Hyperspectral remote sensing image classification method and device based on graph neural network

Abstract

The invention discloses a hyperspectral remote sensing image classification method and device based on a graph neural network, and relates to the technical field of hyperspectral image processing, and the method comprises the steps: taking each pixel point in original hyperspectral data as a node to construct an adjacent matrix, and generating graph structure data; removing abnormal pixels in first-order neighbor nodes by using a cosine similarity measurement function, constructing a new adjacency matrix, performing dimensionality reduction on original data by using a principal component analysis method, extracting main spectral feature components, reducing calculation complexity, and improving the accuracy of spectral feature extraction; and inputting the optimized adjacent matrix and the dimensionality-reduced three-dimensional hyperspectral remote sensing image into a pre-trained graph neural network model, aggregating neighbor features of each pixel point through an average aggregation mode, extracting features of an aggregated central pixel point through a KAN network to obtain embedding of the central pixel point, and finally performing Softmax processing to obtain a classification result. The classification efficiency and accuracy are improved, and a breakthrough progress is brought to the field of hyperspectral remote sensing image processing.

Description

Technical Field

[0001] The present invention relates to a hyperspectral remote sensing image classification method and device based on graph neural network, belonging to the technical field of hyperspectral remote sensing image processing. Background Art

[0002] Hyperspectral remote sensing images are rich in spectral information and have been widely used in modern agriculture, environmental protection, military reconnaissance and other fields. In the field of agriculture, hyperspectral remote sensing images can be used to monitor crop health and evaluate yield and quality; in environmental protection, it helps to monitor vegetation cover, water pollution and soil quality; and in military reconnaissance, hyperspectral remote sensing images can be used for target detection and recognition. However, in the process of collecting and applying hyperspectral data, due to the limitations of sensor performance and the influence of complex environmental factors, the noise problem has become an important challenge that cannot be ignored. The phenomenon of different spectra for the same object and different objects for the same spectrum makes the classification task of hyperspectral remote sensing images more complicated and difficult. In early studies, in order to meet this challenge, researchers mainly relied on traditional statistical methods for hyperspectral remote sensing image classification, such as spectral angular distance and Euclidean distance algorithms, and realized the basic classification of images through spectral feature analysis based on prior knowledge. Although these methods can distinguish different types of objects to a certain extent, they are difficult to deal with classification problems in complex scenes. With the advancement of random machine learning technology, algorithms such as k-nearest neighbor (KNN), support vector machine (SVM), and random forest (RF) have been introduced into hyperspectral remote sensing image classification. These methods significantly improve the accuracy and robustness of classification through pattern recognition and feature selection. In particular, SVM has achieved good results in hyperspectral remote sensing image classification due to its strong generalization ability. The rise of deep learning has brought new opportunities for hyperspectral remote sensing image classification. Deep learning methods represented by convolutional neural networks (CNNs) have made historic breakthroughs in the field of image processing with their powerful feature extraction and abstraction capabilities. CNNs can automatically learn high-level features in images, thereby greatly improving the performance of hyperspectral remote sensing image classification. However, traditional CNN models have certain limitations when processing hyperspectral data. Since CNNs can only process structured data, in hyperspectral remote sensing image classification, training samples and test samples usually need to be processed into neighborhood blocks of the same size. This method ignores the continuity and spatial correlation of the distribution of objects in hyperspectral remote sensing images, especially for nodes at the boundary of categories, the neighborhood blocks often contain pixels from different categories, resulting in a large number of abnormal pixels. These abnormal pixels will have a negative impact on the training and testing of the model, reducing the accuracy and reliability of classification. Summary of the invention

[0003] The purpose of the present invention is to provide a hyperspectral remote sensing image classification method and device based on graph neural network, which can achieve effective classification of original hyperspectral remote sensing images by improving the spectral and spatial feature extraction capabilities, and also significantly improve the classification efficiency and accuracy.

[0004] In order to achieve the above objectives / solve the above technical problems, the present invention is implemented by adopting the following technical solutions.

[0005] In one aspect, the present invention provides a hyperspectral remote sensing image classification method based on graph neural network, comprising the following steps:

[0006] Acquire the three-dimensional hyperspectral remote sensing image to be tested;

[0007] Constructing an adjacency matrix according to each pixel point in the three-dimensional hyperspectral remote sensing image to be measured;

[0008] Reconstructing the spectral features of the objects in the three-dimensional hyperspectral remote sensing image to be measured, identifying abnormal pixels in the adjacency matrix using a cosine similarity measurement function according to the reconstructed spectral features of the objects, and removing the abnormal pixels to obtain an optimized adjacency matrix;

[0009] Perform dimensionality reduction processing on the three-dimensional hyperspectral remote sensing image to be measured;

[0010] The optimized adjacency matrix and the dimensionality-reduced three-dimensional hyperspectral remote sensing image to be tested are input into the pre-trained graph neural network model to obtain the classification result;

[0011] The processing process of the trained graph neural network model includes:

[0012] The dimensionality-reduced three-dimensional hyperspectral remote sensing image to be tested is constructed into graph structure data through the optimized adjacency matrix. The neighbor features of each pixel in the graph structure data are aggregated by average aggregation to obtain the aggregated features. The features of the central pixel are extracted from the aggregated features through the KAN network to obtain the embedding of the central pixel. The embedding of the central pixel is processed by Softmax to obtain the classification result.

[0013] Furthermore, the specific method of constructing an adjacency matrix according to each pixel point in the three-dimensional hyperspectral remote sensing image to be measured is:

[0014] Each pixel in the hyperspectral remote sensing image is connected as a node to its neighboring nodes to construct an adjacency matrix, which is expressed as:

[0015] ;

[0016] in: Pixel The adjacency matrix of is the i-th pixel of the hyperspectral remote sensing image, , is the set of real numbers, is the number of bands of the hyperspectral remote sensing image, represent The neighbor nodes of is the jth neighbor node, ;

[0017] Construct an adjacency matrix to form graph-structured data to better capture the spatial relationship between pixels.

[0018] Furthermore, the method of reconstructing the spectral features of the objects in the three-dimensional hyperspectral remote sensing image to be measured, identifying abnormal pixels in the adjacency matrix using the cosine similarity measurement function according to the reconstructed spectral features of the objects, and removing the abnormal pixels to obtain the optimized adjacency matrix specifically includes:

[0019] The spectral characteristics of the objects in the three-dimensional hyperspectral remote sensing image to be measured are reconstructed, and the expression is:

[0020] ;

[0021] ;

[0022] ;

[0023] in, After reconstruction The spectral characteristics of each pixel, yes The pixel The value of the band, is the spectral feature The mean of is the spectral feature The standard deviation of

[0024] The original spectral data were standardized to ensure that the data distribution conformed to the normal distribution. The spectral features were scaled using a standard normalization function to eliminate the numerical differences between different spectral features and increase the distinction between different categories of spectral features, thereby improving the accuracy of subsequent processing steps.

[0025] Optimize the adjacency matrix, the expression is:

[0026] ;

[0027] ;

[0028] ;

[0029] in, is the spectral feature of the jth neighbor node after reconstruction; yes and The dot product of yes The Euclidean norm of , yes The Euclidean norm of ;

[0030] and yes and No. The value of each band;

[0031] When the cosine similarity metric function When and The edges between them are from the adjacency matrix Remove it and get the optimized adjacency matrix ;

[0032] To avoid the interference of abnormal pixels, the cosine similarity measurement function is used to remove abnormal pixels in the first-order nodes and construct a new adjacency matrix to improve the robustness and accuracy of the adjacency matrix.

[0033] Furthermore, the dimensionality reduction processing of the three-dimensional hyperspectral remote sensing image to be measured specifically includes:

[0034] The mean normalization of each band value of each pixel point is performed, and the expression is:

[0035] ;

[0036] in: is the data matrix of each band value for each pixel point, It is the data matrix of each band value of each pixel after removing the mean;

[0037] Calculate the covariance matrix of the data matrix after removing the mean , the expression is:

[0038] ;

[0039] Where n is the number of samples;

[0040] Covariance matrix Perform eigenvalue decomposition to obtain eigenvalues and the corresponding eigenvector :

[0041] ;

[0042] Among them: eigenvector For the The directions of the principal components, the eigenvalues For the The size of the variance in the direction of the principal components;

[0043] Before selection The eigenvector with the largest eigenvalue , forming a dimensionality reduction transformation matrix, the expression is:

[0044] ;

[0045] in, Is included before The matrix of eigenvectors;

[0046] Project the hyperspectral remote sensing image into a low-dimensional space, and the data after dimensionality reduction for:

[0047] ;

[0048] in: It is the data matrix of each band value of each pixel after dimensionality reduction;

[0049] Using principal component analysis (PCA) will reduce the number of bands, facilitate the extraction of main spectral characteristic components, reduce data dimensions, and reduce computational complexity.

[0050] Furthermore, the three-dimensional hyperspectral remote sensing image to be tested after dimension reduction is constructed into graph structure data through the optimized adjacency matrix, the neighbor features of each pixel in the graph structure data are aggregated by average aggregation to obtain aggregate features, the features of the central pixel are extracted from the aggregate features through the KAN network to obtain the embedding of the central pixel, and the embedding of the central pixel is subjected to Softmax processing to obtain the classification result, which specifically includes:

[0051] ;

[0052] ;

[0053] ;

[0054] in: For the neural network The mean of the neighbor node features of the layer, For the The mean of the layer center node and neighbor nodes, is the optimized adjacency matrix Center and Pixel A set of adjacent pixels. is the optimized adjacency matrix The neighboring pixels of a pixel, It is Pixels Layer embedding, Is Aggregation Neighboring pixel The embedding obtained after layer features,

[0055] It is Pixels Layer embedding, It is Pixels Layer embedding, , yes Middle The features of pixels after dimensionality reduction;

[0056] Hyperspectral remote sensing images The band characteristics are output through the univariate function. is a function in high-dimensional space, It is The final embedding of the pixel output;

[0057] A variety of aggregation methods are used to integrate the first-order spatial features of the nodes, and the Kolmogorov-Arnold Network (KAN) model is combined to extract spectral features to enhance the model's expressiveness, improve the model's classification ability, and improve the model's accuracy.

[0058] In a second aspect, the present invention provides a hyperspectral remote sensing image classification device based on a graph neural network, comprising:

[0059] An acquisition module is used to acquire the three-dimensional hyperspectral remote sensing image to be tested;

[0060] A construction module, used for constructing an adjacency matrix according to each pixel point in the three-dimensional hyperspectral remote sensing image to be measured;

[0061] The optimization module is used to reconstruct the spectral features of the ground objects in the three-dimensional hyperspectral remote sensing image to be measured, identify abnormal pixels in the adjacency matrix using the cosine similarity measurement function according to the reconstructed spectral features of the ground objects, and remove the abnormal pixels to obtain an optimized adjacency matrix;

[0062] Dimensionality reduction module, used to perform dimensionality reduction processing on the three-dimensional hyperspectral remote sensing image to be tested;

[0063] The classification module is used to input the optimized adjacency matrix and the dimensionality-reduced three-dimensional hyperspectral remote sensing image to be tested into the pre-trained graph neural network model to obtain the classification result;

[0064] The processing process of the trained graph neural network model includes:

[0065] The dimensionality-reduced three-dimensional hyperspectral remote sensing image to be tested is constructed into graph structure data through the optimized adjacency matrix. The neighbor features of each pixel in the graph structure data are aggregated by average aggregation to obtain the aggregated features. The features of the central pixel are extracted from the aggregated features through the KAN network to obtain the embedding of the central pixel. The embedding of the central pixel is processed by Softmax to obtain the classification result.

[0066] In a third aspect, the present invention provides a hyperspectral remote sensing image classification system based on a graph neural network, comprising:

[0067] Memory, for storing computer programs / instructions;

[0068] A processor is used to execute the computer program / instructions to implement the steps of the above-mentioned hyperspectral remote sensing image classification method based on graph neural network.

[0069] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program / instruction stored thereon. When the computer program / instruction is executed by a processor, the steps of the above-mentioned hyperspectral remote sensing image classification method based on graph neural network are implemented.

[0070] In a fifth aspect, the present invention provides a computer program product, comprising a computer program / instruction, which, when executed by a processor, implements the steps of the above-mentioned hyperspectral remote sensing image classification method based on graph neural network.

[0071] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: the present invention first regards each pixel point of the hyperspectral remote sensing image as a node, constructs an adjacency matrix, and forms graph structure data, so as to better capture the spatial relationship between pixels; then, the original spectral data is standardized to ensure that the data distribution conforms to the normal distribution, thereby improving the accuracy of subsequent processing steps; on this basis, the cosine similarity measurement function is used to identify and eliminate abnormal pixels in the first-order neighbor nodes, optimize the adjacency matrix, and reduce noise interference in the classification process; then, the principal component analysis (PCA) method is applied to reduce the dimension of the data and simplify the feature space; on this basis, a graph neural network model is constructed, and the first-order spatial features of the nodes are integrated by multiple aggregation methods, and the spectral features are extracted in combination with the Kolmogorov-Arnold Network (KAN) network model to enhance the expression ability of the model, and finally through the trained advanced classification model, remote sensing image detection and classification are performed. This method not only realizes the effective classification of the original hyperspectral remote sensing images, but also significantly improves the classification efficiency and accuracy, bringing breakthrough progress to the field of hyperspectral remote sensing image processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 It is a flow chart of the present invention.

[0073] Figure 2 It is a framework diagram of the present invention.

[0074] Figure 3 It is an RGB pseudo-color image of the Indian Pines hyperspectral remote sensing image in an embodiment of the present invention.

[0075] Figure 4 It is a real label image of the Indian Pines hyperspectral remote sensing image in an embodiment of the present invention.

[0076] Figure 5 This is a schematic diagram of the results of the SVM classification method;

[0077] Figure 6 This is a schematic diagram of the results of the 2D-CNN classification method;

[0078] Figure 7 This is a schematic diagram of the results of the 3D-CNN classification method;

[0079] Figure 8 This is a schematic diagram of the results of the SSRN classification method;

[0080] Fig. 9 It is a schematic diagram of the results of the SSFTT classification method;

[0081] Fig.10 This is a schematic diagram of the results of the GraphTransformer classification method;

[0082] Fig.11 It is a schematic diagram of the results of the GAT classification method;

[0083] Fig.12 This is a schematic diagram of the results of the GraphSAGE classification method;

[0084] Fig.13 Schematic diagram of the results of the GCN classification method;

[0085] Fig.14 Schematic diagram of the classification method results of this embodiment. DETAILED DESCRIPTION

[0086] The technical solution of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations on the technical solution of the present invention. The embodiments of the present invention and the technical features in the embodiments may be combined with each other unless there is a conflict.

[0087] The term "and / or" is only a description of the association relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " generally indicates that the related objects are in an "or" relationship.

[0088] Example 1

[0089] like Figure 1 to Figure 2 An embodiment shown in the figure provides a hyperspectral remote sensing image classification method based on a graph neural network, comprising the following steps:

[0090] Step S1: obtaining a three-dimensional hyperspectral remote sensing image to be tested, and constructing an adjacency matrix according to each pixel point in the three-dimensional hyperspectral remote sensing image to be tested, specifically comprising:

[0091] Acquiring 3D hyperspectral remote sensing images , each pixel Constructing an adjacency matrix as a node , each node With The 48 neighboring nodes connected;

[0092] Adjacency Matrix The expression is:

[0093] ;

[0094] in: Pixel The adjacency matrix of is the i-th pixel of the hyperspectral remote sensing image, , is the number of bands of the hyperspectral remote sensing image, represent The neighbor nodes of is the jth neighbor node, , are the length and width of the hyperspectral remote sensing image respectively.

[0095] Step S2: reconstructing the spectral features of the objects in the three-dimensional hyperspectral remote sensing image to be measured, identifying abnormal pixels in the adjacency matrix using the cosine similarity measurement function according to the reconstructed spectral features of the objects, and removing the abnormal pixels to obtain an optimized adjacency matrix, specifically comprising:

[0096] Step S2.1: reconstructing the spectral features of the objects in the three-dimensional hyperspectral remote sensing image to be measured, specifically includes:

[0097] The spectral information in the original three-dimensional hyperspectral remote sensing image data to be measured Perform feature scaling and use the standard normalization function to normalize the spectral features. The expression is:

[0098] ;

[0099] ;

[0100] ;

[0101] in, After reconstruction The spectral characteristics of each pixel, is the i-th pixel of the hyperspectral remote sensing image, yes The pixel The value of the band, is the spectral feature The mean of is the spectral feature The standard deviation of is the number of bands of the hyperspectral remote sensing image;

[0102] Step S2.2: The first-order neighbor nodes constructed based on the spatial hypothesis may contain pixels of other categories. In order to avoid abnormal pixels The cosine similarity measurement function is used to remove abnormal pixels in the first-order nodes. The expression of the optimized adjacency matrix is:

[0103] ;

[0104] ;

[0105] ;

[0106] in, is the spectral feature of the jth neighbor node after reconstruction; yes and The dot product of yes The Euclidean norm of , yes The Euclidean norm of ;

[0107] and yes and No. The value of each band;

[0108] When the cosine similarity metric function When and The edges between them are from the adjacency matrix Remove it and get the optimized adjacency matrix .

[0109] Step S3: performing dimensionality reduction processing on the three-dimensional hyperspectral remote sensing image to be measured, specifically including:

[0110] Original 3D hyperspectral remote sensing image Each pixel in The spectrum of the image was reduced from 200 bands to 70 bands using principal component analysis (PCA);

[0111] The expression for normalizing the mean value of each band value at each pixel is:

[0112] ;

[0113] in: is the mean of the band, is the data matrix of each band value for each pixel point, It is the data matrix of each band value of each pixel after removing the mean;

[0114] Calculate the covariance matrix of the data matrix after removing the mean , the expression is:

[0115] ;

[0116] Where n is the number of samples;

[0117] Covariance matrix Perform eigenvalue decomposition to obtain eigenvalues and the corresponding eigenvector :

[0118] ;

[0119] Among them: eigenvector For the The directions of the principal components, the eigenvalues For the The size of the variance in the direction of the principal components;

[0120] Before selection The eigenvector with the largest eigenvalue , forming a dimensionality reduction transformation matrix, the expression is:

[0121] ;

[0122] in, Is included before The matrix of eigenvectors;

[0123] Project the hyperspectral remote sensing image into a low-dimensional space, and the data after dimensionality reduction for:

[0124] ;

[0125] in: It is the data matrix of each band value of each pixel after dimensionality reduction.

[0126] Step S4: Input the optimized adjacency matrix and the three-dimensional hyperspectral remote sensing image to be tested after dimension reduction into a pre-trained graph neural network model to obtain a classification result; wherein the processing process of the trained graph neural network model includes: constructing the three-dimensional hyperspectral remote sensing image to be tested after dimension reduction into graph structure data through the optimized adjacency matrix, aggregating the neighbor features of each pixel in the graph structure data by average aggregation to obtain aggregate features, extracting the features of the central pixel from the aggregate features through the KAN network to obtain the embedding of the central pixel, and performing Softmax processing on the embedding of the central pixel to obtain the classification result, specifically including:

[0127] Step S4.1: The graph neural network model aggregates the information of adjacent pixels through an aggregation strategy, extracts the spectral features of each pixel through a neural network, and After the iterative aggregation, each pixel can capture its The structural information within the jump neighborhood, formally, the first The layers are represented as follows:

[0128] ;

[0129] ;

[0130] in: For the neural network The mean of the neighbor node features of the layer, For the The mean of the layer center node and neighbor nodes, is the optimized adjacency matrix Center and Pixel A set of adjacent pixels. is the optimized adjacency matrix The neighboring pixels of a pixel, yes Pixel No. Layer embedding, Is Aggregation Neighboring pixel The embedding obtained after layer features;

[0131] It is Pixels Layer embedding, It is Pixels The layer embedding, and , yes Middle The features of pixels after dimensionality reduction;

[0132] In order to more fully extract the spectral features, the KAN classification model is used to extract the features of the central pixel of the aggregated information. The expression is:

[0133] ;

[0134] in, is a univariate function, representing the The band characteristics are output through the univariate function. is a function in high-dimensional space, used to combine multiple The result is output after multi-layer graph convolution. Final embedding of the node .

[0135] The training method of the graph neural network model includes: obtaining a three-dimensional hyperspectral remote sensing image and constructing a training set, inputting the training set into the graph neural network model to be trained, optimizing the loss function using the Adam optimizer, adjusting the network model parameters, and stopping the training when the loss no longer decreases, so that the model can accurately distinguish the spectral features of different categories, and obtaining a trained graph neural network model;

[0136] (1) Initialize the estimated values ​​of the first-order moment and the second-order moment:

[0137] ;

[0138] ;

[0139] (2) In each iteration, calculate the gradient :

[0140] (3) Update the first-order moment estimate (momentum):

[0141] ;

[0142] (4) Update the second-order moment estimate;

[0143] ;

[0144] (5) Calculate the bias-corrected first-order moment estimate and the second moment estimate :

[0145] ;

[0146] ;

[0147] (6) Parameter update formula:

[0148] ;

[0149] like Figure 3 As shown, in this embodiment, the Indian Pines dataset was collected by the AVIRIS sensor in a rural area in northwestern Indiana, USA in 1992, with a wavelength range of 0.4-2.45 , with a spatial resolution of 20 meters, including 145 145 pixels, 224 bands, some of which are water absorption bands and useless bands. These bands were removed in the experiment, and the final data set contained 200 bands in total; the data set contains a total of 21025 pixels, 16 kinds of crops, of which 10249 pixels are labeled. The real labels of the ground objects are as follows Figure 4 shown.

[0150] like Figures 5 to 14 The classification diagrams of different methods are shown. For the convenience of comparison, this embodiment shows the false color pictures of Indian Pines and the real labels. 1% of the training samples are randomly selected from each category in the Indian Pines data. The present invention is compared with several advanced graph neural network methods. The experimental results are shown in Table 1. It can be seen that the OA, AA, and K of the method in this paper are higher than those of the 9 comparison methods. Among the 16 types of objects, the classification accuracy of 6 types of objects exceeds that of other methods. The OA of SVM is 59.23%. The best performance of the four CNN-based classification models (i.e., 2D-CNN, 3D-CNN, SSRN, and SSFTT) is SSRN, with an OA of 86.16%. The best performance of the four GCN-based methods (i.e., GraphTransformer, GAT, GraphSAGE, and GCN) is GCN, with an OA of 88.85%. The NAGIN we proposed is 91.07%, which is 2.22% higher than GCN.

[0151] Table 1. Per-class accuracy and classification metrics obtained by different methods on the Indian Pines dataset.

[0152]

[0153] From the above table, we can see that SVM has the worst classification effect, and we can see the obvious salt-and-pepper phenomenon. Other methods use spatial features and have better noise control, but there will be patches of misclassified areas, especially at the category boundaries. This misclassification phenomenon is more obvious. Relatively speaking, the overall classification performance of this method is better.

[0154] Example 2

[0155] This embodiment provides a hyperspectral remote sensing image classification device based on a graph neural network, comprising:

[0156] An acquisition module is used to acquire the three-dimensional hyperspectral remote sensing image to be tested;

[0157] A construction module, used for constructing an adjacency matrix according to each pixel point in the three-dimensional hyperspectral remote sensing image to be measured;

[0158] The optimization module is used to reconstruct the spectral features of the ground objects in the three-dimensional hyperspectral remote sensing image to be measured, identify abnormal pixels in the adjacency matrix using the cosine similarity measurement function according to the reconstructed spectral features of the ground objects, and remove the abnormal pixels to obtain an optimized adjacency matrix;

[0159] Dimensionality reduction module, used to perform dimensionality reduction processing on the three-dimensional hyperspectral remote sensing image to be tested;

[0160] The classification module is used to input the optimized adjacency matrix and the dimensionality-reduced three-dimensional hyperspectral remote sensing image to be tested into the pre-trained graph neural network model to obtain the classification result;

[0161] The processing process of the trained graph neural network model includes:

[0162] The dimensionality-reduced three-dimensional hyperspectral remote sensing image to be tested is constructed into graph structure data through the optimized adjacency matrix. The neighbor features of each pixel in the graph structure data are aggregated by average aggregation to obtain the aggregated features. The features of the central pixel are extracted from the aggregated features through the KAN network to obtain the embedding of the central pixel. The embedding of the central pixel is processed by Softmax to obtain the classification result.

[0163] Example 3

[0164] This embodiment provides a hyperspectral remote sensing image classification system based on a graph neural network, including:

[0165] Memory, for storing computer programs / instructions;

[0166] A processor is used to execute the computer program / instructions to implement the steps of the above-mentioned hyperspectral remote sensing image classification method based on graph neural network.

[0167] Example 4

[0168] This embodiment provides a computer-readable storage medium having a computer program / instruction stored thereon. When the computer program / instruction is executed by a processor, the steps of the above-mentioned hyperspectral remote sensing image classification method based on graph neural network are implemented.

[0169] Example 5

[0170] This embodiment provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the above-mentioned hyperspectral remote sensing image classification method based on graph neural network.

[0171] The above embodiments confirm the feasibility of the present invention in the classification of hyperspectral remote sensing images. The present invention solves the classification performance bottleneck caused by the fixed spatial structure in the classification of remote sensing images, improves the quality of training samples and test samples, and makes the classification performance more stable when the training samples are randomly selected, thereby effectively improving the accuracy of hyperspectral remote sensing image classification.

[0172] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented 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.

[0173] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0174] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1A function specified in one or more boxes.

[0175] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0176] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the enlightenment of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the purpose of the present invention and the claims, which all fall within the protection of the present invention.

Claims

1. A hyperspectral remote sensing image classification method based on graph neural network, characterized in that: The following steps are involved: Acquire the three-dimensional hyperspectral remote sensing image to be tested; Constructing an adjacency matrix according to each pixel point in the three-dimensional hyperspectral remote sensing image to be measured; Reconstructing the spectral features of the objects in the three-dimensional hyperspectral remote sensing image to be measured, identifying abnormal pixels in the adjacency matrix using a cosine similarity measurement function according to the reconstructed spectral features of the objects, and removing the abnormal pixels to obtain an optimized adjacency matrix; Perform dimensionality reduction processing on the three-dimensional hyperspectral remote sensing image to be measured; The optimized adjacency matrix and the dimensionality-reduced three-dimensional hyperspectral remote sensing image to be tested are input into the pre-trained graph neural network model to obtain the classification result; The processing process of the trained graph neural network model includes: The dimensionality-reduced three-dimensional hyperspectral remote sensing image to be tested is constructed into graph structure data through the optimized adjacency matrix. The neighbor features of each pixel in the graph structure data are aggregated by average aggregation to obtain the aggregated features. The features of the central pixel are extracted from the aggregated features through the KAN network to obtain the embedding of the central pixel. The embedding of the central pixel is processed by Softmax to obtain the classification result.

2. The hyperspectral remote sensing image classification method based on graph neural network according to claim 1 is characterized in that: The step of constructing an adjacency matrix according to each pixel point in the three-dimensional hyperspectral remote sensing image to be measured specifically includes: Each pixel in the hyperspectral remote sensing image is connected as a node to its neighboring nodes to construct an adjacency matrix, which is expressed as: ； in: Pixel The adjacency matrix of is the i-th pixel of the hyperspectral remote sensing image, , is the set of real numbers, is the number of bands of the hyperspectral remote sensing image, represent The neighbor nodes of is the jth neighbor node, .

3. The hyperspectral remote sensing image classification method based on graph neural network according to claim 2 is characterized in that: The method of reconstructing the spectral features of the objects in the three-dimensional hyperspectral remote sensing image to be measured, identifying abnormal pixels in the adjacency matrix using a cosine similarity measurement function according to the reconstructed spectral features of the objects, and removing the abnormal pixels to obtain an optimized adjacency matrix specifically includes: The spectral characteristics of the objects in the three-dimensional hyperspectral remote sensing image to be measured are reconstructed, and the expression is: ； ； ； in, After reconstruction The spectral characteristics of each pixel, yes The pixel The value of the band, is the spectral feature The mean of is the spectral feature The standard deviation of Optimize the adjacency matrix, the expression is: ； ； ； in, is the spectral feature of the jth neighbor node after reconstruction; yes and The dot product of yes The Euclidean norm of , yes The Euclidean norm of ; and yes and No. The value of each band; When the cosine similarity metric function When and The edges between them are from the adjacency matrix Remove it and get the optimized adjacency matrix .

4. The hyperspectral remote sensing image classification method based on graph neural network according to claim 3 is characterized in that: The dimensionality reduction processing of the three-dimensional hyperspectral remote sensing image to be measured specifically includes: The mean normalization of each band value of each pixel point is performed, and the expression is: ； in: is the data matrix of each band value for each pixel point, It is the data matrix of each band value of each pixel after removing the mean; Calculate the covariance matrix of the data matrix after removing the mean , the expression is: ； Where n is the number of samples; Covariance matrix Perform eigenvalue decomposition to obtain eigenvalues and the corresponding eigenvector : ； Among them: eigenvector For the The directions of the principal components, the eigenvalues For the The size of the variance in the direction of the principal components; Before selection The eigenvector with the largest eigenvalue , forming a dimensionality reduction transformation matrix, the expression is: ； in, Is included before The matrix of eigenvectors; Project the hyperspectral remote sensing image into a low-dimensional space, and the data after dimensionality reduction for: ； in: It is the data matrix of each band value of each pixel after dimensionality reduction.

5. The hyperspectral remote sensing image classification method based on graph neural network according to claim 4 is characterized in that: The method constructs the dimensionally reduced three-dimensional hyperspectral remote sensing image to be tested into graph structure data through the optimized adjacency matrix, aggregates the neighbor features of each pixel in the graph structure data through average aggregation to obtain aggregate features, extracts the features of the central pixel from the aggregate features through the KAN network to obtain the embedding of the central pixel, and performs Softmax processing on the embedding of the central pixel to obtain the classification result, specifically including: ； ； ； in: For the neural network The mean of the neighbor node features of the layer, For the The mean of the layer center node and neighbor nodes, is the optimized adjacency matrix Center and Pixel A set of adjacent pixels. is the optimized adjacency matrix The neighboring pixels of a pixel, It is Pixels Layer embedding, Is Aggregation Neighboring pixel The embedding obtained after layer features, It is Pixels Layer embedding, It is Pixels Layer embedding, , yes Middle The features of pixels after dimensionality reduction; Hyperspectral remote sensing images The band characteristics are output through the univariate function. is a function in high-dimensional space, It is The final embedding of the pixel output.

6. A hyperspectral remote sensing image classification device based on graph neural network, characterized in that: include: An acquisition module is used to acquire the three-dimensional hyperspectral remote sensing image to be tested; A construction module, used for constructing an adjacency matrix according to each pixel point in the three-dimensional hyperspectral remote sensing image to be measured; The optimization module is used to reconstruct the spectral features of the ground objects in the three-dimensional hyperspectral remote sensing image to be measured, identify abnormal pixels in the adjacency matrix using the cosine similarity measurement function according to the reconstructed spectral features of the ground objects, and remove the abnormal pixels to obtain an optimized adjacency matrix; Dimensionality reduction module, used to perform dimensionality reduction processing on the three-dimensional hyperspectral remote sensing image to be tested; The classification module is used to input the optimized adjacency matrix and the dimensionality-reduced three-dimensional hyperspectral remote sensing image to be tested into the pre-trained graph neural network model to obtain the classification result; The processing process of the trained graph neural network model includes: The dimensionality-reduced three-dimensional hyperspectral remote sensing image to be tested is constructed into graph structure data through the optimized adjacency matrix. The neighbor features of each pixel in the graph structure data are aggregated by average aggregation to obtain the aggregated features. The features of the central pixel are extracted from the aggregated features through the KAN network to obtain the embedding of the central pixel. The embedding of the central pixel is processed by Softmax to obtain the classification result.

7. A hyperspectral remote sensing image classification system based on graph neural network, characterized in that: include: Memory, for storing computer programs / instructions; A processor, configured to execute the computer program / instructions to implement the steps of the hyperspectral remote sensing image classification method based on graph neural network according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instruction is executed by a processor, the steps of the hyperspectral remote sensing image classification method based on graph neural network described in any one of claims 1 to 5 are implemented.

9. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the hyperspectral remote sensing image classification method based on graph neural network described in any one of claims 1 to 5 are implemented.

Images