Hyperspectral image classification method and device based on subgraph adaptive neural network

Abstract

The invention discloses a hyperspectral image classification method and device based on a subgraph adaptive neural network. The method comprises the following steps: S1, constructing a paired probability graph structure based on an SSAPGCN method; s2, reading a feature index and generating a feature matrix, performing PCA dimension reduction on the input data, and performing data preprocessing; s3, a global graph structure S is obtained based on dimension reduction preprocessing, a Metis graph segmentation algorithm is called for graph segmentation, and the global probability graph structure is divided into a plurality of sub-graphs; s4, calculating an intra-class distance intra (C) as a self-adaptive feedback threshold value, determining an optimal image convolution layer number k according to sub-image spectral difference, and avoiding over-smoothing or under-smoothing; and S5, training and testing the data by using the optimal image convolution layer number k to complete hyperspectral image classification. The method is mainly used for processing large image classification, the calculation cost can be reduced, spectrum and spatial features are fully mined, and the classification accuracy and efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of hyperspectral image classification, and more particularly to a hyperspectral image classification method and device based on a subgraph adaptive neural network. Background Art

[0002] Hyperspectral image classification is an important research direction in the field of remote sensing. Traditional hyperspectral image classification methods are mainly based on spectral features, such as the maximum likelihood classification method, support vector machines, etc. These methods have certain effects when dealing with simple scenarios and small amounts of data. However, with the continuous increase in the amount of hyperspectral image data and the improvement of scene complexity, traditional methods gradually expose some limitations. On the one hand, traditional methods often ignore the rich spatial information in hyperspectral images and rely only on spectral features for classification, which easily leads to inaccurate classification results, especially when dealing with ground objects with similar spectral features but different spatial distributions, the classification effect is poor. On the other hand, traditional methods have low computational efficiency when dealing with large-scale hyperspectral image data and are difficult to meet the requirements of practical applications.

[0003] In recent years, deep learning technology has achieved remarkable results in the field of image classification. Some methods based on Convolutional Neural Networks (hereinafter referred to as CNN) have been applied to hyperspectral image classification. These methods can automatically extract the features of hyperspectral images and improve the classification accuracy to a certain extent. However, CNN is mainly based on local convolution operations, insufficiently utilizes the global structural information of images, and has certain limitations when dealing with the complex topological structures of hyperspectral images.

[0004] To overcome the above problems, researchers have begun to explore hyperspectral image classification methods based on graph Neural Networks (hereinafter referred to as GNN). GNN can effectively process graph-structured data, represent pixel points in hyperspectral images as nodes in a graph, and represent the relationships between pixels as edges, so that spectral information and spatial information can be fully utilized for classification. However, existing GNN methods still face problems such as high computational complexity and long training time when dealing with large-scale hyperspectral images.

[0005] Therefore, the researchers began to focus on Subgraph Neural Networks (hereinafter referred to as SGNN). However, although the method based on SGNN can retain the structural information between nodes and reduce the loss of structural information to a certain extent, the intra-class spectral variability and inter-class spectral similarity existing in Hyperspectral Image (HSI) lead to the imbalance of the topological structure within the subgraph. In the message passing mechanism, it is difficult to effectively smooth the feature information within the subgraph by adopting the aggregation strategy of fixing the K-order neighbor information for the subgraph with an unbalanced graph structure, which may lead to over-smoothing or under-smoothing phenomena. Summary of the Invention

[0006] A hyperspectral image classification method and device based on a subgraph adaptive neural network provided by the present invention can effectively reduce the calculation cost, fully exploit the spectral and spatial features of the image, and significantly improve the accuracy and efficiency of classification, and can at least solve one of the above technical problems.

[0007] To solve the above technical problems, the present invention adopts the following technical solutions:

[0008] The hyperspectral image classification method based on a subgraph adaptive neural network includes the following steps:

[0009] S1. Construct a pairwise probability graph structure based on the SSAPGCN method to effectively integrate the spatial information and spectral information of the hyperspectral image;

[0010] S2. Read the feature indexes of each spectral band in the original spectral image and generate a feature matrix containing spectral information, perform PCA dimensionality reduction on the input feature matrix data, and perform data preprocessing to remove redundant information;

[0011] S3. Based on the globally reduced preprocessed graph structure S, call the Metis graph segmentation algorithm for graph segmentation, and divide the globally obtained probability graph structure into multiple local subgraphs to mine local features;

[0012] S4. Calculate the intra-class distance intra(C) as an adaptive feedback threshold, and determine the optimal number of graph convolution layers k according to the spectral difference of the subgraph to avoid over-smoothing or under-smoothing;

[0013] S5. Use the optimal number of graph convolution layers k to train and test the globally obtained probability graph structure data to complete hyperspectral image classification.

[0014] Further, the S1 further includes:

[0015] S11. Input the spectral feature matrix and spatial coordinate matrix of the hyperspectral image:

[0016] X ∈ R N×d

[0017] L ∈ R N×2

[0018] Among them, in this formula, X is the spectral feature matrix, L is the spatial coordinate matrix, N is the number of pixels, and d is the spectral dimension;

[0019] S12. To establish a graph topological relationship that takes into account spectral similarity and spatial proximity, a probabilistic graph structure A ∈ R is constructed by minimizing the following objective function N×N :

[0020]

[0021] Among them, in this formula:

[0022] a ij is the similarity between pixels i and j, which is jointly determined by the spectral difference ‖x i - x j ‖ and the spatial distance ‖l i - l j ‖;

[0023] P ∈ R N×N is the guidance matrix based on expert prior;

[0024] δ is the balance parameter of spectral information and spatial information.

[0025] Furthermore, in S2, to remove redundant information, PCA dimensionality reduction is performed:

[0026] X PCA = X · V top

[0027] Among them, in this formula, X PCA is the dimensionality-reduced feature matrix, and V top is the d x 60 eigenvector matrix;

[0028] Retain the first 60 principal component quantities to obtain the dimensionality-reduced feature matrix X PCA ∈ R n×60 .

[0029] Furthermore, in S3, the divided multiple subgraphs are respectively:

[0030] A = [A1, A2,..., A m , X = [X1, X2,..., X m

[0031] Among them, in this formula, each subgraph A i ∈ R n×n maintains the local topological structure and contains the corresponding node features Xi ∈R n ×60 , realizing data parallel processing.

[0032] Furthermore, in S4, for the heterogeneity of the subgraphs, for each subgraph i, the following dynamic mechanism is used to iteratively determine the optimal number of graph convolution layers k:

[0033] S41. Initialize the feature matrix

[0034] S42. The node feature update rule from the original feature of the 0th layer to the (k + 1)th layer is:

[0035]

[0036] Among them, in this formula:

[0037] is the feature vector of node i at the lth layer;

[0038] a ij is the adjacency matrix A i is the connection weight between nodes i and j in it;

[0039] n is the total number of nodes in the current subgraph;

[0040] S43. Introduce the degree matrix residual mechanism, and embed the degree information of the nodes as residuals into the feature representation. The formula is as follows:

[0041]

[0042] Among them, in this formula:

[0043] De d ∈R n×n is the subgraph degree matrix, and the diagonal elements are the node degrees;

[0044] β is the balance coefficient, used to control the weights of the original feature and the degree information;

[0045] S44. Design an adaptive adjustment feedback threshold ξ to obtain the optimal number of graph convolution layers k according to the varying spectral differences within different subgraphs;

[0046] S45. To determine the optimal number of graph convolution layers k, take ξ = intra(C) as the early stopping condition for adaptive subgraph iteration, and calculate the intra-class distance intra(C):

[0047]

[0048] Among them, in this formula, C is the category set, |C| is the total number of categories, and c i is the ith category;

[0049] The number of convolutional layers required to connect sparse or dense subgraphs is different. If the number of convolutional layers is too small, the feature information within the subgraph will not be fully aggregated, resulting in large differences in the feature of samples of the same category and the inability to effectively capture the intra-class consistency, which is the under-smoothing phenomenon. If the number of convolutional layers is too large, the feature information within the subgraph will be over-aggregated, resulting in the feature of samples of different categories tending to be similar and losing discriminability, which is the over-smoothing phenomenon;

[0050] Use the intra-class distance intra(C) as a threshold to determine whether the optimal number of convolutional layers is reached. The intra-class distance intra(C) shows a trend of first decreasing and then increasing during the convolution process. Find the optimal number of convolutional layers, that is, the optimal graph convolutional layer number k, according to the change trend;

[0051] S46. In each training round, calculate the intra-class distance intra(C) of the current round, and compare the intra-class distance intra(C) of the t-th round t and the intra-class distance intra(C) of the (t + 1)-th round t+1 . If intra(C) t < intra(C) t+1 , continue the iteration. If intra(C) t > intra(C) t+1 , it is considered that the node features embedded with excessive information residuals have reached a reasonable similarity, stop the iteration, and record the current iteration layer number;

[0052] S47. When intra(C) t > intra(C) t+1 appears for the first time, it is considered that the intra-class distance is at the minimum value, stop the iteration, and prevent the over-smoothing phenomenon from occurring.

[0053] Further, the testing process in S5 further includes:

[0054] S51. Concatenate the final features of each subgraph into a global feature matrix and use a fully connected layer and the Softmax function to generate classification probabilities:

[0055]

[0056] where, in this formula, W cls is the classification weight matrix, with a dimension of 60x|C|;

[0057] S52. Set the maximum number of iterations to 100, and take the average of the obtained overall accuracy, average accuracy, and KAP coefficient results respectively to verify the classification accuracy.

[0058] A computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above-mentioned hyperspectral image classification method based on a subgraph adaptive neural network.

[0059] The beneficial effects of the present invention are reflected in:

[0060] 1. Reduce computational cost: By graph segmentation, the processing of large-scale graphs is transformed into the processing of multiple small-scale subgraphs, reducing computational complexity and storage requirements, and improving the efficiency of processing large graphs.

[0061] 2. Fully exploit features: Using the SSAPGCN method to integrate spatial information and spectral information into the graph structure, and combining with adaptive subgraph convolution, can fully exploit the spectral and spatial features of hyperspectral images and improve the accuracy of feature representation.

[0062] 3. Avoid the smoothing problem: By adaptively determining the optimal number of convolutional layers for each subgraph, the over-smoothing or under-smoothing phenomenon caused by the fixed K-order neighbor information aggregation strategy is avoided, and the effective smoothness of feature information is improved.

[0063] 4. Improve classification performance: Introducing the adaptive subgraph convolutional neural network model into hyperspectral image classification significantly improves the classification effect and accuracy of hyperspectral images, providing more reliable support for the practical application of hyperspectral images. Description of the Drawings

[0064] The drawings described herein are used to provide a further understanding of the present application, form a part of the present application, and the schematic embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation of the present application.

[0065] Figure 1 It is the overall flow block diagram of the classification method of the embodiment of the present invention.

[0066] Figure 2 It is the overall flow schematic diagram of the classification method of the embodiment of the present invention.

[0067] Figure 3 It is the simplified flow chart of the adaptive algorithm of the embodiment of the present invention.

[0068] Figure 4 It is the structural block diagram of the computer device of the embodiment of the present invention. Detailed Embodiments

[0069] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Without conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0070] It should be noted that the meaning of "and / or" appearing throughout the text includes three parallel solutions. Taking "A and / or B" as an example, it includes solution A, or solution B, or the solution where A and B are satisfied simultaneously. In addition, "a plurality of" means two or more. In addition, the technical solutions between the embodiments can be combined with each other, but it must be based on the ability of those of ordinary skill in the art to implement. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the protection scope required by the present invention.

[0071] Hyperspectral images contain rich spectral information and can provide important data support for many fields such as ground object recognition, environmental monitoring, and agricultural yield estimation. By accurately classifying hyperspectral images, the characteristics and distribution information of different ground objects can be effectively extracted, and it has broad application prospects in resource exploration, urban planning, ecological protection, etc. This method uses an adaptive subgraph neural network to process hyperspectral images, aiming to improve the accuracy and efficiency of classification. Specifically, the method process is as follows:

[0072] See Figure 1 , the embodiments of the present invention provide a hyperspectral image classification method based on a subgraph adaptive neural network, including the following steps:

[0073] S1. Construct a pairwise probability graph structure based on the SSAPGCN method to effectively integrate the spatial information and spectral information of the hyperspectral image, providing a basis for subsequent analysis;

[0074] S2. Read the feature indexes of each spectral band in the original spectral image and generate a feature matrix containing spectral information. Perform PCA dimensionality reduction on the input feature matrix data for data preprocessing, remove redundant information, reduce the computational complexity, and improve the processing efficiency;

[0075] S3. Based on the globally reduced preprocessing to obtain the global graph structure S, call the Metis graph segmentation algorithm for graph segmentation, and divide the globally obtained probability graph structure into multiple local subgraphs to facilitate the mining of local features;

[0076] S4. Calculate the intra-class distance intra(C) as the adaptive feedback threshold, and determine the optimal number of graph convolutional layers k based on the spectral differences of the subgraphs to avoid over-smoothing or under-smoothing;

[0077] S5. Use the optimal number of graph convolutional layers k to train and test the global probability graph structure data, complete the hyperspectral image classification, and improve the classification accuracy.

[0078] The present invention provides a hyperspectral image classification method based on an adaptive subgraph neural network. On the one hand, by decomposing a large-scale graph into multiple subgraphs, the processing complexity is reduced, enabling the model to focus more on local features. On the other hand, an adaptive method for determining the optimal number of convolutional layers is adopted to cleverly solve the problem of feature smoothing under different subgraph structures, avoiding the common over-smoothing or under-smoothing problems in traditional methods. In this way, the internal information of hyperspectral images can be more accurately mined, and more accurate image classification can be achieved.

[0079] Among them, the introductions of some professional terms in this application are as follows:

[0080] SSAPGCN method: Used to construct a graph structure in hyperspectral image (HSI) classification, adaptively learn the local graph structure and output features, and optimize the graph construction through a feedback mechanism.

[0081] PCA dimensionality reduction: Project the original high-dimensional data (such as 220 bands) into a low-dimensional space (such as 60 dimensions) through a linear transformation, and retain the principal components with the largest variance.

[0082] Metis graph partitioning algorithm: Through multi-level coarsening, partitioning, and refinement steps, the original graph is partitioned into subgraphs with higher density, reducing the edge connections between subgraphs.

[0083] See Figure 2 , in this embodiment, the S1 further includes:

[0084] S11. Input the spectral feature matrix and spatial coordinate matrix of the hyperspectral image:

[0085] X ∈ R N×d

[0086] L ∈ R N×2

[0087] Among them, in this formula, X is the spectral feature matrix, L is the spatial coordinate matrix, N is the number of pixels, and d is the spectral dimension;

[0088] S12. To establish a graph topological relationship that takes into account both spectral similarity and spatial proximity, construct the probability graph structure A ∈ R by minimizing the following objective function N×N :

[0089]

[0090] Among them, in this formula:

[0091] a ij is the similarity between pixels i and j, which is jointly determined by the spectral difference ‖x i -x j ‖ and the spatial distance ‖l i -l j ‖;

[0092] P ∈ R N×N is the guidance matrix based on expert prior;

[0093] δ is the balance parameter between spectral information and spatial information.

[0094] By minimizing the above objective function, a graph structure that simultaneously considers spectral information and spatial information can be constructed.

[0095] In this step, the SSAPGCN method is used to fuse spatial information and spectral information through a pairwise probability graph, which can effectively enhance the representation ability of the graph structure.

[0096] See Figure 2 , in this embodiment, in S2, in order to remove redundant information, reduce high-dimensional noise, and improve computational efficiency, PCA dimensionality reduction is performed on the input image data:

[0097] X PCA = X · V top

[0098] Among them, in this formula, X PCA is the dimensionality-reduced feature matrix, and V top is the d x 60 eigenvector matrix;

[0099] Retain the first 60 principal component quantities to obtain the dimensionality-reduced feature matrix X PCA ∈ R n×60 , and this low-dimensional feature not only retains the main spectral information but also provides a suitable dimension for subsequent graph neural network processing.

[0100] The above-mentioned dimensionality-reduced feature matrix X PCA is used for subsequent graph segmentation and convolution operations.

[0101] See Figure 2 , in this embodiment, in S3, the multiple subgraphs divided are respectively:

[0102] A = [A1, A2,..., A m , X = [X1, X2,..., X m

[0103] ​Among them, in this formula, each sub-graph A i ∈R n×n maintains the local topological structure and contains the corresponding node features X i ∈R n ×60 , and realizes data parallel processing.

[0104] In this step, after obtaining the overall graph structure, the Metis graph partitioning algorithm is used for graph partitioning.

[0105] See Figure 2 - Figure 3 , in this embodiment, since common SGNN-based methods usually use convolutions with a fixed number of layers for sub-graph convolutions, and due to the imbalance of sub-graph structures, it is more reasonable to use adaptive sub-graph convolutions for different sub-graphs. Therefore, in S4, for the heterogeneity of sub-graphs, for each sub-graph i, the following dynamic mechanism is used to iteratively determine the optimal number of graph convolution layers k:

[0106] S41. Initialize the feature matrix

[0107] S42. The node feature update rule from the original feature of the 0th layer to the (k + 1)th layer is:

[0108]

[0109] Among them, in this formula:

[0110] is the feature vector of node i at the lth layer;

[0111] a ij is the connection weight between nodes i and j in the adjacency matrix A i ;

[0112] n is the total number of nodes in the current sub-graph;

[0113] The feature representation of each node not only contains its own feature information, but also fuses the information of its neighbor nodes;

[0114] S43. To further enhance the effect of the convolution layer and alleviate the over-smoothing problem in the deep training process, a degree matrix residual mechanism is introduced, and the degree information of the nodes is embedded as a residual into the feature representation. This mechanism can better represent the adjacency connections and topological structures of the nodes. The formula is as follows:

[0115]

[0116] Among them, in this formula:

[0117] De d ∈R n×nis the subgraph degree matrix, and the diagonal elements are the node degrees;

[0118] β is the balance coefficient, which is used to control the weights of the original features and degree information;

[0119] Dynamically balance the original features and topological features through the learnable parameter β to alleviate the feature assimilation problem caused by deep propagation;

[0120] S44. Considering that the determination of the optimal number of graph convolution layers k will significantly affect the quality of neighborhood information aggregation, an adaptive adjustment feedback threshold ξ is designed to obtain the optimal number of graph convolution layers k according to the varying spectral differences within different subgraphs;

[0121] S45. To determine the optimal number of graph convolution layers k, take ξ = intra(C) as the early stopping condition for adaptive subgraph iteration, and calculate the within-class distance intra(C):

[0122]

[0123] where, in this formula, C is the set of categories, |C| is the total number of categories, and c i is the i-th category;

[0124] The number of convolution layers required to connect sparse or dense subgraphs is different. If the number of convolution layers is too small, the feature information within the subgraph cannot be fully aggregated, resulting in large differences in the sample features of the same category and being unable to effectively capture the within-class consistency, which is the under-smoothing phenomenon. If the number of convolution layers is too large, the feature information within the subgraph will be over-aggregated, resulting in the sample features of different categories tending to be similar and losing discriminability, which is the over-smoothing phenomenon;

[0125] Use the within-class distance intra(C) as the threshold to judge whether the optimal number of convolution layers is reached. The within-class distance intra(C) shows a trend of first decreasing and then increasing as the number of convolution layers increases during the convolution process. Find the optimal number of convolution layers, that is, the optimal number of graph convolution layers k, according to the change trend;

[0126] The within-class distance intra(C) is used to evaluate the feature consistency within the homogeneous region. Sparse-connected subgraphs may require more convolution layers, while dense-connected subgraphs may only need fewer convolution layers to achieve effective feature smoothing;

[0127] S46. In each training epoch, calculate the within-class distance intra(C) of the current epoch, and compare the within-class distance intra(C) t of the t-th epoch and the within-class distance intra(C) t+1 of the (t + 1)-th epoch. If intra(C) t < intra(C) t+1, then continue the iteration. If intra(C) t > intra(C) t+1 , it is considered that the node features embedded with degree information residuals have reached a reasonable similarity, stop the iteration, and record the current iteration level;

[0128] S47. When intra(C) t > intra(C) t+1 first appears, it is considered that the intra-class distance is at the minimum value, stop the iteration to prevent the over-smoothing phenomenon.

[0129] In this step, the intra-class distance intra(C) is used as the adaptive adjustment feedback threshold for the clustering performance metric. By comparing the magnitudes of intra(C) calculated for different convolutional layer numbers, the optimal graph convolutional layer number k can be obtained and then perform adaptive convolution.

[0130] See Figure 2 , in this embodiment, the test process in S5 further includes:

[0131] S51. Concatenate the final features of each subgraph into a global feature matrix and use a fully connected layer and the Softmax function to generate classification probabilities:

[0132]

[0133] where, in this formula, W cls is the classification weight matrix, with a dimension of 60x|C|;

[0134] S52. Set the maximum number of iterations to 100, and take the average of the obtained overall accuracy, average accuracy, and KAP coefficient results respectively to verify the classification accuracy.

[0135] To verify the effectiveness of this method, experiments were conducted on three public datasets, Indian pines, PaviaU, and Houston, and the detection results are shown in Tables 1 - 3 below:

[0136] Table 1 OA(STD)(%), AA(STD)(%), KC(STD)(%) OF DIFFERENT COMPARED METHODS AND ASGCN METHOD FOR INDIAN DATASET

[0137]

[0138]

[0139] Table 2 OA(STD)(%), AA(STD)(%), KC(STD)(%) OF DIFFERENT COMPARED METHODS AND ASGCN METHOD FOR HOUSTON DATASET

[0140]

[0141]

[0142] Table 3 OA(STD)(%), AA(STD)(%), KC(STD)(%) OF DIFFERENT COMPARED METHODS AND ASGCN METHOD FOR PAVIAU DATASET

[0143] Class DPRN DSGSF morph Former GraphGST ASGCN 1 90.95(8.12) 86.98(7.33) 91.64(6.04) 91.54(6.23) 84.48(4.63) 2 97.81(2.18) 85.35(6.15) 89.91(5.96) 91.64(6.73) 85.67(6.80) 3 97.33(2.21) 99.33(1.30) 86.88(11.55) 83.99(24.79) 98.96(2.86) 4 90.89(2.08) 88.04(2.58) 90.27(5.31) 90.61(2.94) 85.28(4.0) 5 92.57(4.61) 87.55(3.65) 97.76(5.62) 97.34(7.96) 98.29(0.70) 6 50.18(35.83) 97.76(2.29) 60.78(38.45) 69.57(32.00) 88.70(5.99) 7 70.13(6.81) 72.18(7.27) 67.74(12.20) 68.91(9.80) 84.66(6.75) 8 68.92(9.26) 70.99(13.01) 63.73(7.36) 64.74(7.97) 59.01(5.93) 9 66.55(10.94) 57.87(14.95) 62.48(12.83) 75.24(10.01) 63.44(8.89) 10 82.49(11.12) 56.36(10.57) 65.34(21.69) 73.68(12.51) 91.89(4.72) 11 74.81(9.71) 57.36(17.70) 76.30(12.23) 76.32(12.56) 86.25(6.53) 12 76.55(7.04) 54.93(15.38) 61.80(25.30) 73.26(17.76) 77.07(6.43) 13 60.10(9.86) 67.60(19.08) 61.25(23.48) 61.80(25.30) 74.19(7.73) 14 91.11(5.68) 94.03(4.74) 98.40(1.77) 91.90(18.09) 99.38(1.95) 15 97.79(3.01) 94.47(6.24) 97.99(3.87) 100.00(0.00) 97.94(2.61) OA 81.09(2.24) 74.49(6.58) 78.71(1.84) 80.85(2.34) 83.54(1.17) AA 80.26(2.11) 78.05(5.60) 78.84(3.00) 80.67(2.43) 85.01(1.14) KC 79.93(2.55) 72.39(7.13) 76.98(1.99) 79.29(2.52) 82.21(1.26)

[0144] Combined with the above Tables 1 - 3, it can be seen that the detection results obtained on the three public datasets Indian pines, PaviaU, and Houston can reach the overall accuracies of 88.76%, 93.17%, and 83.54% respectively.

[0145] The embodiment of the present invention also provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the processor is caused to execute the steps of the above hyperspectral image classification method based on a subgraph adaptive neural network.

[0146] See Figure 4 , the embodiment of the present invention also provides a computer device including a memory and a processor, the memory storing a computer program, and when the computer program is executed by the processor, the processor is caused to execute the steps of the above hyperspectral image classification method based on a subgraph adaptive neural network.

[0147] The embodiment of the present invention also provides a computer program product containing instructions, and when it runs on a computer, the computer is caused to execute the steps of the above hyperspectral image classification method based on a subgraph adaptive neural network.

[0148] It can be understood that the system, device, and storage medium provided by the embodiment of the present invention correspond to the method provided by the embodiment of the present invention, and the explanations, examples, and beneficial effects of the relevant content can refer to the corresponding parts in the above hyperspectral image classification method based on a subgraph adaptive neural network.

[0149] It should be noted that those of ordinary skill in the art can understand that all or part of the steps implemented in the embodiments of the present invention can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using hardware, it can be implemented in whole or in part in the form of purchasing standard parts or modified parts. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that the computer can access or a data storage device such as a server or data center that includes one or more integrated available media. The available medium can be a magnetic medium (such as a floppy disk, hard disk, magnetic tape), an optical medium (such as a DVD), or a semiconductor medium (such as a solid state disk (SSD)).

[0150] In summary, in view of the problems existing in the existing hyperspectral image classification methods when processing large-scale image data, such as high computational complexity, insufficient utilization of global and local structural information, and the aggregation strategy of fixed K-order neighbor information, which is difficult to effectively smooth the feature information within the subgraph and may lead to over-smoothing or under-smoothing phenomena, etc., therefore, the present invention provides a hyperspectral image classification method based on an adaptive subgraph neural network, which can effectively reduce the computational cost and fully exploit the spectral and spatial features of the image when processing large-scale image data, thereby significantly improving the accuracy and efficiency of classification.

[0151] It should be understood that the examples and embodiments described herein are only for illustration and are not intended to limit the present invention. Those skilled in the art can make various modifications or changes according to it. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A hyperspectral image classification method based on a subgraph adaptive neural network, characterized in that, The steps include the following: S1. Construct a pairwise probability graph structure based on the SSAPGCN method to effectively integrate the spatial information and spectral information of the hyperspectral image; S2. Read the feature indices of each spectral band in the original spectral image and generate a feature matrix containing spectral information. Perform PCA dimensionality reduction on the input feature matrix data for data preprocessing to remove redundant information; S3. Based on the globally reduced and preprocessed graph structure S, call the Metis graph segmentation algorithm for graph segmentation to divide the globally obtained probability graph structure in S1 into multiple local subgraphs for local feature mining; S4. Calculate the intra-class distance intra(C) as an adaptive feedback threshold, and determine the optimal number of graph convolutional layers k according to the spectral differences of the subgraphs to avoid over-smoothing or under-smoothing; S5. Use the optimal number of graph convolutional layers k to train and test the globally obtained probability graph structure data to complete hyperspectral image classification.

2. The hyperspectral image classification method based on the subgraph adaptive neural network according to claim 1, wherein S1 further includes: S11. Input the spectral feature matrix and spatial coordinate matrix of the hyperspectral image: X ∈ R N×d L∈R N×2 Among them, in this formula, X is the spectral feature matrix, L is the spatial coordinate matrix, N is the number of pixels, and d is the spectral dimension; S12. To establish a graph topological relationship that takes into account both spectral similarity and spatial proximity, a probability graph structure A ∈ R is constructed by minimizing the following objective function N×N :[[]] Among them, in this formula: a ij is the similarity between pixels i and j, which is jointly determined by the spectral difference ‖xi - x j ‖ and the spatial distance ‖l i - l j ‖; P ∈ R N×N is the guidance matrix based on expert prior knowledge; δ is the balance parameter between spectral information and spatial information.

3. The hyperspectral image classification method based on the subgraph adaptive neural network according to claim 1, characterized in that, In S2, for removing redundant information, PCA dimensionality reduction is performed: X PCA = X · V top Among them, in this formula, X PCA is the feature matrix after dimensionality reduction, and V top is the d x 60 feature vector matrix; Retain the number of the first 60 principal components to obtain the feature matrix X after dimensionality reduction PCA ∈R n×60 .

4. The hyperspectral image classification method based on subgraph adaptive neural network according to claim 1, wherein, In S3, the multiple divided subgraphs are respectively: A = [A1, A2,..., A m , X = [X1, X2,..., X m ​ Among them, in this formula, each sub-graph A i ∈R n×n maintains the local topological structure and contains the corresponding node feature X i ∈R n×60 , and realizes data parallel processing.

5. The hyperspectral image classification method based on the subgraph adaptive neural network according to claim 1, wherein In S4, for the heterogeneity of the subgraphs, for each subgraph i, the following dynamic mechanism is used to iteratively determine the optimal number of graph convolutional layers k: S41. Initialize the feature matrix S42. The node feature update rule from the original features of the 0th layer to the (k + 1)th layer is: Among them, in this formula: is the feature vector of node i at the l-th layer; a ij is the adjacency matrix A i is the connection weight between nodes i and j in n is the total number of nodes in the current subgraph; S43. Introduce a degree matrix residual mechanism to embed the degree information of the nodes into the feature representation as residuals, and the formula is as follows: Among them, in this formula: De d ∈R n×n is the subgraph degree matrix, and the diagonal elements are the node degrees; β is the balance coefficient used to control the weights of the original features and degree information; S44. Design an adaptive adjustment feedback threshold ξ to obtain the optimal number of graph convolutional layers k according to the varying spectral differences within different subgraphs; S45. To determine the optimal number of graph convolutional layers k, take ξ = intra(C) as the early stopping condition for adaptive subgraph iteration, and calculate the intra-class distance intra(C): Among them, in this formula, C is the set of categories, |C| is the total number of categories, and c i is the i-th category; The number of convolutional layers required to connect sparse or dense subgraphs is different. If the number of convolutional layers is too small, the feature information within the subgraph cannot be fully aggregated, resulting in still large differences in the sample features of the same category and unable to effectively capture the intra-class consistency, which is the under-smoothing phenomenon. If the number of convolutional layers is too large, the feature information within the subgraph will be over-aggregated, resulting in the sample features of different categories tending to be similar and losing discriminability, which is the over-smoothing phenomenon; Use the intra-class distance intra(C) as the threshold to judge whether the optimal number of convolutional layers is reached. The intra-class distance intra(C) shows a trend of first decreasing and then increasing during the convolution process. Find the optimal number of convolutional layers, that is, the optimal number of graph convolutional layers k, according to the change trend; S46. In each training round, calculate the intra-class distance intra(C) of the current round, and compare the intra-class distance intra(C) in the t-th round t with the intra-class distance intra(C) in the (t + 1)-th round t+1 . If intra(C) t < intra(C) t+1 , continue the iteration. If intra(C) t > intra(C) t+1 , it is considered that the node features after degree information residual embedding have reached a reasonable similarity, stop the iteration and record the current iteration layer number; S47. When intra(C) first appears t > intra(C) t+1 it is considered that the within-class distance is at the minimum value, and the iteration is stopped to prevent over-smoothing.

6. The hyperspectral image classification method and device based on the subgraph adaptive neural network according to claim 1, characterized in that, The testing process in S5 further includes: S51. Concatenate the final features of each sub - figure into a global feature matrix Use a fully - connected layer and the Softmax function to generate classification probabilities: Among them, in this formula, W cls is the classification weight proof, and the dimension is 60x|C|; Set the maximum number of iterations to 100, and average the obtained overall accuracy, average accuracy, and KAP coefficient results respectively to verify the classification accuracy.

7. A computer device, characterized in that, It includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor is caused to execute the steps of the hyperspectral image classification method based on the subgraph adaptive neural network according to any one of claims 1-6.

Images