Hyperspectral change detection method based on joint attention and diffusion model

Abstract

The invention belongs to the technical field of remote sensing image processing, and discloses a hyperspectral change detection method based on a joint attention and diffusion model. The method comprises the following steps: firstly, performing data extraction and dimension reduction on a dual-time-phase hyperspectral image through a dual-branch network; performing differential operation on the dimension-reduced dual-time-phase features to obtain a preliminary difference feature map, and performing operation by adopting a multi-scale feature fusion strategy, convolution kernels of different scales and global pooling; designing and applying a joint channel-space attention module to perform modeling on the enhanced difference features; and applying the deeply optimized condition features as guidance information to the generation process of a diffusion model, and generating a target binary change mask image in an end-to-end manner. According to the method, the joint attention and diffusion model is introduced, deep modeling of channel and space levels on hyperspectral data is ensured, the interference resistance of the model to noise and tiny disturbance is improved, and the reliability and robustness of the model to tiny edge data processing are improved.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing and relates to a method for detecting hyperspectral changes based on a joint attention and diffusion model. Background Technology

[0002] In the field of remote sensing, long-term monitoring of natural surface changes and human-induced alterations has always been a crucial task. Change detection (CD) is a fundamental task in remote sensing, with its core objective being the accurate identification of the spatiotemporal evolution of land features in images of the same area acquired at different time phases. CD plays a significant role in practical applications such as urban sprawl monitoring, agricultural disaster assessment, and mineral resource surveys. With the continuous advancement of remote sensing technology, the spectral resolution and number of bands in acquired image data are constantly improving. Hyperspectral images (HSI), in particular, have become a hot topic in change detection research due to their ability to provide rich, continuous narrowband bands. HSI typically contains 100–300 bands in the visible, near-infrared, and even short-wave infrared range. The complete spectral curve carried by each pixel provides more information compared to optical images, not only helping to more finely distinguish land cover types but also sensitively capturing subtle spectral changes in materials, thus significantly improving the accuracy of change detection.

[0003] Traditional hyperspectral variation detection methods (HSI-CD) still face multiple challenges in practical applications: hyperspectral images have extremely high dimensionality and rich redundant information, and are highly sensitive to noise, making it difficult for algebraic methods based on pixel difference, principal component analysis, or change vector analysis to accurately capture fine-grained spectral-spatial-temporal variations. With the development of deep learning, researchers are attempting to apply deep learning methods to CD.

[0004] In 2024, Luo et al. constructed a differential fusion attention subnetwork in "DCENet: Diff-Feature Contrast Enhancement Network for Semi-Supervised Hyperspectral Change Detection" to extract temporal features from the initial input HSI image and enhance these features using a dual-branch enhancement module. However, this method relies on convolutional neural networks, which makes it difficult for the model to effectively capture long-range dependencies between the two temporal images. Furthermore, the attention mechanism proposed in this method is not specifically designed for HSI-CD, making it difficult to guarantee the deep fusion of high-dimensional spectral redundancy and fine-grained detail changes. In early 2025, Jiang et al. proposed a hybrid attention block in their paper "Adaptive Center-Focused Hybrid Attention Network for Change Detection in Hyperspectral Images," which includes an adaptive spatial-spectral hybrid self-attention module that dynamically adjusts spatial-spectral feature weights and a center-focusing attention module to improve the accuracy and boundary accuracy of hyperspectral image change detection. However, its inherent problems of blurred boundaries and inaccurate detail extraction still exist in pixel-level classification.

[0005] In recent years, diffusion models have demonstrated powerful capabilities in image generation and restoration. The inverse denoising process of diffusion models has proven highly effective for image deblurring and high-frequency detail reconstruction. In the field of hyperspectral remote sensing imagery, diffusion models have been used for super-resolution and image restoration.

[0006] However, existing research has not systematically integrated customized high-dimensional joint attention mechanisms with generative diffusion models possessing strong boundary recovery capabilities in an end-to-end manner. Existing generative methods focus on robust feature extraction and semantic association, but lack a unified framework to simultaneously address the three major challenges of hyperspectral redundancy, cross-temporal alignment, and the generation of high-precision boundary maps. Therefore, using a unified, redundancy-aware spatiotemporal-spectral joint attention mechanism as an efficient feature encoder and leveraging a diffusion model as a refinement generator represents a key gap in current research on hyperspectral change detection. Summary of the Invention

[0007] The purpose of this invention is to propose a hyperspectral change detection method based on joint attention and diffusion models. This method aims to fully explore and jointly model the channel and spatial features of hyperspectral images, and to realize an end-to-end generative framework by introducing a diffusion model, so as to solve the hyperspectral change detection problem efficiently and accurately.

[0008] The technical solution of the present invention: A hyperspectral change detection method based on joint attention and diffusion model is proposed. An end-to-end deep learning architecture is constructed, which includes a dual-branch dimensionality reduction and differential feature extraction module, a joint channel-spatial attention enhancement module, and a diffusion model generation module connected in sequence. First, the dual-temporal hyperspectral images are input into the dual-branch dimensionality reduction and differential feature extraction module's dual-branch dimensionality reduction network, which includes linear and nonlinear branches. The linear branch achieves feature decoupling and denoising based on the minimum noise separation principle, obtaining linear spectral features; the nonlinear branch obtains nonlinear spectral features based on convolution operations. The linear and nonlinear branches work together to achieve data dimensionality reduction, obtaining a dual-temporal dimensionality reduction feature representation with a high signal-to-noise ratio. Subsequently, the initial differential features are obtained by calculating the absolute difference of the dual-temporal dimensionality reduction feature representation. Then, a multi-scale feature fusion strategy is adopted, utilizing multiple convolution branches with different receptive fields and a global average pooling branch to enhance the information of the initial differential features, generating multi-scale enhanced differential features containing rich semantic information. Next, the joint channel-spatial attention enhancement module is applied to refine the multi-scale enhancement difference features by capturing the nonlinear correlation between spectral bands and the long-distance dependence of pixel space, and output deep guidance condition features. Finally, the deep guidance condition features are injected as condition guidance information into the diffusion model generation module to assist the reverse iterative denoising process of the diffusion model, and finally generate a binary change mask map that reflects the real changes of ground features.

[0009] The beneficial effects of this invention are: 1. Introduction of innovative architecture and generative framework: This invention is the first to combine the joint channel-space attention mechanism with the diffusion model for hyperspectral change detection tasks, and proposes an end-to-end generative change detection model, which greatly enriches the research paradigm of hyperspectral change detection.

[0010] 2. Efficient Dimensionality Reduction and Robust Feature Enhancement: This invention employs a bi-branch feature dimensionality reduction operation to efficiently remove redundancy and noise from hyperspectral data while ensuring no loss of key information. Subsequently, a multi-scale feature fusion module significantly enhances the discriminative power of differences between two temporal features, and a joint channel-spatial attention mechanism enables comprehensive and refined modeling of the differential features.

[0011] 3. Excellent noise resistance and direct result generation: This invention utilizes the powerful generation and denoising capabilities of the diffusion model to achieve direct generation of the result image with the changing mask, avoiding the complex post-processing steps in traditional methods. Simultaneously, the diffusion model's inherent denoising mechanism enhances the noise resistance and robustness of the final generated image, improving detection accuracy and stability in complex environments (such as low signal-to-noise ratios). Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the system framework of the present invention; Figure 2 This is a schematic diagram of the multi-scale difference feature fusion module proposed in this invention; Figure 3 This is a flowchart of the hyperspectral change detection system based on the joint attention and diffusion model described in an embodiment of the present invention. Detailed Implementation

[0013] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0014] A hyperspectral change detection method based on joint attention and diffusion models is proposed, and an end-to-end deep learning architecture is constructed, such as... Figure 1 As shown, the system includes a dual-branch dimensionality reduction and differential feature extraction module, a joint channel-spatial attention enhancement module, and a diffusion model generation module connected in sequence. The input of the entire system is a dual-temporal hyperspectral image, representing the image before and after the change, respectively. The output is a binary change mask image of the target, which enables accurate identification of the changed region.

[0015] First, the dual-temporal hyperspectral images are input into the dual-branch dimensionality reduction and differential feature extraction module's dual-branch dimensionality reduction network, which includes linear and nonlinear branches. The linear branch achieves feature decoupling and denoising based on the minimum noise separation principle, obtaining linear spectral features; the nonlinear branch obtains nonlinear spectral features based on convolution operations. The linear and nonlinear branches work together to achieve data dimensionality reduction, obtaining a dual-temporal dimensionality reduction feature representation with a high signal-to-noise ratio. Subsequently, the initial differential features are obtained by calculating the absolute difference of the dual-temporal dimensionality reduction feature representation. Then, a multi-scale feature fusion strategy is adopted, utilizing multiple convolution branches with different receptive fields and a global average pooling branch to enhance the information of the initial differential features, generating multi-scale enhanced differential features containing rich semantic information. Step 1: Based on the minimum shrinkage covariance noise separation algorithm, such as... Figure 3 As shown, the image before the change in the dual-temporal hyperspectral image is... and the changed image in the dual-temporal hyperspectral image The data is concatenated along the batch dimension, and then rearranged in both the batch and spatial dimensions to construct a merged sample matrix. ;in, For batch size, This represents the number of spectral bands. The spatial dimensions of the dual-temporal hyperspectral images This represents the total number of spatial pixels in the merged sample matrix; then, the merged sample matrix... Centralization is performed along the channel dimension to obtain the merged sample matrix. The mean vector of all samples along the channel dimension ,in, Represents the merged sample matrix One of the channel vectors; and from the merged sample matrix Subtract the mean vector from each column of samples The centered feature matrix is ​​obtained. ; The spatial difference method is used to estimate noise in dual-temporal hyperspectral images, and the horizontal difference feature used to represent the first-order spatial gradient change of image noise is calculated. and vertical difference features And the Laplacian operator features that reflect the degree of local second-order abrupt changes in image noise. ; Horizontal difference features Vertical difference features Features of the Laplace operator Horizontal concatenation is performed along the sample dimension to construct the noise estimation sample matrix. ; Horizontal difference features Vertical difference features Features of the Laplace operator The calculation formula is defined as follows: in, and Representing elements respectively In the corresponding centered feature matrix The first-dimensional index and the second-dimensional index; Further shrinkage covariance estimation is performed, and the signal covariance matrix is ​​calculated. and noise covariance matrix Simultaneously, a shrinkage estimate is introduced to calculate the noise shrinkage covariance estimate. : in, The shrinkage coefficient, For regularization parameters, It is the identity matrix. It is a diagonal matrix operator; Then, noise whitening is performed, starting with the noise shrinkage covariance estimation. Perform eigenvalue decomposition, represented as ,in, It is a noise shrinkage covariance estimation eigenvector matrix, Estimated by noise shrinkage covariance A diagonal matrix composed of eigenvalues; a whitening transformation matrix is ​​constructed based on the eigenvalue decomposition results. And according to the whitening transformation matrix and signal covariance matrix Calculate the covariance matrix of the whitened signal. ; covariance matrix of whitened signal Eigenvalue decomposition is performed, and the resulting eigenvalues ​​are used to characterize the covariance matrix of the whitened signal. The signal-to-noise ratio along the eigenvector direction corresponding to each eigenvalue; selecting the covariance matrix of the whitened signal. Center front The projection matrix is ​​composed of the eigenvectors corresponding to the largest eigenvalues. This leads to the linear transformation matrix based on the minimum noise separation algorithm for shrinkage covariance. The image before transformation is obtained through this linear transformation matrix. and the image after the change Mapped to linear dimensionality reduction features before change and linear dimensionality reduction features after transformation ; Build based on Nonlinear Refinement Mapping Networks for Convolution Operations ; Image before change and the image after the change Input nonlinear fine mapping network Obtain the nonlinear dimensionality reduction features before the change and nonlinear dimensionality reduction characteristics after transformation Convolution operations ensure nonlinear fine-mapping networks The number of output channels is set to ; Combining linear and nonlinear dimensionality reduction features along the channel dimension yields the original dimensionality reduction features. and dimensionality reduction features after transformation The number of channels for both is ; Finally, calculate the dimensionality reduction features before the change. and dimensionality reduction features after transformation The absolute difference between them yields the initial difference features. : Step 2, Multi-scale differential feature fusion module; such as Figure 2 As shown, the multi-scale difference feature fusion module; for the initial difference features obtained in step 1 Parallel passage , and The convolution operation branch and the global average pooling branch are set together, and the number of output channels in each branch is uniformly set to... The characteristics of the four branches are calculated as follows: in, This represents the convolution operation. Indicates batch normalization, Represents the ReLU activation function. This indicates a global average pooling operation. This indicates bilinear interpolation upsampling; Finally, the features calculated from the above four branches are... By stitching along the channel dimension, a multi-scale enhanced differential feature integrating multi-scale receptive field information is constructed. Its expression is as follows: Next, the joint channel-spatial attention enhancement module is applied to refine the multi-scale enhancement difference features by capturing the nonlinear correlation between spectral bands and the long-distance dependence of pixel space, and output deep guidance condition features. Enhanced differential features at multiple scales Introducing a joint channel-spatial attention enhancement module to obtain deep guidance condition features. The joint channel-space attention enhancement module includes a local adaptive channel attention mechanism and a global contextual spatial attention mechanism; among them, the local adaptive channel attention mechanism is used to enhance multi-scale differential features in the spectral channel dimension. It also calculates adaptive weights for each spectral channel; a global contextual attention mechanism is used to enhance differential features at multiple scales. The spatial dependencies are modeled to enhance the spatial structure information of the changing region and suppress noise interference; Step 1: Construct a local adaptive channel attention module; Firstly, a spatial guidance strategy is used to enhance multi-scale differential features. Weighted processing is performed; average pooling and max pooling operations are performed on the channel dimension to aggregate the spectral information of each spatial location and obtain a spatial statistical mapping map. in, and It is a multi-scale enhanced differential feature The spatial location in is The pixel on, the pixel in Calculate the average and maximum values ​​for each channel; The two spatial statistical mapping maps mentioned above are concatenated along the channel dimension, and then... Convolution operation and Sigmoid activation function Generate a spatially guided attention weight map : Then, the generated spatial guidance attention weight map is used. Enhanced differential features at multiple scales Perform element-wise multiplication to obtain the spatially weighted features. : in, This represents element-wise multiplication based on dimension. Based on spatial weighted features Average pooling and max pooling operations are performed in the spatial dimension to obtain the channel description vector, including the global semantic background. and local saliency features : Construct the channel self-attention computation process; incorporate the global semantic background. As input, through the linear mapping matrix Perform feature mapping to obtain the channel query vector. Channel key vector and channel value vector : in, Represents the channel query vector Channel key vector and channel value vector Feature dimensions; Calculate the channel query vector With channel key vector The scaled dot product similarity between channels is used to obtain the correlation matrix between channels, and then based on the channel value vectors... Perform weighted aggregation to obtain channel attention features. : in, Represents the normalized exponential activation function. This is the scaling factor; To further study channel attention features Adjustments are made by converting them into self-attention weights through a lightweight gating transformation. : in, This represents the Sigmoid activation function. This is a linear transformation operation; global semantic background and local saliency features Combining at the channel dimension, and through... The gating weights of the multilayer perceptron are obtained by performing nonlinear transformations on the multilayer perceptron structure composed of convolutions. , is represented as: Through element-wise multiplication Self-attention weights and multilayer sensor gating weights Multiply to obtain the channel gating weights. , is represented as: Finally, the channel gating weights are enhanced through a residual enhancement structure. Applied to spatial weighted features Local adaptive channel attention features are obtained. : Step 2: Construct a global context space attention module; First, to reduce computational complexity and expand the feature receptive range, multi-scale enhancement of differential features in the input is performed. implement Average pooling downsampling is used to obtain downsampling features. : in, Is adopted Average pooling downsampling operation of convolution kernels; Based on downsampling features A nonlocal self-attention computation process is constructed to capture long-range spatial dependencies between pixels; three sets of... Convolutional layers downsample features Projected as spatial query vectors Spatial key vector With spatial value vector : From spatial query vector With spatial key vector The transpose of the dot product yields the correlation matrix between spatial locations, based on the spatial value vector. Perform weighted aggregation to obtain spatial attention features. : in, Spatial query vector Spatial key vector With spatial value vector Feature dimensions; Through the above calculations, the downsampling features are made Each spatial location feature can be integrated with information from the entire map to establish long-distance relationships between spatial locations; then, through... Convolution operations restore the number of feature channels, and bilinear interpolation upsampling is used to adjust the spatial size of the spatial attention features to match the multi-scale enhanced difference features. Consistency, resulting in nonlocal enhancement features : Subsequently, nonlocal enhancement features Multiscale Enhanced Differential Features Nonlocal spatial fusion features are obtained by combining them using residual fusion. The process is represented as follows: Nonlocal spatial fusion features Average pooling and max pooling operations are performed at the channel level, and pixel-level spatial attention weight maps are generated through convolution operations. : Finally, element-wise multiplication is used to weight the spatial attention graph. Multiply by nonlocal spatial fusion features To obtain global context space attention features : Step 3: Cross-attention interaction and deep fusion; To achieve information interaction between different features, a nonlinear interaction method based on ternary element-wise multiplication is adopted to enhance the differential features at multiple scales. Local adaptive channel attention features and global context space attention features By fusing the features, we obtain the cross-attention features: To prevent multiple calculation processes from leading to multi-scale enhancement of differential features To address the loss, a learnable scalar parameter is introduced. Initialized to 1.0, spatial-spectral fusion features are obtained through residual connections. : Finally, through Convolution operation for spatial spectrum fusion features Nonlinear spatial spectrum reconstruction is performed, and batch normalization and ReLU activation function are combined to output deep guiding condition features for the diffusion model generation module. : Finally, the deep guidance condition features are injected as condition guidance information into the diffusion model generation module to assist the reverse iterative denoising process of the diffusion model, and finally generate a binary change mask map that reflects the real changes of ground features. Deep guidance condition characteristics As conditionally guided information input into the diffusion model, iterative denoising is used to achieve end-to-end change region detection; the diffusion model removes random noise... Initially, based on the characteristics of deep guidance conditions Guided by this principle, and through an iterative denoising process, the binary transformation mask map is ultimately generated end-to-end. ; The diffusion model is a parameterized denoising network that uses a symmetric UNet network, consisting of an encoder, intermediate feature layers, and a decoder; at a given time step Next, the noise reduction network reception time step Noisy change mask image and deep guidance condition characteristics For noisy change mask images Perform noise reduction processing and output the time step. Noisy change mask image ; The encoder is responsible for transforming the noisy, modified mask image. The encoder employs multi-level downsampling, comprising six layers, each consisting of a ResBlock network and a downsampling network. The decoder, through these six layers of ResBlock and upsampling networks, progressively reconstructs the noisy, modified mask image. Spatial resolution; The diffusion model minimizes the prediction noise. With actual noise The mean square error between them, learning the deep guidance condition features Denoising mapping under constraints; the objective function of the diffusion model is defined as: By minimizing the objective function Diffusion model learning in deep guiding condition features Under constraints, noise is accurately predicted and eliminated, and the change mask prediction is gradually obtained. Finally, based on the threshold The change mask prediction is converted into a binary state to obtain the final binary change mask map. The process is represented as: To further constrain network parameters, the following regularization constraints are introduced: in, For nonlinear refined mapping networks The learnable parameter weight matrix; At the same time, a regularization term based on the L2 norm is introduced. : in, The weight tensor parameters of each parameter layer in the multi-scale difference feature fusion module and the joint channel-spatial attention enhancement module; Therefore, the total loss function is expressed as: in, and Hyperparameters are used to balance the weights of each task.

Claims

1. A method for detecting hyperspectral changes based on a joint attention and diffusion model, characterized in that, The hyperspectral change detection method based on joint attention and diffusion model constructs an end-to-end deep learning architecture, including a dual-branch dimensionality reduction and differential feature extraction module, a joint channel-spatial attention enhancement module, and a diffusion model generation module connected in sequence. First, the dual-temporal hyperspectral images are input into the dual-branch dimensionality reduction and differential feature extraction module. The dual-branch dimensionality reduction network includes linear and nonlinear branches. The linear branch achieves feature decoupling and noise reduction based on the principle of minimum noise separation, and obtains linear spectral features. The nonlinear branch, based on convolution operations, obtains nonlinear spectral features; the linear and nonlinear branches work together to achieve data dimensionality reduction, obtaining a dual-temporal dimensionality reduction feature representation with high signal-to-noise ratio; Subsequently, the initial difference features are obtained by calculating the absolute difference of the bi-temporal dimensionality-reduced feature representation; Furthermore, a multi-scale feature fusion strategy is adopted, which utilizes multiple convolutional branches with different receptive fields and a global average pooling branch to enhance the information of the initial differential features and generate multi-scale enhanced differential features containing rich semantic information. Next, the joint channel-spatial attention enhancement module is applied to refine the multi-scale enhancement difference features by capturing the nonlinear correlation between spectral bands and the long-distance dependence of pixel space, and output deep guidance condition features. Finally, the deep guidance condition features are injected as condition guidance information into the diffusion model generation module to assist the reverse iterative denoising process of the diffusion model, and finally generate a binary change mask map that reflects the real changes of ground features.

2. The hyperspectral change detection method based on a joint attention and diffusion model according to claim 1, characterized in that, The specific operations of the dual-branch dimensionality reduction and differential feature extraction module are as follows: Step 1: Based on the minimum shrinkage covariance noise separation algorithm, separate the images before changes from the dual-temporal hyperspectral images. and the changed image in the dual-temporal hyperspectral image The data is concatenated along the batch dimension, and then rearranged in both the batch and spatial dimensions to construct a merged sample matrix. ;in, For batch size, This represents the number of spectral bands. The spatial dimensions of the dual-temporal hyperspectral images This represents the total number of spatial pixels in the merged sample matrix; then, the merged sample matrix... Centralization is performed along the channel dimension to obtain the merged sample matrix. The mean vector of all samples along the channel dimension ,in, Represents the merged sample matrix One of the channel vectors; and from the merged sample matrix Subtract the mean vector from each column of samples The centered feature matrix is ​​obtained. ; The spatial difference method is used to estimate noise in dual-temporal hyperspectral images, and the horizontal difference feature used to represent the first-order spatial gradient change of image noise is calculated. and vertical difference features And the Laplacian operator features that reflect the degree of local second-order abrupt changes in image noise. ; Horizontal difference features Vertical difference features Features of the Laplace operator Horizontal concatenation is performed along the sample dimension to construct the noise estimation sample matrix. ; Horizontal difference features Vertical difference features Features of the Laplace operator The calculation formula is defined as follows: in, and Representing elements respectively In the corresponding centered feature matrix The first-dimensional index and the second-dimensional index; Further shrinkage covariance estimation is performed, and the signal covariance matrix is ​​calculated. and noise covariance matrix Simultaneously, a shrinkage estimate is introduced to calculate the noise shrinkage covariance estimate. : in, The shrinkage coefficient, For regularization parameters, It is the identity matrix. It is a diagonal matrix operator; Then, noise whitening is performed, starting with the noise shrinkage covariance estimation. Perform eigenvalue decomposition, represented as ,in, It is a noise shrinkage covariance estimation eigenvector matrix, Estimated by noise shrinkage covariance A diagonal matrix composed of eigenvalues; a whitening transformation matrix is ​​constructed based on the eigenvalue decomposition results. And according to the whitening transformation matrix and signal covariance matrix Calculate the covariance matrix of the whitened signal. ; covariance matrix of whitened signal Eigenvalue decomposition is performed, and the resulting eigenvalues ​​are used to characterize the covariance matrix of the whitened signal. The signal-to-noise ratio along the eigenvector direction corresponding to each eigenvalue; selecting the covariance matrix of the whitened signal. Center front The projection matrix is ​​composed of the eigenvectors corresponding to the largest eigenvalues. This leads to the linear transformation matrix based on the minimum noise separation algorithm for shrinkage covariance. The image before transformation is obtained through this linear transformation matrix. and the image after the change Mapped to linear dimensionality reduction features before change and linear dimensionality reduction features after transformation ; Build based on Nonlinear Refinement Mapping Networks for Convolution Operations ; Image before change and the image after the change Input nonlinear fine mapping network Obtain the nonlinear dimensionality reduction features before the change and nonlinear dimensionality reduction characteristics after transformation Convolution operations ensure nonlinear fine-mapping networks The number of output channels is set to ; Combining linear and nonlinear dimensionality reduction features along the channel dimension yields the original dimensionality reduction features. and dimensionality reduction features after transformation The number of channels for both is ; Finally, calculate the dimensionality reduction features before the change. and dimensionality reduction features after transformation The absolute difference between them yields the initial difference features. : Step 2: Multi-scale differential feature fusion module; Multi-scale differential feature fusion module; Processing the initial differential features obtained in Step 1. Parallel passage , and The convolution operation branch and the global average pooling branch are set together, and the number of output channels in each branch is uniformly set to... The characteristics of the four branches are calculated as follows: in, This represents the convolution operation. Indicates batch normalization, Represents the ReLU activation function. This indicates a global average pooling operation. This indicates bilinear interpolation upsampling; Finally, the features calculated from the above four branches are... By stitching along the channel dimension, a multi-scale enhanced differential feature integrating multi-scale receptive field information is constructed. Its expression is as follows: 。 3. The hyperspectral change detection method based on a joint attention and diffusion model according to claim 1 or 2, characterized in that, The specific implementation process of the joint channel-spatial attention enhancement module is as follows: Enhanced differential features at multiple scales Introducing a joint channel-spatial attention enhancement module to obtain deep guidance condition features. ; The joint channel-space attention enhancement module includes a local adaptive channel attention mechanism and a global contextual spatial attention mechanism; among them, the local adaptive channel attention mechanism is used to enhance multi-scale differential features along the spectral channel dimension. It also calculates adaptive weights for each spectral channel; a global contextual attention mechanism is used to enhance differential features at multiple scales. The spatial dependencies are modeled to enhance the spatial structure information of the changing region and suppress noise interference; Step 1: Construct a local adaptive channel attention module; Firstly, a spatial guidance strategy is used to enhance multi-scale differential features. Weighted processing is performed; average pooling and max pooling operations are performed on the channel dimension to aggregate the spectral information of each spatial location and obtain a spatial statistical mapping map. in, and It is a multi-scale enhanced differential feature The spatial location in is The pixel on, the pixel in Calculate the average and maximum values ​​for each channel; The two spatial statistical mapping maps mentioned above are concatenated along the channel dimension, and then... Convolution operation and Sigmoid activation function Generate a spatially guided attention weight map : Then, the generated spatial guidance attention weight map is used. Enhanced differential features at multiple scales Perform element-wise multiplication to obtain the spatially weighted features. : in, This represents element-wise multiplication based on dimension. Based on spatial weighted features Average pooling and max pooling operations are performed in the spatial dimension to obtain the channel description vector, including the global semantic background. and local saliency features : Construct the channel self-attention computation process; incorporate the global semantic background. As input, through the linear mapping matrix Perform feature mapping to obtain the channel query vector. Channel key vector and channel value vector : in, Represents the channel query vector Channel key vector and channel value vector Feature dimensions; Calculate the channel query vector With channel key vector The scaled dot product similarity between channels is used to obtain the correlation matrix between channels, and then based on the channel value vectors... Perform weighted aggregation to obtain channel attention features. : in, Represents the normalized exponential activation function. This is the scaling factor; To further study channel attention features Adjustments are made by converting them into self-attention weights through a lightweight gating transformation. : in, This represents the Sigmoid activation function. This is a linear transformation operation; global semantic background and local saliency features Combining at the channel dimension, and through... The gating weights of the multilayer perceptron are obtained by performing nonlinear transformations on the multilayer perceptron structure composed of convolutions. , is represented as: Through element-wise multiplication Self-attention weights and multilayer sensor gating weights Multiply to obtain the channel gating weights. , is represented as: Finally, the channel gating weights are enhanced through a residual enhancement structure. Applied to spatial weighted features Local adaptive channel attention features are obtained. : Step 2: Construct a global context space attention module; First, to reduce computational complexity and expand the feature receptive range, multi-scale enhancement of differential features in the input is performed. implement Average pooling downsampling is used to obtain downsampling features. : in, Is adopted Average pooling downsampling operation of convolution kernels; Based on downsampling features A nonlocal self-attention computation process is constructed to capture long-range spatial dependencies between pixels; three sets of... Convolutional layers downsample features Projected as spatial query vectors Spatial key vector With spatial value vector : From spatial query vector With spatial key vector The transpose of the dot product yields the correlation matrix between spatial locations, based on the spatial value vector. Perform weighted aggregation to obtain spatial attention features. : in, Spatial query vector Spatial key vector With spatial value vector Feature dimensions; Through the above calculations, the downsampling features are made Each spatial location feature can be integrated with information from the entire map to establish long-distance relationships between spatial locations; then, through... Convolution operations restore the number of feature channels, and bilinear interpolation upsampling is used to adjust the spatial size of the spatial attention features to match the multi-scale enhanced difference features. Consistency, resulting in nonlocal enhancement features : Subsequently, nonlocal enhancement features Multiscale Enhanced Differential Features Nonlocal spatial fusion features are obtained by combining them using residual fusion. The process is represented as follows: Nonlocal spatial fusion features Average pooling and max pooling operations are performed at the channel level, and pixel-level spatial attention weight maps are generated through convolution operations. : Finally, element-wise multiplication is used to weight the spatial attention graph. Multiply by nonlocal spatial fusion features To obtain global context space attention features : Step 3: Cross-attention interaction and deep fusion; To achieve information interaction between different features, a nonlinear interaction method based on ternary element-wise multiplication is adopted to enhance the differential features at multiple scales. Local adaptive channel attention features and global context space attention features By fusing the features, we obtain the cross-attention features: To prevent multiple calculation processes from leading to multi-scale enhancement of differential features To address the loss, a learnable scalar parameter is introduced. Initialized to 1.0, spatial-spectral fusion features are obtained through residual connections. : Finally, through Convolution operation for spatial spectrum fusion features Nonlinear spatial spectrum reconstruction is performed, and batch normalization and ReLU activation function are combined to output deep guiding condition features for the diffusion model generation module. : 。 4. The hyperspectral change detection method based on a joint attention and diffusion model according to claim 3, characterized in that, The specific implementation process of the diffusion model generation module is as follows: Deep guidance condition characteristics As conditionally guided information input into the diffusion model, iterative denoising is used to achieve end-to-end change region detection; the diffusion model removes random noise... Initially, based on the characteristics of deep guidance conditions Guided by this principle, and through an iterative denoising process, the binary transformation mask map is ultimately generated end-to-end. ; The diffusion model is a parameterized denoising network that uses a symmetric UNet network, consisting of an encoder, intermediate feature layers, and a decoder; at a given time step Next, the noise reduction network reception time step Noisy change mask image and deep guidance condition characteristics For noisy change mask images Perform noise reduction processing and output the time step. Noisy change mask image ; The encoder is responsible for transforming the noisy, modified mask image. The encoder employs multi-level downsampling, comprising six layers, each consisting of a ResBlock network and a downsampling network. The decoder, through these six layers of ResBlock and upsampling networks, progressively reconstructs the noisy, modified mask image. Spatial resolution; The diffusion model minimizes the prediction noise. With actual noise The mean square error between them, learning the deep guidance condition features Denoising mapping under constraints; the objective function of the diffusion model is defined as: By minimizing the objective function Diffusion model learning in deep guiding condition features Under constraints, noise is accurately predicted and eliminated, and the change mask prediction is gradually obtained. Finally, based on the threshold The change mask prediction is converted into a binary state to obtain the final binary change mask map. The process is represented as: To further constrain network parameters, the following regularization constraints are introduced: in, For nonlinear refined mapping networks The learnable parameter weight matrix; At the same time, a regularization term based on the L2 norm is introduced. : in, The weight tensor parameters of each parameter layer in the multi-scale difference feature fusion module and the joint channel-spatial attention enhancement module; Therefore, the total loss function is expressed as: in, and Hyperparameters are used to balance the weights of each task.

Images