpixel (i.e., superpixel features) are considered as vertices of the intra-class graph. To construct an intra-class graph Each class's internal edge From the vertices of the class graph The vertices of the same type constitute the superpixel's higher-order intra-class association. (2) Selection of similar vertices: First, the PCA-Kmeans method is used to filter the initial feature matrix. Perform unsupervised clustering to distinguish between changing and invariant superpixels; then calculate the similarity matrix using cosine similarity. For each vertex, select the K vertices with the highest, medium, and lowest similarity in the same class to form 3 edges connecting vertices of the same class (avoid full connections to reduce computation). (3) Calculation of intra-class edge weights: Intra-class edge weights The weights represent the degree of connection between vertices, and the formula is as follows: ; in, , They are respectively in the initial characteristic matrix The corresponding feature vector in. inner edge of class The number of vertices in the middle, where N is the number of edges. Calculate the Euclidean distance. This represents the average feature distance of all vertices of the same type.

[0099] Intra-class aggregation constraint formula: using intra-class edge weights Constructing intra-class aggregation constraints This is achieved by minimizing the Euclidean eigendistance between vertices of the same type. The aggregation of similar features within a class, with the following intra-class aggregation constraint expression: ; Where, in the formula, and These are the vertices of the intra-class graph. and In discriminative features (matrix) The corresponding feature vector in Represents the trace operation of a matrix. Let be the graph Laplace matrix.

[0100] The constraint logic is as follows: in the initial feature matrix The smaller the feature distance between vertices of the same class (i.e., the more similar the variation features of vertices of the same class), the higher their intra-class edge weights. The larger, in The stronger the constraints (in) The more similar the vertices should be, the more adaptively these vertices will be brought closer together, thus promoting intra-class aggregation.

[0101] In this embodiment, firstly, edges can connect any number of superpixel vertices, enabling the model to simultaneously capture the global clustering structure among multiple superpixels of the same type, thus preserving the inherent distribution patterns of similar regions more completely. Secondly, for clusters of the same type with highly consistent features (high confidence), the model applies strong aggregation pressure; for clusters whose features are relatively dispersed (potentially containing noise or boundary regions), the model applies weaker aggregation pressure, avoiding forced convergence that could lead to information distortion, achieving adaptive weighted clustering and enhancing model robustness. Thirdly, set-weighted aggregation constraints force superpixels belonging to the same edge to maintain feature similarity, reducing noise and making the edges of ground features in the change detection results clearer and the regions more complete.

[0102] In one embodiment, taking the feature similarity between dissimilar superpixels among the features of each superpixel as a reference, an inter-class separation constraint expression for the discriminative variation component is constructed, including: Each superpixel feature is used as a vertex in the inter-class graph to construct the inter-class graph; heterogeneous vertices are filtered in the inter-class graph, and the inter-class edge weights of each edge in the inter-class graph are determined; the expression of the inter-class edge weights includes the average feature similarity of heterogeneous vertices; based on the inter-class edge weights, an inter-class separation constraint expression is constructed for the discriminative change components in the initial change feature matrix.

[0103] In this context, each vertex in the inter-class graph can represent a superpixel feature.

[0104] Inter-class graphs can represent a general graph structure where an edge can connect two vertices.

[0105] Inter-class edge weights are assigned a numerical value to each edge, representing the overall aggregation or importance among all vertices (superpixels) within that edge.

[0106] Specifically, constraint terms By utilizing the changing feature subspace relationships, the feature matrix to be optimized can be optimized. The features of heterogeneous superpixels are made as dissimilar as possible (to achieve inter-class separation). The core logic is: using an initial feature matrix... Using the feature similarity of heterogeneous superpixels as a reference, graph regularization constraints are applied. The feature similarity of the corresponding heterogeneous superpixels.

[0107] Inter-class graph construction process: (1) Graph structure definition: with the initial feature matrix Feature vectors of each superpixel (i.e., superpixel features) are vertices of the inter-class graph. Construct inter-class graph Each inter-class edge From the vertices of the inter-class graph (2) Selection of outlier vertices: Based on the PCA-Kmeans clustering results and similarity matrix. Selection. For each vertex, select the K vertices with the highest, medium, and lowest cosine similarity from the heterogeneous set to form edges (each vertex forms 3K edges, constructing an inter-class graph structure). (3) Calculation of inter-class edge weights: Inter-class edges The weight is used to measure the degree of connection between connected vertices, and the calculation formula is as follows: ; In the formula, , They are respectively in the initial characteristic matrix The corresponding feature vector. For inter-class edges The number of vertices in the array Let be the total number of edges. calculate and Cosine similarity between them This represents the average feature similarity among all out-of-class vertices.

[0108] Inter-class separation constraint formula: using inter-class weights Constructing inter-class separation constraints By minimizing the feature similarity of out-of-class vertices, we can achieve... Separation of vertices of different classes. The inter-class separation constraint expression is as follows: ; Where, in the formula, and These are the vertices of the inter-class graph. and In discriminative features (matrix) The corresponding feature vector. Representation matrix The matrix after column normalization based on column norm. It is a diagonal matrix with diagonal elements as follows: The norm of each column, A is the graph The adjacency matrix.

[0109] The constraint logic is as follows: in the initial feature matrix The smaller the similarity between vertices (i.e., the less similar the variation features between vertices), the higher the weight of the inter-class edge. The larger, in The stronger the constraints (in) The more dissimilar the vertices should be, the more adaptively these vertices will be separated, thus improving the separability between classes.

[0110] In this embodiment, firstly, the inter-class graph explicitly characterizes the proximity relationships and topological structure between vertices of different classes by connecting vertices of different classes, enabling the optimization model to constrain the similarity of vertices of different classes, thereby enhancing the separability between different classes. Secondly, the lower the similarity (the less similar) between two vertices of different classes, the larger the edge weight, and the stronger the separation constraint imposed by the model on the vertices within that edge; conversely, the edge weight of vertices of different classes with high similarity is small, and the separation constraint is weak. This adaptive weighting mechanism makes the separation constraint more intelligent and efficient. Thirdly, for hard-to-distinguish samples that are located at class boundaries and have similar features, the graph inter-class constraint considers the complex associations formed by multiple intermediate vertices and mines the high-order topological relationships between vertices in the graph, thereby enhancing the ability to distinguish hard-to-distinguish samples.

[0111] In one embodiment, during a preset iteration process, an unsupervised training is performed on a deep unfolded network containing expressions for auxiliary variables and expressions for change components. Upon meeting the iteration termination condition, the final discriminative change component is output, including: In the first iteration, pseudo-labels generated using traditional unsupervised methods on the training set are used, and the top N percent of pseudo-labels with the highest confidence ranking are selected as the sample set. A deep unfolded network containing expressions for auxiliary variables and changing components is trained based on this sample set. The deep unfolded network includes a PnP network corresponding to the expressions for the auxiliary variables.

[0112] In the iteration of the target number of iterations, the PnP network corresponding to the auxiliary variable expression is normalized to output pseudo-labels, and the pseudo-labels with the highest confidence ranking of A are selected as the target sample set. The fusion result of the target sample set and the sample set is used as the training sample in the next iteration. The final discriminative change component is output when the deep unfolded network meets the iteration termination condition. Where the target number of iterations is greater than 1, A = N × the current iteration number.

[0113] In this context, pseudo-labels can be simulated labels automatically generated for unlabeled data when no real labels are available. Specifically, in the first iteration, a traditional unsupervised method is used to generate a change detection result map. In subsequent iterations, the PnP network outputs a change probability map for the input image. After a normalization layer, high-confidence predictions are used as new pseudo-labels.

[0114] The sample set can be a collection of high-confidence samples selected from pseudo-labels, used to train the PnP network.

[0115] A normalization layer can be an operational layer in a neural network that scales the input data to a fixed numerical range (such as [0,1] or [-1,1]).

[0116] In the t-th iteration, the target sample set is a new set composed of the samples with the highest confidence ranking of A from the pseudo-labels output by the current PnP network.

[0117] Training samples, the actual dataset used for the next training iteration, are formed by fusing the sample set from the previous round with the target sample set for the current round. Specifically, the fusion method can be to take the union of the two sets to increase the training sample size, or to assign different weights to samples from different sources (e.g., high weight for initial pseudo-labels, low weight for subsequent pseudo-labels). The fusion method can be flexibly chosen according to the actual situation and is not limited here.

[0118] In one embodiment, before training the model, constructing an optimization model for the discriminative variation components based on intra-class aggregation constraint expressions, inter-class separation constraint expressions, and sparsity constraint expressions can specifically be as follows: ; in, To optimize the objective (the optimized discriminative features), The initial characteristic matrix, For data fidelity items, For the sparsity constraint expression, For discriminative features (matrix) For the graph Laplace matrix, and The regularization parameter used to balance the constraint strength, Represents the trace operation of a matrix. It is a diagonal matrix with diagonal elements as follows: The norm of each column, A is the graph The adjacency matrix.

[0119] After obtaining the optimized model, the semi-quadratic splitting method (HQS) can be used to optimize and solve the model. First, auxiliary variables are introduced. The augmented Lagrangian function of the optimization model is constructed as follows: ; in, This is the penalty parameter.

[0120] In this way, the optimization model is decoupled into two sub-problems, as follows: ; ; When optimizing the solution to the two subproblems, the variables... and The solution process involves alternating updates over a total of T iterations, with optimization as follows: Figure 4 As shown. The optimization model is transformed into one that includes variation components. Expressions and auxiliary components The expression's deep unfolding network, where The subproblem is updated using gradient descent. The subproblems are updated through a self-trained PnP network.

[0121] Specifically, for The subproblem can be solved using gradient descent, as shown below: ; In the formula, The learning rate controls the step size for updating various parameters. Indicates the first This is the second iteration. In each iteration t, the result is updated once using this formula. No labeled samples are required for training.

[0122] for The subproblem can be solved by combining it with an unlabeled PnP network, letting ,but The optimization problem of the subproblem can be rewritten as: ; Based on the deep PnP method, The subproblem is used as a Moro approximation operator, and the optimization problem is solved using an advanced deep learning network. This enables the unsupervised hyperspectral change detection method of this application to combine the deep image information mining capabilities of deep learning with the high interpretability of the optimization model. At the same time, the PnP network allows for a flexible network structure, eliminating the need to design complex self-supervised training strategies. Discriminative change feature extraction can be achieved simply by flexibly injecting pseudo-label information during the optimization process.

[0123] Design a self-training PnP network Optimization in an unsupervised manner Sub-problem. The PnP network extracts local-global features through a two-branch Graph Convolutional Network (GCN) and a Graph Transformer, suppressing feature interference in B. In each iteration t, the updated features are... enter The network performs E rounds of self-training based on pseudo-labels, and outputs updated features after training is complete. Deep learning networks can be used flexibly to uncover potential change information and remove interfering information irrelevant to change detection from the change features.

[0124] Specifically, The network is self-trained using progressively high-confidence pseudo-labels and the prior constraint unsupervised loss proposed in this embodiment, without the need for real label (GT) supervision.

[0125] In a total of T optimization iterations: In the first iteration (t=1), the difference features including band enhancement are used. The training set was used to generate pseudo-labels using traditional unsupervised methods, and the top 0.5% with high confidence were selected as the pseudo-label sample set. train .

[0126] In the t-th iteration (t>1), The pseudo-labels obtained from the softmax output of the network are selected, with the top (0.5×t)% of high-confidence samples chosen as the target sample set. Target sample set With pseudo-label sample set The samples are fused to obtain the training samples for the next iteration. , For use in the next iteration Network training. The final discriminative change component is output when the deep unfolded network meets the iteration termination condition.

[0127] As the number of iterations increases, the high-confidence pseudo-label sample set is gradually expanded, thereby improving the network training effect.

[0128] The proposed prior-constrained unsupervised loss function for training PnP networks It consists of four parts, and the expression is as follows: ; Among them, sparsity loss : Used to implement Global sparsity constraints in subproblems, for A penalty is applied to the element values ​​of superpixels that are related to the changes.

[0129] Binary cross-entropy loss Based on the pseudo-label sample set Ω (t=1) or the training samples in the next iteration (t>1) Calculate.

[0130] Comparative loss Used to reduce Increasing the feature distance between samples within a class and increasing the feature distance between samples between classes helps in the learning of discriminative feature representations.

[0131] Deep Embedding Clustering (DEC) Loss Used to uncover potential information in unlabeled samples, prompting The network clusters similar embeddings to the same center and separates different categories, driving the model to learn compact and discriminative features without labeled information.

[0132] Variables during optimization and Alternating updates. This is used to obtain discriminative features in the absence of ground truth (GT). ,by The performance in detecting changes on the pseudo-labeled sample set is used as the criterion for stopping the iteration. The difference in overall accuracy between adjacent iterations is denoted as... If more than three times in a row If the value is no longer greater than 0, the iteration stops. The final updated value is... This refers to the discriminative features extracted by the optimization model. .

[0133] The optimized model at the end of the iteration transforms the initial change feature matrix D into discriminative features. Then, it is converted into pixel-level features, and the PCA-Kmeans method is applied to obtain the change detection result map corresponding to the output hyperspectral dual-temporal image through clustering.

[0134] In this embodiment, firstly, a flexible model-data jointly driven unsupervised hyperspectral change detection framework is constructed. This framework can incorporate prior knowledge of change detection through optimized modeling and obtain potential change information that is difficult for prior knowledge to capture through deep image information mining. Secondly, a priori relationship of change feature subspaces is proposed, describing the potential relationship between changed and unchanged features in different subspaces. Utilizing this change prior in optimized modeling can achieve intra-class feature aggregation and inter-class feature separation, improving the distinguishability between changed and invariant features. Thirdly, an interpretable self-trained network is proposed. By embedding a plug-and-play (PnP) network into the optimized modeling process, the interpretability and simplicity of the unsupervised deep learning network are improved. This network can flexibly inject pseudo-label information under unlabeled conditions to deeply mine potential change information in the data.

[0135] In this embodiment, firstly, high-confidence pseudo-labels ranking in the top N percent are selected as the sample set, effectively filtering out noise and errors in the output of traditional methods, preventing the model from learning incorrect patterns in the initial stage, and reducing the risk of training divergence. Secondly, new pseudo-labels are generated using the currently trained PnP network. These pseudo-labels are usually of better quality than those from the previous round (because the model is improving), forming a positive loop of model evolution leading to pseudo-label evolution. Thirdly, the old sample set (from traditional methods or earlier iterations) is merged with the new sample set (from the current model), preserving historical high-confidence knowledge. The model does not lose the correct patterns it has already learned while learning new knowledge.

[0136] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0137] Based on the same inventive concept, this application also provides a model- and data-driven unsupervised hyperspectral change detection device for implementing the model- and data-driven unsupervised hyperspectral change detection method described above. The solution provided by this device is similar to the implementation described in the above method. Therefore, the specific limitations of one or more model- and data-driven unsupervised hyperspectral change detection device embodiments provided below can be found in the limitations of the model- and data-driven unsupervised hyperspectral change detection method described above, and will not be repeated here.

[0138] In one embodiment, such as Figure 5 As shown, an unsupervised hyperspectral change detection device jointly driven by a model and data is provided, comprising: a feature acquisition module 501, a model construction module 502, a model decoupling module 503, and a processing module 504, wherein: The feature acquisition module 501 is used to acquire multiple dual-temporal hyperspectral remote sensing images and extract an initial change feature matrix from the dual-temporal hyperspectral remote sensing images; wherein, the initial change feature matrix includes multiple superpixel features; The model building module 502 is used to build an optimization model for the discriminative change components in the initial change feature matrix, with reference to the feature distance between similar superpixels among the superpixel features and the feature similarity between dissimilar superpixels among the superpixel features. The model decoupling module 503 decouples the optimization model based on the auxiliary variables to obtain the auxiliary variable expression and the change component expression; The processing module 504 is used to perform unsupervised training on a deep unfolded network containing the auxiliary variable expression and the change component expression during a preset iteration process, and output the final discriminative change component when the iteration termination condition is met.

[0139] The output module 505 is used to output the change detection result map of the dual-temporal hyperspectral remote sensing image based on the final discriminative change component through a clustering method.

[0140] The modules in the unsupervised hyperspectral change detection device driven by the aforementioned model and data can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0141] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 6As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores superpixel feature data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements a model- and data-driven unsupervised hyperspectral change detection method.

[0142] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0143] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method steps.

[0144] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described method steps.

[0145] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the above-described method steps.

[0146] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0147] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0148] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A model- and data-driven unsupervised hyperspectral change detection method, characterized in that, The method includes: Multiple dual-temporal hyperspectral remote sensing images are acquired, and an initial variation feature matrix is ​​extracted from the dual-temporal hyperspectral remote sensing images; wherein, the initial variation feature matrix includes multiple superpixel features; Using the feature distance between similar superpixels among the superpixel features as a reference, an intra-class aggregation constraint expression is constructed for the discriminative change component in the initial change feature matrix; using the feature similarity between dissimilar superpixels among the superpixel features as a reference, an inter-class separation constraint expression is constructed for the discriminative change component; penalties are applied to non-zero change-related elements in each superpixel feature to obtain a sparsity constraint expression for the discriminative change feature component; based on the intra-class aggregation constraint expression, the inter-class separation constraint expression, and the sparsity constraint expression, an optimization model is constructed for the discriminative change component; Wherein, "same type superpixels" means that two superpixels belong to the same variable class or the same invariant class; "different type superpixels" means that two superpixels belong to the variable class and the invariant class, respectively. Based on the auxiliary variables, the optimization model is decoupled to obtain the expressions for the auxiliary variables and the changing components. During the preset iteration process, the deep unfolded network containing the auxiliary variable expression and the change component expression is trained in an unsupervised manner. When the iteration termination condition is met, the final discriminative change component is output. Based on the final discriminative change components, a clustering method is used to output the change detection result map of the dual-temporal hyperspectral remote sensing image.

2. The method according to claim 1, characterized in that, The step of extracting the initial change feature matrix from the dual-temporal hyperspectral remote sensing image includes: Feature extraction was performed on the first and second phase remote sensing images in the dual-temporal hyperspectral remote sensing images using a dual-branch convolutional network, resulting in first-phase features and second-phase features. The difference features are obtained by performing a difference operation on the first phase feature and the second phase feature; Among multiple difference features, the bands corresponding to the difference features are filtered to obtain the depth difference features corresponding to the target bands that meet the band filtering conditions. Based on the band attention weight, the depth difference features are subjected to band enhancement processing to obtain the band-enhanced difference features; The enhanced difference features of the band are segmented into superpixels to obtain an initial variation feature matrix containing multiple superpixel features.

3. The method according to claim 1, characterized in that, The step of constructing an intra-class aggregation constraint expression for the discriminative change components in the initial changed feature matrix, using the feature distance between similar superpixels among the superpixel features as a reference, includes: The superpixel features are used as vertices of the intra-class graph to construct an intra-class graph. In the intra-class graph, vertices of the same type are filtered, and the intra-class edge weights of each type of intra-class edge in the intra-class graph are determined; wherein, the expression of the intra-class edge weights includes the average feature distance of vertices of the same type; Based on the intra-class edge weights, an intra-class aggregation constraint expression is constructed for the discriminative change components in the initial change feature matrix.

4. The method according to claim 1, characterized in that, Using the feature similarity between dissimilar superpixels among the aforementioned superpixel features as a reference, an inter-class separation constraint expression is constructed for the discriminative change component, including: The superpixel features are used as vertices of the inter-class graph to construct the inter-class graph; In the inter-class graph, out-of-class vertices are filtered, and the inter-class edge weights of each edge in the inter-class graph are determined; wherein, the expression for the inter-class edge weights includes the average feature similarity of out-of-class vertices; Based on the inter-class edge weights, an inter-class separation constraint expression is constructed for the discriminative change components in the initial change feature matrix.

5. The method according to claim 1, characterized in that, In the preset iteration process, the deep unfolded network containing the auxiliary variable expression and the change component expression is trained unsupervised. When the iteration termination condition is met, the final discriminative change component is output, including: In the first iteration, pseudo-labels generated by traditional unsupervised methods on the training set are used, and the top N percent of pseudo-labels with confidence ranking are selected as the sample set. A deep unfolding network containing the auxiliary variable expression and the change component expression is trained based on the sample set. The deep unfolding network includes the PnP network corresponding to the auxiliary variable expression. In the iteration of the target number of iterations, the PnP network corresponding to the auxiliary variable expression is passed through a normalization layer to obtain pseudo-labels, and the pseudo-labels with the highest confidence ranking of A are selected as the target sample set. The fusion result of the target sample set and the sample set is used as the training sample in the next iteration. The final discriminative change component is output when the deep unfolded network meets the iteration termination condition. Wherein, the target number of iterations is greater than 1, and A = N × the current iteration number.

6. An unsupervised hyperspectral change detection device jointly driven by model and data, characterized in that, The device includes: The feature acquisition module is used to acquire multiple dual-temporal hyperspectral remote sensing images and extract an initial change feature matrix from the dual-temporal hyperspectral remote sensing images; wherein, the initial change feature matrix includes multiple superpixel features; The model building module is used to construct intra-class aggregation constraint expressions for the discriminative change components in the initial change feature matrix, using the feature distance between similar superpixels among the superpixel features as a reference; construct inter-class separation constraint expressions for the discriminative change components, using the feature similarity between dissimilar superpixels among the superpixel features as a reference; apply penalties to non-zero change-related elements in each superpixel feature to obtain sparsity constraint expressions for the discriminative change feature components; and construct an optimization model for the discriminative change components based on the intra-class aggregation constraint expressions, the inter-class separation constraint expressions, and the sparsity constraint expressions; wherein, similar superpixels indicate that two superpixels belong to the same change class or the same invariant class; dissimilar superpixels indicate that two superpixels belong to the change class and the invariant class, respectively; The model decoupling module decouples the optimization model based on auxiliary variables, obtaining expressions for the auxiliary variables and expressions for the changing components; The processing module is used to perform unsupervised training on a deep unfolded network containing the auxiliary variable expression and the change component expression during a preset iteration process, and output the final discriminative change component when the iteration termination condition is met. The output module is used to output the change detection result map of the dual-temporal hyperspectral remote sensing image based on the final discriminative change component through a clustering method.

Images