Hyperspectral anomaly detection method fusing graph attention and beta wavelet graph network

Abstract

The application discloses a hyperspectral anomaly detection method fusing graph attention and beta wavelet graph network, which regards the pixel of a hyperspectral image as a graph node, regards spectral features as node attributes, establishes edge connection through a K nearest neighbor algorithm, and constructs a graph data structure containing spatial and spectral information. Firstly, the graph attention network is used to dynamically aggregate neighborhood information and enhance spatial context features. Secondly, a beta wavelet filter bank is constructed based on a graph Laplacian matrix, multi-scale frequency domain filtering is performed on node features, and the right shift phenomenon of spectral energy caused by an abnormal target is effectively captured. Finally, multi-scale filtering features are fused and mapped into an abnormal probability, and pixel-level anomaly detection is realized. Through the joint mining of nonlinear spatial correlation and unique spectral anomaly features of hyperspectral data by the graph neural network, the accuracy and robustness of anomaly detection in a complex scene are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of hyperspectral remote sensing anomaly detection technology, and more specifically, to a hyperspectral anomaly detection method using a fused graph attention and beta wavelet graph network. Background Technology

[0002] Hyperspectral remote sensing technology can simultaneously acquire spatial information and continuous, detailed spectral information of ground objects, making it possible to discover unknown, small-sized anomalous targets in complex backgrounds. It has important application value in fields such as military reconnaissance, environmental monitoring, and mineral exploration.

[0003] Traditional hyperspectral anomaly detection methods mainly include statistical models, collaborative representations, and low-rank sparse decomposition. These methods typically rely on specific data distribution assumptions, such as Gaussian distribution, or use fixed local windows for background modeling. Their performance is limited when processing hyperspectral data with complex nonlinear structures, non-Gaussian distributions, or noise interference. In recent years, deep learning-based methods have made some progress by learning features in a data-driven manner.

[0004] However, existing deep learning methods are mostly based on convolutional neural networks, which struggle to effectively model long-range dependencies between pixels and do not fully utilize the spatial-spectral joint characteristics of hyperspectral data, resulting in complex models and computational inefficiencies. Graph neural networks offer a new approach to modeling complex spatial relationships between pixels in hyperspectral data, but current research focuses primarily on classification tasks, and effective spectral feature extraction mechanisms for anomaly detection tasks remain to be explored.

[0005] Therefore, there is an urgent need for a new detection method that can efficiently integrate the spatial and spectral information of hyperspectral images and accurately capture the unique characteristics of abnormal targets. Summary of the Invention

[0006] In view of this, the present invention proposes a hyperspectral anomaly detection method that integrates graph attention and beta wavelet graph networks, including the following steps: S1, Graph structure modeling step: The three-dimensional hyperspectral image data cube is modeled as graph structure data, where each pixel is regarded as a node in the graph, the feature vector of the node is its spectral vector, and edge connections are established for the nodes based on spectral similarity to simultaneously embed the spatial and spectral information of the image; S2, Spatial feature enhancement step: The graph structure data is processed using a graph attention network, and the node neighbor information is aggregated through an adaptive attention mechanism to dynamically learn and enhance the spatial context features related to anomaly detection; S3, Spectral anomaly capture step: Based on the graph Laplacian matrix, a set of beta wavelet filters is constructed to form a graph wavelet transform group, and the node features enhanced in step S2 are subjected to multi-scale frequency domain filtering to specifically capture the right shift of spectral energy caused by anomalous targets; S4, Anomaly discrimination and output step: The filtered spectral features are fused and mapped, the probability of each pixel being an anomaly is calculated, and the final anomaly detection result image is generated.

[0007] Furthermore, in step S1, the K-nearest neighbor algorithm is used to establish edge connections for each node, that is, to calculate the spectral distance between all pairs of nodes and to select the K other nodes with the closest spectral distance to each node to establish connection edges.

[0008] Furthermore, the operation of the graph attention network in step S2 includes: performing a linear transformation on the features of each node; calculating the attention weights between the node and each of its neighboring nodes; performing a weighted summation of the transformed features of the neighboring nodes based on the attention weights; and using an activation function to obtain the updated spatial augmentation features of the node.

[0009] Furthermore, the beta wavelet filter described in step S3 is constructed based on a parameterized beta function, which is defined as follows: ,in Let be the shape parameter, and satisfy . C is the wavelet order.

[0010] Furthermore, the beta wavelet filter bank W consists of C+1 filters of different scales, denoted as... ,in One is a low-pass filter, and the rest are band-pass or high-pass filters.

[0011] Furthermore, in step S3, the phenomenon of right shift of spectral energy refers to the fact that when there is an abnormal target in the image, the energy distribution of the signal in the Fourier transform domain will shift to the high-frequency part, and the beta wavelet filter is configured to selectively enhance the high-frequency energy.

[0012] Further, step S4 specifically includes: aggregating all the wavelet-scale filtered features from step S3; mapping the aggregated features to a scalar using a multilayer perceptron; and applying a Sigmoid activation function to convert the scalar into a value representing the probability of anomalies.

[0013] Furthermore, the method performs end-to-end training by minimizing a weighted cross-entropy loss function, which is defined as: Where, 𝑁 represents the total number of pixels, 𝑦 𝑖 For the true label of pixel i, 𝑝 𝑖 This represents the probability that pixel i is an anomalous, as predicted by the model.

[0014] Further, step S1 specifically includes: S11, Input 3D hyperspectral data cube And the number of nearest neighbors K; S12. Reconstruct the data cube into a two-dimensional matrix. ,in Each row represents the spectral feature vector of one pixel; S13. Define each pixel as a graph node. This forms the node set V; S14. For each node Calculate the Euclidean distance of its spectral characteristics with all other nodes, select the K nodes with the smallest distance as its neighbors, and establish an edge. This forms the edge set ε; S15. Construct the adjacency matrix A and degree matrix D based on the edge set, and calculate the Laplacian matrix L=DA of the graph accordingly. S16, Output Diagram Structure And its Laplace matrix L.

[0015] Further, step S2 specifically includes: S21, Input Graph Structure and their adjacency relationships; S22. Initialize node characteristics ; S23. For each node Perform the following operations: S231. Apply a learnable weight matrix W to perform a linear transformation: ; S232, For each neighbor node Calculate the raw attention score , where a is a learnable weight vector, and || denotes vector concatenation; S233. Normalize the attention scores of neighboring nodes using the softmax function: ; S234. Aggregate neighbor features based on normalized attention coefficients. , where σ is a nonlinear activation function; S24. Collect all updated node features to form the spatially enhanced feature matrix H′.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By modeling hyperspectral images as graph structures, this paper organically integrates graph attention networks and beta wavelet graph neural networks for the first time, achieving joint learning and enhancement of spatial context relationships and nonlinear spectral anomaly features within the same framework, overcoming the limitations of traditional methods that separate the processing of spatial and spectral information. 2. The graph attention mechanism adaptively focuses on the spatial neighborhood most relevant to anomaly detection, enhancing the model's discriminative power against complex local backgrounds. Beta wavelet transform is specifically designed to address the physical phenomenon of rightward shift of spectral energy caused by anomalies, accurately locating and amplifying anomalous signals in the frequency domain while effectively suppressing background interference. This results in higher detection accuracy and stronger robustness across various complex scenarios. 3. Compared to traditional spectral methods that require complex matrix decomposition, the beta wavelet used in this invention can be implemented through recursive computation, avoiding explicit eigenvalue decomposition and significantly reducing computational complexity. Furthermore, the entire method is based on a clear physical mechanism of spatial correlation and spectral right shift, making the model's decision-making process more interpretable. 4. By using the K-nearest neighbor algorithm instead of a fixed window to construct graph connections, non-locally similar pixels can be included, reducing the interference of noise and neighboring outliers. This enhances the model's adaptability to noise and anomaly spatial distribution, making it more suitable for real and complex hyperspectral remote sensing scenarios. Attached Figure Description

[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a flowchart of a hyperspectral anomaly detection method using a fusion graph attention and beta wavelet graph network provided in an embodiment of the present invention; Figure 2 A summary diagram of the algorithm runtime of the hyperspectral anomaly detection method using fused graph attention and beta wavelet graph networks provided in the embodiments of the present invention; Figure 3The ablation experiment results are shown in the figure for the hyperspectral anomaly detection method of fusion graph attention and beta wavelet graph network provided in the embodiments of the present invention. Detailed Implementation

[0018] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] Please see Figure 1 As can be seen, this embodiment provides a hyperspectral anomaly detection method that integrates graph attention and beta wavelet graph networks, including the following steps: S1, Graph structure modeling step: The three-dimensional hyperspectral image data cube is modeled as graph structure data, where each pixel is regarded as a node in the graph, the feature vector of the node is its spectral vector, and edge connections are established for the nodes based on spectral similarity to simultaneously embed the spatial and spectral information of the image; S2, Spatial feature enhancement step: The graph structure data is processed using a graph attention network, and the node neighbor information is aggregated through an adaptive attention mechanism to dynamically learn and enhance the spatial context features related to anomaly detection; S3, Spectral anomaly capture step: Based on the graph Laplacian matrix, a set of beta wavelet filters is constructed to form a graph wavelet transform group, and the node features enhanced in step S2 are subjected to multi-scale frequency domain filtering to specifically capture the right shift of spectral energy caused by abnormal targets; S4, Anomaly discrimination and output step: The filtered spectral features are fused and mapped, the probability of each pixel being an anomaly is calculated, and the final anomaly detection result image is generated.

[0020] Therefore, the overall implementation process of the hyperspectral anomaly detection method based on the fusion graph attention and beta wavelet graph network provided by this invention is as follows: Step 1: Hyperspectral Image Input and Preprocessing: Obtain the three-dimensional hyperspectral image data cube to be detected. Where H is the height, W is the width, and B is the number of spectral bands. Optional preprocessing includes bad band removal, noise smoothing, etc. This example uses a 120×120×189 sub-image from the AVIRIS-I dataset.

[0021] Step 2: Graph Structure Construction: The preprocessed data cube is reconstructed into a two-dimensional matrix. Where N = H × W is the total number of pixels, and each row Represents the spectral vector of a pixel. For each pixel is regarded as a node in the graph , and the node feature is its spectral vector . The K-nearest neighbor algorithm is used to construct edges: Calculate the Euclidean distance between all node pairs , for each node select the K nodes with the smallest distance as its neighbors, and establish undirected edges , forming the edge set ε. Based on this, construct the adjacency matrix , if , then , otherwise it is 0. Calculate the degree matrix D (diagonal matrix, ), and the Laplacian matrix L = D - A. In this embodiment, K = 30 is set.

[0022] Step 3: Spatial feature extraction based on graph attention: Input the graph and the node feature matrix X into a graph attention layer.

[0023] First, perform a shared linear transformation on the features of each node: , where is the learnable weight matrix, F is the dimension of the transformed features, and in this embodiment, F = 128 is set.

[0024] Secondly, calculate the attention coefficient between node i and its neighbor : where is the learnable vector, || represents the concatenation operation, and the negative slope of LeakyReLU is 0.2.

[0025] Then, use the softmax function to normalize the attention coefficient: Finally, perform weighted aggregation on the neighbor features and apply the activation function: where σ is the ELU activation function. The output obtains the spatially enhanced node feature matrix .

[0026] Step 4: Spectral anomaly detection based on the beta wavelet graph network: This step aims to capture the anomaly features in the spectral domain, especially the energy right shift phenomenon: First, construct the beta wavelet filter bank. Define the beta function . In the spectral graph domain, define the scaled beta kernel , and the corresponding graph wavelet filter is This can be achieved efficiently by recursively calculating powers of L, avoiding eigenvalue decomposition.

[0027] Set the wavelet order C=2. Construct a filter bank containing C+1=3 filters: W= .in, It is a low-pass filter. For bandpass filters, It is a high-pass filter, specifically designed to capture high-frequency abnormal signals.

[0028] Secondly, the features of each node output in step 300 Perform multi-scale filtering. The image is mapped to the spectral filtering space through a multilayer perceptron (MLP), and then processed by three wavelet filters respectively: MLP is a single-layer network with the same output dimension as the input dimension F.

[0029] Next, the filtering results from all scales are aggregated: In this embodiment, the AGG function is a concatenation operation, therefore The frequency domain feature matrix is ​​obtained. .

[0030] Step 5: Anomaly probability prediction and result generation.

[0031] The aggregated frequency domain features Input another MLP (a single-layer network in this example), map it to a scalar, and then use the Sigmoid activation function to obtain the probability that the pixel is an anomaly: The probability values ​​of all pixels constitute an anomaly probability map. Binarization can be performed by setting a threshold (such as 0.5) to obtain the final anomaly detection binary image.

[0032] Step 6: Model Training During the training phase, a dataset with real anomaly labels is required. The loss function is defined as weighted cross-entropy loss. in, ∈{0,1} represents the true label of pixel i (1 indicates anomaly, 0 indicates background), and w is a weighting factor used to alleviate the problem of imbalance between the number of anomaly and background samples. In this embodiment, we set... Using the Adam optimizer, with an initial learning rate of 0.001, we trained for 200 epochs.

[0033] Furthermore, to verify the effectiveness of each module and the parameter selection of this invention, the following experiments were conducted: 1. Analysis of parameter C (wavelet order): Performance was tested on the AVIRIS-I dataset with C values ​​ranging from 2 to 10. The results show that the AUC value reaches its peak (0.9937) when C=2 or 3. Increasing C enhances spectral locality but reduces spatial locality and increases computational cost. Therefore, C=2 is preferred.

[0034] 2. Analysis of parameter K (nearest neighbor number): Tests were conducted on the abu-urban-3 dataset with K=10, 20, 30, 40, and 50. Experiments showed that K=10 achieved the optimal AUC (0.9982). Too small a K value makes the dataset susceptible to noise, while too large a K value introduces irrelevant background information. For different datasets, the optimal K can be determined through a simple grid search.

[0035] 3. Ablation Experiments: BWGNN Only: The Graph Attention Network (GAN) was removed, and the original spectral features were directly input into the BWGNN module. On the AVIRIS-II dataset, the AUC decreased from 0.9986 to 0.9972. GAN Only: The BWGNN module was removed, and only the features extracted by the Graph Attention Network were used for linear classification. On the abu-beach-2 dataset, the AUC decreased significantly from 0.9961 to 0.9426.

[0036] Experimental results are as follows Figure 3 The results clearly demonstrate the necessity of the graph attention module and the beta wavelet module working together, as they are crucial for improving spatial discrimination and spectral anomaly detection capabilities, respectively.

[0037] And a performance comparison with traditional and deep learning methods: This invention (GAN-BWGNN) is compared with several mainstream methods, including traditional methods (GRX, LRX, CRD) and advanced deep learning methods (PDBSNet, BockNet, DirectNet, NL2Net), on seven public datasets (AVIRIS-I / II, abu-beach-2 / urban-3, Salinas-simulate, GrandIsle, Cri). The evaluation metrics are AUC and runtime.

[0038] Detection accuracy: This invention achieves the highest AUC values ​​on most datasets. For example, it reaches 0.9986 on AVIRIS-II, 0.9982 on abu-urban-3, and 0.9999 on Salinas-simulate, significantly outperforming the comparison methods.

[0039] Computational efficiency: such as Figure 2 As shown, the detection time of this invention is only 0.20-0.28 seconds on most datasets, which is 100-500 times faster than traditional methods (LRX, CRD) and 6-8 times faster than some deep learning models (such as NL2Net), demonstrating excellent real-time processing potential.

[0040] Specific use case examples: The trained GAN-BWGNN model is deployed in a hyperspectral remote sensing image processing system. The system receives raw data from airborne or spaceborne hyperspectral sensors, performs preprocessing such as radiometric calibration and atmospheric correction, and then calls upon the model of this invention for real-time anomaly detection. The anomaly probability map or binary map of the detection results can be overlaid on the original image, assisting operators in quickly locating potential man-made targets such as military facilities, illegal vessels, or natural anomalies such as pollutant leaks and mineralized alteration zones. The model's efficiency enables it to process large-scale inspection data.

[0041] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0042] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0043] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0044] These computer program instructions can also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0045] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A hyperspectral anomaly detection method fusing graph attention and beta wavelet graph networks, characterized in that, Includes the following steps: S1. Graph structure modeling: The three-dimensional hyperspectral image data cube is modeled as graph structure data, where each pixel is regarded as a node in the graph, the feature vector of the node is its spectral vector, and edge connections are established for the nodes based on spectral similarity to simultaneously embed the spatial and spectral information of the image. S2. Spatial Feature Enhancement: The graph structure data is processed using a graph attention network. Node neighbor information is aggregated through an adaptive attention mechanism to dynamically learn and enhance spatial context features related to anomaly detection. S3. Spectral Anomaly Capture: Based on the graph Laplacian matrix, a set of beta wavelet filters is constructed to form a graph wavelet transform group. Multi-scale frequency domain filtering is performed on the node features enhanced in step S2 to specifically capture the right shift of spectral energy caused by anomalous targets. S4. Anomaly Detection and Output: The filtered spectral features are fused and mapped to calculate the probability of each pixel being an anomaly and generate the final anomaly detection result image.

2. The hyperspectral anomaly detection method according to claim 1, which integrates graph attention and beta wavelet graph networks, is characterized in that... In step S1, the K-nearest neighbor algorithm is used to establish edge connections for each node, that is, to calculate the spectral distance between all pairs of nodes and to select the K other nodes with the closest spectral distance to each node to establish connection edges.

3. The hyperspectral anomaly detection method according to claim 2, which integrates graph attention and beta wavelet graph networks, is characterized in that... The operation of the graph attention network in step S2 includes: performing a linear transformation on the features of each node; calculating the attention weights between the node and each of its neighboring nodes; performing a weighted summation of the transformed features of the neighboring nodes based on the attention weights; and using an activation function to obtain the updated spatial augmentation features of the node.

4. The hyperspectral anomaly detection method according to claim 1, which integrates graph attention and beta wavelet graph networks, is characterized in that... The beta wavelet filter described in step S3 is constructed based on a parameterized beta function, which is defined as follows: ,in Let be the shape parameter, and satisfy . C is the wavelet order.

5. The hyperspectral anomaly detection method according to claim 4, which integrates graph attention and beta wavelet graph networks, is characterized in that... The beta wavelet filter bank W consists of C+1 filters of different scales, denoted as... ,in One is a low-pass filter, and the rest are band-pass or high-pass filters.

6. The hyperspectral anomaly detection method according to claim 1, which integrates graph attention and beta wavelet graph networks, is characterized in that... In step S3, the right shift of spectral energy specifically means that when there is an abnormal target in the image, the energy distribution of the signal in the Fourier transform domain will shift to the high-frequency part, and the beta wavelet filter is configured to selectively enhance the high-frequency energy.

7. The hyperspectral anomaly detection method according to claim 1, which integrates graph attention and beta wavelet graph networks, is characterized in that... Step S4 specifically includes: aggregating all the wavelet-scale filtered features from step S3; mapping the aggregated features to a scalar using a multilayer perceptron; and applying the Sigmoid activation function to convert the scalar into a value representing the probability of anomalies.

8. The hyperspectral anomaly detection method according to claim 1, characterized in that, The method performs end-to-end training by minimizing a weighted cross-entropy loss function, which is defined as: Where, 𝑁 represents the total number of pixels, 𝑦 𝑖 For the true label of pixel i, 𝑝 𝑖 This represents the probability that pixel i is an anomalous, as predicted by the model.

9. The hyperspectral anomaly detection method according to claim 2, which integrates graph attention and beta wavelet graph networks, is characterized in that... Step S1 specifically includes: S11, Input 3D hyperspectral data cube And the number of nearest neighbors K; S12. Reconstruct the data cube into a two-dimensional matrix. ,in Each row represents the spectral feature vector of one pixel; S13. Define each pixel as a graph node. This forms the node set V; S14. For each node Calculate the Euclidean distance of its spectral characteristics with all other nodes, select the K nodes with the smallest distance as its neighbors, and establish an edge. This forms the edge set ε; S15. Construct the adjacency matrix A and degree matrix D based on the edge set, and calculate the Laplacian matrix L=DA of the graph accordingly. S16, Output Diagram Structure And its Laplace matrix L.

10. The hyperspectral anomaly detection method according to claim 3, which integrates graph attention and beta wavelet graph networks, is characterized in that... Step S2 specifically includes: S21, Input Graph Structure and their adjacency relationships; S22. Initialize node characteristics ; S23. For each node Perform the following operations: S231. Apply a learnable weight matrix W to perform a linear transformation: ; S232, For each neighbor node Calculate the raw attention score , where a is a learnable weight vector, and || denotes vector concatenation; S233. Normalize the attention scores of neighboring nodes using the softmax function: ; S234. Aggregate neighbor features based on normalized attention coefficients. , where σ is a nonlinear activation function; S24. Collect all updated node features to form the spatially enhanced feature matrix H′.

Images