Hyperspectral image classification method based on graph convolutional neural network and state space model
By combining graph convolutional neural networks and state-space models, a multi-hop graph-graph convolutional-state-space model MGMN is constructed, which solves the problem of balancing accuracy and efficiency in hyperspectral image classification and achieves efficient hyperspectral image classification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIDIAN UNIV
- Filing Date
- 2026-03-31
- Publication Date
- 2026-06-30
AI Technical Summary
Existing hyperspectral image classification methods suffer from low computational efficiency while maintaining accuracy. In particular, traditional methods that ignore spatial information or Transformer-based methods have high computational complexity and cannot effectively improve detection speed.
By combining graph convolutional neural networks with state-space models, spatial features are extracted through a multi-hop graph structure, and a multi-hop graph-graph convolution-state-space model MGMN is constructed to integrate spatial and spectral information and improve computational efficiency.
While maintaining the accuracy of hyperspectral image classification, it significantly improves computational efficiency, achieving higher detection speed and better classification results.
Smart Images
Figure CN122313271A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of image processing technology, and further relates to hyperspectral image classification technology, specifically a hyperspectral image classification method based on graph convolutional neural networks and state-space models, which can be used in environmental monitoring, military and medical fields. Background Technology
[0002] Hyperspectral image classification is one of the important tasks in hyperspectral image processing. Its purpose is to accurately classify different land cover at the pixel level to further guide production practices. This technology is widely used in military, civilian, environmental monitoring, and agricultural fields. In recent years, among existing popular hyperspectral image classification methods, deep learning-based methods have attracted much attention from researchers and achieved satisfactory performance.
[0003] Currently, traditional KNN algorithms are being introduced into hyperspectral image classification to classify images based on spectral information, while random forests are being used to extract spectral information for classification. Both approaches focus solely on spectral information while neglecting spatial information, resulting in poor classification performance due to spectral variability. In contrast, existing literature on 2DCNN [X. Zhu, D. Cheng, Z. Zhang, S. Lin, and J. Dai, “An Empirical Study of Spatial Attention Mechanisms in Deep Networks,” in 2019 IEEE / CVF International Conference on Computer Vision (ICCV), Seoul, Korea (South), Oct. 2019] combines spatial and spectral information for hyperspectral classification, achieving better results. HybridSN [SK Roy, G.Krishna, SR Dubey, and BB Chaudhuri, “HybridSN: Exploring 3-D–2-D CNNFeature Hierarchy for Hyper-spectral Image Classification,” IEEE Geoscience and Remote Sensing Letters, pp. 277–281, Feb. 2020] and IRTS-3DCNN [C.-I. Chang, KY Ma, C.-C. Liang, Y.-M. Kuo, S. Chen, and S. Zhong, “Iterative RandomTraining Sampling Spectral Spatial Classification for Hyperspectral Images,” IEEE Journal of Selected Topics in Applied Earth Observations and RemoteSens-ing, vol. 13, pp. 3986–4007, Jan. In 2020, 3DCNN was used to classify hyperspectral images by combining spatial and spectral data, which significantly improved the classification effect. However, 3DCNN has a limited receptive field and cannot focus on global spectral details, so there is still room for improvement in its classification performance.
[0004] The CasRNN [R. Hang, Q. Liu, D. Hong, and P. Ghamisi, “Cascaded Recurrent Neural Networks for Hyperspectral Image Classification,” IEEE Transactions on Geoscience and Remote Sensing, pp. 5384–5394, Aug. 2019] innovatively uses RNNs for spectral classification and achieves good results. SpectralFormer [D. Hong et al., “SpectralFormer: Rethinking Hyperspectral Image Classification with transformers,” IEEE Transactions on Geoscience and Remote Sensing, pp. 1–15, Jan. 2022] and GraphGST [M. Jiang et al., "GraphGST: Graph Generative Structure-Aware transformer for Hyperspectral Image Classification," in IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1-16, 2024, Artno. 5504016] both use Transformers for hyperspectral image classification and have achieved state-of-the-art results in this field. However, Transformers suffer from high computational complexity, which limits their computational efficiency and restricts their application scenarios.
[0005] In summary, traditional hyperspectral classification models such as KNN and SVM typically focus only on spectral information while ignoring spatial information, thus limiting the improvement of classification accuracy due to spatial variability. Transformer-based hyperspectral image classification methods generally have high computational complexity, limiting their detection speed and lacking the ability to perceive spectral details. CNN-based hyperspectral image classification methods, due to the limitation of their receptive field, cannot perceive global spatial information. Therefore, in hyperspectral image classification tasks, how to improve computational efficiency while ensuring classification accuracy is a problem worth exploring. Summary of the Invention
[0006] The purpose of this invention is to address the shortcomings of existing technologies by proposing a hyperspectral image classification method based on graph convolutional neural networks and a state-space model, thus solving the problems of low detection efficiency or limited classification accuracy in existing hyperspectral image classification models. First, spatial features are extracted from a multi-hop image obtained through superpixel image segmentation. Then, the obtained spatial features are fused with the original hyperspectral image using the multi-hop image-graph convolutional-state-space model designed in this invention to obtain the final hyperspectral image classification result. This invention overcomes the deficiencies of existing technologies, effectively improving the computational efficiency of hyperspectral image classification models while maintaining classification accuracy.
[0007] The specific steps for achieving the above objectives according to the present invention include the following:
[0008] (1) Input the original hyperspectral image Dimensionality reduction is then performed, followed by superpixel segmentation of the dimensionality-reduced image to generate a superpixel matrix Q. Flattened into a two-dimensional hyperspectral image The image is then fused with the segmented image to obtain a hyperspectral image H with a graph structure.
[0009] (2) Construct a multi-hop graph generation mechanism, generate the basic weight matrix W based on the superpixel matrix Q, and obtain the weight matrix W through the depth-first search (DFS) algorithm. Skip graph adjacency matrix ;
[0010] (3) A graph convolution-state space module GM for extracting hyperspectral image features is constructed by combining the graph convolution model GCN with the state space model SS2D. The implementation process of this module is as follows:
[0011] (3.1) Encoded hyperspectral image and Skip Adjacency Matrix Input a graph convolutional model GCN, output a graph convolutional model GCN. Adding them together gives ;
[0012] (3.2) will Input state-space model SS2D, output and Adding them together gives ;
[0013] (3.3) will The input is processed by a self-attention mechanism, passed through a linear layer, and then added to itself to obtain the final output. ;
[0014] (4) Construct a multi-hop graph-graph convolution-state space model MGMN based on the graph convolution-state space module GM:
[0015] (4.1) Obtain the multi-hop adjacency matrix according to step (2.2). , and ;
[0016] (4.2) will Encode with Q to obtain initial features ;
[0017] (4.3) Stack three GM modules to extract features and obtain the final features. :
[0018] ,
[0019] ,
[0020] ,
[0021] (4.4) Final features The data is fed into a multilayer perceptron (MLP) classifier, which outputs the classification result.
[0022] (5) The MGMN model is trained using the cross-entropy loss function to obtain a trained classification model, which is then used to classify hyperspectral images.
[0023] Compared with the prior art, the present invention has the following advantages:
[0024] First, this invention is the first to combine the State Space Model (SSM) with the Graph Convolutional Network (GCN) in the hyperspectral image classification task, and constructs a hyperspectral image classification network MGMN based on the graph convolutional neural network and the state space model. It has the advantages of both the graph convolutional network and the state space model, and can fuse spatial and spectral information, thereby significantly improving computational efficiency while ensuring classification accuracy.
[0025] Secondly, since the invention uses a multi-hop graph approach for spatial feature extraction, it effectively overcomes the limitation that graph convolution and state space models can only extract local features.
[0026] Third, this invention proposes a graph convolution-state space model, Graph-Mamba, to extract hyperspectral image features. This model can be further used in all networks that require graph convolution to enhance the long sequence information processing capability of graph convolution, thereby enabling the fusion of features extracted by the multi-hop graph mechanism with spectral information. Attached Figure Description
[0027] Figure 1 This is a flowchart illustrating the overall implementation of the method of the present invention;
[0028] Figure 2This is a schematic diagram of the convolutional-state-space model architecture in this invention;
[0029] Figure 3 The diagram shows the classification results of the present invention and existing methods on the Pavia University dataset. The labels (a)-(k) are the classification results of existing algorithms ground truth, KNN, RF, 1DCNN, 2DCNN, HybridSN, IRTS-3DCNN, CasRNN, ViT, SpectralFormer and GraphGST, respectively. (l) is the classification result of the MGMN algorithm proposed in the present invention.
[0030] Figure 4 The diagram shows the classification results of the present invention and existing methods on the Salinas Valley dataset. The labels (a)-(k) are the classification results of existing algorithms ground truth, KNN, RF, 1DCNN, 2DCNN, HybridSN, IRTS-3DCNN, CasRNN, ViT, SpectralFormer and GraphGST, respectively. (l) is the classification result of the MGMN algorithm proposed in the present invention. Detailed Implementation
[0031] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0032] Example 1: Refer to Figure 1 This invention proposes a hyperspectral image classification method based on graph convolutional neural networks and state-space models, specifically including the following steps:
[0033] Step 1) Input the raw hyperspectral image Dimensionality reduction is then performed, followed by superpixel segmentation of the dimensionality-reduced image to generate a superpixel matrix Q. Flattened into a two-dimensional hyperspectral image The image is then fused with the segmented image to obtain a hyperspectral image H with a graph structure. The process includes:
[0034] (1.1) Input the original hyperspectral image PCA dimensionality reduction is then performed on it to reduce redundant information and lower computational complexity;
[0035] (1.2) Perform superpixel segmentation on the dimension-reduced image to obtain the superpixel matrix. Where Z represents the number of superpixel blocks, and h and w represent the height and width of the hyperspectral image, respectively; in this embodiment, the superpixel segmentation of the dimension-reduced image is implemented using the SLIC algorithm, which is used to reduce the data dimension while preserving the image edge information.
[0036] (1.3) The original hyperspectral image Flattening yields a two-dimensional hyperspectral image. Where B represents the band number; then... Multiplying by the transpose of the normalized superpixel matrix yields a graph-structured hyperspectral image. .
[0037] In this embodiment, a matrix is obtained by superpixel segmentation of the hyperspectral image in this step. At the same time, hyperspectral images After flattening, you get It can be seen that Q and There is a one-to-one correspondence in the first dimension. The values of the column vector of Q represent their corresponding... The column vector (spectral) superpixel blocks are obtained. This results in a hyperspectral image H constructed as a graph structure, which serves as the basis for extracting hyperspectral spatial and spectral information in this invention; the specific calculation formula is as follows:
[0038] , <1> ,
[0039] in, This represents the normalized Q.
[0040] Step 2) Construct a multi-hop graph generation mechanism, generate the basic weight matrix W based on the superpixel matrix Q, and obtain the weight matrix W using the depth-first search (DFS) algorithm. Skip graph adjacency matrix .
[0041] In this embodiment, the aforementioned basic weight matrix W is specifically calculated using a Gaussian kernel function, as shown in the following formula:
[0042] ; <2> ;
[0043] in, and This represents the eigenvectors of any two superpixel nodes in the superpixel matrix Q. Represents a superpixel node and Similarity weights between them; Represents a superpixel node and The Euclidean distance between two vectors decreases monotonically as the distance between them increases; This represents the bandwidth parameter of the Gaussian kernel; express The set of neighboring nodes.
[0044] The above Skip graph adjacency matrix Obtained through the following methods:
[0045] For each superpixel node Use the Depth-First Search (DFS) algorithm to find and record the distance from the selected center node. All paths to jump, with the endpoints of the paths represented as... ; through computing nodes and The average of the path weights is obtained. Skip Adjacency Matrix :
[0046] ; <3> ;
[0047] in, This represents the intermediate nodes of the path. Based on the above method, the present invention can generate a series of k-hop graph adjacency matrices with weights A1, A2, ...; in this embodiment, the first three hop graphs are used as an example, namely A1, A2, and A3.
[0048] Step 3) A graph convolution-state space module (GM) for extracting hyperspectral image features is constructed by combining the graph convolution model GCN and the state space model SS2D. This module combines the local graph structure modeling capability of GCN, the long-range dependency handling capability of SS2D, and the global association capture capability of self-attention; the state space model SS2D is used to efficiently handle long-range dependencies between graph nodes. The specific implementation process of the above graph convolution-state space module GM in this embodiment is as follows:
[0049] (3.1) Encoded hyperspectral image and Skip Adjacency Matrix Input a graph convolutional model GCN, output a graph convolutional model GCN. Adding them together gives ;
[0050] (3.2) will Input state-space model SS2D, output and Adding them together gives ;
[0051] (3.3) will The input is processed by a self-attention mechanism, passed through a linear layer, and then added to itself to obtain the final output. .
[0052] Reference Figure 2In this embodiment, a graph convolutional-state-space module (GM) is designed in this step to extract features from the hyperspectral image. GCN excels at handling graph-structured data and can effectively capture the complex relationships between nodes in the hop graph after superpixel segmentation, while SSM excels at handling long sequence data. Graph-Mamba accepts two inputs: the output of the feedforward network and a multi-hop graph adjacency matrix. For the k-th GM module, the input is the encoded hyperspectral image information. Adjacency matrix of the first multi-hop graph .Will and The input is fed into a GCN for graph convolution, and then the result is combined with... Adding them together gives Subsequently, After entering SS2D, then... Add them together to get Then After being input into the self-attention mechanism proposed in Chapter 3, the input passes through a linear layer and then interacts with... Add them together to get the final output. :
[0053] , <4> ;
[0054] , <5> ;
[0055] , <6> ;
[0056] ; <7> ;
[0057] Step 4) Construct the Multi-hop Graph Mamba Network (MGMN) based on the Graph Convolution-State Space Module (GM):
[0058] (4.1) Obtain the multi-hop adjacency matrix according to step (2.2). , and ;
[0059] (4.2) Using the encoding function to... Encode with Q to obtain initial features :
[0060] , <8> ;
[0061] in, This is the normalized result of the superpixel matrix Q. Indicates the transpose operation; This is the flattened vector of the original hyperspectral image.
[0062] (4.3) Stack three GM modules to extract features and obtain the final features. :
[0063] , <9> ;
[0064] , <10> ;
[0065] , <11> ;
[0066] In this embodiment, the three GM modules mentioned above use residual connections to add the output of each GM module to the features of the previous layer, so as to ensure efficient information transmission and avoid gradient vanishing.
[0067] (4.4) Final features The data is fed into a multilayer perceptron (MLP) classifier, which outputs the classification result.
[0068] Step 5) Train the MGMN model using the cross-entropy loss function to obtain the trained classification model, and use the model to complete the hyperspectral image classification.
[0069] In this embodiment, the cross-entropy loss mentioned above is used to minimize the difference between the model's predicted probability and the true label, thereby optimizing classification accuracy. Specifically, it is expressed as follows:
[0070] , <12> ;
[0071] in, The one-hot vector of the real label; Predict probabilities for the model; This indicates the number of samples in the training set.
[0072] The effects of the present invention will be further explained below with reference to experiments.
[0073] 1. Experimental conditions:
[0074] The hardware configuration for implementing this invention is as follows: an E5-2650V4 CPU (2.2 GHz 12x2 cores), 512GB of memory, and a workstation with 8 NVIDIA RTX Titan graphics cards were used for the experiment. The software platform configuration is as follows: Ubuntu 16.04, Python 3.7, and PyTorch 1.7.
[0075] 2. Experiment Content:
[0076] The present invention was qualitatively and quantitatively compared with ten popular algorithms on two public datasets for hyperspectral image classification. The publicly available hyperspectral anomaly detection datasets used in the experiments included those from Pavia University and Salinas Valley. The ten popular algorithms compared were: KNN [W. Li, Q. Du, F. Zhang, and W. Hu, “Collaborative representation based k-nearest neighbor classifier for hyperspectral imaging,” in 2014 6th Workshop on Hyperspectral Image and Signal Processing: Evolution in Remote Sensing (WHISPERS), Lausanne, Switzerland, Jun. 2014.], RF [S. Amini, S. Homayouni, A. Safari, and AADarvishsefat, “Object-based classification of hyperspectral data using RandomForest algorithm,” Geo-spatial Information Science, pp. 127–138, Apr. 2018], 1DCNN, and 2DCNN [X. Zhu, D. Cheng, Z. Zhang, S. Lin, and J. Dai, “An Empirical Study of Spatial Attention Mechanisms in Deep Networks,” in 2019 IEEE / CVFInternational Conference on Computer Vision (ICCV), Seoul, Korea (South),Oct. 2019], HybridSN [SK Roy, G. Krishna, SR Dubey, and BBChaudhuri,“HybridSN: Exploring 3-D–2-D CNN Feature Hierarchy for Hyper-spectral ImageClassification,” IEEE Geoscience and Remote Sensing Letters, pp. 277–281,Feb. 2020]、IRTS-3DCNN[C.-I. Chang, K. Y. Ma, C.-C. Liang, Y.-M. Kuo, S. Chen,and S. Zhong, “Iterative Random Training Sampling Spectral SpatialClassification for Hyperspectral Images,” IEEE Journal of Selected Topics inApplied Earth Observations and Remote Sens-ing, vol. 13, pp. 3986–4007, Jan.2020]、CasRNN[R. Hang, Q. Liu, D. Hong, and P. Ghamisi, “Cascaded RecurrentNeural Networks for Hyperspectral Image Classification,” IEEE Transactions onGeoscience and Remote Sensing, pp. 5384–5394, Aug. 2019]、VIT[J. Li, K. Zheng,W. Liu, Z. Li, H. Yu and L. Ni, "Model-Guided Coarse-to-Fine Fusion Networkfor Unsupervised Hyperspec-tral Image Super-Resolution," in IEEE Geoscienceand Remote Sensing Letters, vol. 20, pp. 1-5, 2023, Art no. 5508605]、SpectralFormer[D. Hong et al., "SpectralFormer: Rethinking HyperspectralImage Classification with transformers," IEEE Transactions on Geoscience and Remote Sensing, pp. 1–15, Jan. 2022] and GraphGST [M. Jiang et al., "GraphGST: Graph Generative Structure-Aware transformer for Hyperspectral ImageClassification," in IEEE Transactions on Geoscience and Remote Sensing, vol.62, pp. 1-16, 2024, Art no. 5504016]. .
[0077] 3. Simulation Results and Analysis:
[0078] The classification results of this invention and existing methods on the Pavia University dataset are as follows: Figure 3 As shown, (a)-(k) are the classification results of existing algorithms ground truth, KNN, RF, 1DCNN, 2DCNN, HybridSN, IRTS-3DCNN, CasRNN, ViT, SpectralFormer and GraphGST respectively, and (l) is the classification result of the MGMN algorithm proposed in this invention;
[0079] The classification results of this invention and existing methods on the SalinasValley dataset are as follows: Figure 4 As shown, (a)-(k) are the classification results of existing algorithms ground truth, KNN, RF, 1DCNN, 2DCNN, HybridSN, IRTS-3DCNN, CasRNN, ViT, SpectralFormer and GraphGST respectively, and (l) is the classification result of the MGMN algorithm proposed in this invention.
[0080] The quantitative comparison results of this invention with ten popular comparison algorithms on two datasets are shown in Tables 1 and 2. In each table, the first column represents the pixel category, and the remaining columns represent the classification accuracy of the corresponding method for a specific category. The evaluation metrics used are Overall Accuracy (OA), Average Accuracy (AA), and Kappa Coefficient (κ). These three metrics evaluate the accuracy of the algorithm from different perspectives; a higher value indicates higher detection accuracy. As can be seen from Tables 1 and 2, this invention exhibits the best detection performance.
[0081] Table 1. Comparison of classification results on the Pavia University dataset.
[0082]
[0083] Table 2. Comparison of classification results on the Salinas dataset.
[0084]
[0085] This invention innovatively combines graph convolutional neural networks with a state-space model, fully mining the spatial correlation information of hyperspectral images through a multi-hop graph structure. The unique "multi-hop graph-graph convolution-state-space model" exhibits stronger robustness and superior classification performance compared to existing technologies, demonstrating outstanding technological innovation and advancement. It effectively solves the pain point of "difficulty in balancing accuracy and efficiency" in existing technologies. Experiments show that this method significantly improves computational efficiency while maintaining high classification accuracy (up to 94.36% overall accuracy, up to 96.71% average accuracy, and up to 0.9374 Kappa coefficient). It can be widely applied in environmental monitoring, national defense security, and medical health fields, possessing significant application value in scenarios such as land cover classification, military target recognition, and medical image-assisted diagnosis, generating substantial social and economic benefits. The simulation analysis above proves the correctness and effectiveness of the proposed method.
[0086] The parts of this invention not described in detail are common knowledge to those skilled in the art. The above description is only a preferred embodiment of the present invention and is not intended to limit the invention. Obviously, those skilled in the art, after understanding the content and principles of the present invention, may make various modifications and changes in form and detail without departing from the principles and structure of the present invention. However, these modifications and changes based on the concept of the present invention are still within the protection scope of the claims of the present invention.
Claims
1. A hyperspectral image classification method based on graph convolutional neural networks and a state-space model, characterized in that, Includes the following steps: (1) Input the original hyperspectral image Dimensionality reduction is then performed, followed by superpixel segmentation of the dimensionality-reduced image to generate a superpixel matrix Q. Flattened into a two-dimensional hyperspectral image The image is then fused with the segmented image to obtain a hyperspectral image H with a graph structure. (2) Construct a multi-hop graph generation mechanism, generate the basic weight matrix W based on the superpixel matrix Q, and obtain the weight matrix W through the depth-first search (DFS) algorithm. Skip graph adjacency matrix ; (3) A graph convolution-state space module GM for extracting hyperspectral image features is constructed by combining the graph convolution model GCN with the state space model SS2D. The implementation process of this module is as follows: (3.1) Encoded hyperspectral image and Skip Adjacency Matrix Input a graph convolutional model GCN, output a graph convolutional model GCN. Adding them together gives ; (3.2) will Input state-space model SS2D, output and Adding them together gives ; (3.3) will The input is processed by a self-attention mechanism, passed through a linear layer, and then added to itself to obtain the final output. ; (4) Construct a multi-hop graph-graph convolution-state space model MGMN based on the graph convolution-state space module GM: (4.1) Obtain the multi-hop adjacency matrix according to step (2.2). , and ; (4.2) will Encode with Q to obtain initial features ; (4.3) Stack three GM modules to extract features and obtain the final features. : ， ， ， (4.4) Final features The data is fed into a multilayer perceptron (MLP) classifier, which outputs the classification result. (5) The MGMN model is trained using the cross-entropy loss function to obtain a trained classification model, which is then used to classify hyperspectral images.
2. The method according to claim 1, characterized in that: The hyperspectral image H of the graph structure described in step (1) is obtained specifically according to the following steps: (1.1) Input the original hyperspectral image PCA dimensionality reduction is then performed on it to reduce redundant information and lower computational complexity; (1.2) Perform superpixel segmentation on the dimension-reduced image to obtain the superpixel matrix. , where Z represents the number of superpixel blocks, and h and w represent the height and width of the hyperspectral image, respectively; (1.3) The original hyperspectral image Flattening yields a two-dimensional hyperspectral image. Where B represents the band number; then... Multiplying by the transpose of the normalized superpixel matrix yields a graph-structured hyperspectral image. .
3. The method according to claim 2, characterized in that: The superpixel segmentation of the dimensionality-reduced image described in step (1.2) is implemented using the SLIC algorithm, which is used to reduce the data dimensionality while preserving the image edge information.
4. The method according to claim 1, characterized in that: The basic weight matrix W mentioned in step (2) is specifically calculated using the Gaussian kernel function, as shown in the following formula: ； in, and This represents the eigenvectors of any two superpixel nodes in the superpixel matrix Q. Represents a superpixel node and Similarity weights between them; Represents a superpixel node and The Euclidean distance between them; This represents the bandwidth parameter of the Gaussian kernel; express The set of neighboring nodes.
5. The method according to claim 4, characterized in that: The steps described in step (2) Skip graph adjacency matrix It is obtained through the following methods: For each superpixel node Use the Depth-First Search (DFS) algorithm to find and record the distance from the selected center node. All paths to jump, with the endpoints of the paths represented as... ; through computing nodes and The average of the path weights is obtained. Skip Adjacency Matrix : ， in, Indicates the intermediate nodes of the path.
6. The method according to claim 1, characterized in that: The graph convolution-state space module DM mentioned in step (3) combines the local graph structure modeling capability of GCN, the long-range dependency processing capability of SS2D, and the global association capture capability of self-attention; the state space model SS2D is used to efficiently process the long-range dependencies between graph nodes.
7. The method according to claim 1, characterized in that: The initial features described in step (4.2) Generated by the encoding function: ， in, This is the normalized result of the superpixel matrix Q. Indicates the transpose operation; This is the flattened vector of the original hyperspectral image.
8. The method according to claim 1, characterized in that: The three GM modules mentioned in step (4.3) use residual connections to add the output of each GM module to the features of the previous layer to ensure efficient information transmission and avoid gradient vanishing.
9. The method according to claim 1, characterized in that: The cross-entropy loss mentioned in step (5) is used to minimize the difference between the model's predicted probability and the true label, thereby optimizing the classification accuracy. Specifically, it is expressed as follows: ， in, The one-hot vector of the real label; Predict probabilities for the model; This indicates the number of samples in the training set.