CN117934975A - Full-variation regular guide graph convolution unsupervised hyperspectral image classification method - Google Patents

Abstract

The invention relates to an unsupervised hyperspectral image classification method based on total variation regular guide graph convolution. Compared with the prior art, the method solves the problem that the hyperspectral image of the complex scene is difficult to accurately classify under the condition of no label. The invention comprises the following steps: denoising pretreatment of hyperspectral images based on relative total variation; constructing a joint model of a total variation regularization guide graph convolution module and a spatial spectrum self-encoder module; training a joint model of a total variation regularization guide graph convolution module and a spatial spectrum self-encoder module; and obtaining an unsupervised classification result of the hyperspectral image. The method aims to solve the problem of lack of sample labels in hyperspectral image classification, and by utilizing a total variation regularization term, the smoothness and consistency of an image space are maintained in a graph rolling network in a learning process, and spatial spectrum self-encoder modules are utilized to extract spatial spectrum local context information, so that the model can maintain robustness and generalization in complex unsupervised classification tasks.

Description

An unsupervised hyperspectral image classification method based on total variation regularized guided graph convolution

Technical Field

The invention relates to the technical field of hyperspectral remote sensing image processing, and in particular to an unsupervised hyperspectral image classification method based on total variation regularized guided graph convolution.

Background technique

In recent years, with the rapid development of hyperspectral imaging technology and the continuous expansion of its application fields, hyperspectral images have been widely and deeply applied in agriculture, environmental monitoring, geological exploration and other fields. Its uniqueness lies in its multi-spectral band information, which enables it to provide extremely rich details and spectral characteristics of the surface of objects, bringing unprecedented information depth to remote sensing image analysis. However, although hyperspectral images have shown great potential in many fields, their processing and analysis still face a series of challenges, one of the most prominent of which is the unsupervised classification problem of hyperspectral images.

In the field of hyperspectral image classification, traditional methods mainly rely on supervised learning, which requires a large number of labeled samples to train classifiers. However, the process of obtaining these large-scale labeled samples is not only time-consuming and laborious, but also difficult to obtain sufficient label information in some special scenarios. To cope with this dilemma, unsupervised learning has gradually become an important way to solve the problem of hyperspectral image classification. Unsupervised learning methods do not rely on pre-labeled samples, but instead classify images by mining the intrinsic structure and features of the data itself, thereby better adapting to different scenarios and application requirements.

However, the current unsupervised learning methods still have some urgent problems to be solved in the classification of hyperspectral images. First, due to the high dimensionality and complex spectral characteristics of hyperspectral images, traditional feature extraction methods are difficult to fully explore the potential information of the image. Secondly, the noise, illumination changes and similarities between different categories in the image make it difficult for unsupervised learning to achieve accurate classification. In order to overcome these problems, it is urgent to propose a hyperspectral image classification method that is insensitive to noise and adaptable to complex scenes to better meet the needs of complex practical application scenarios. This involves the introduction of advanced technologies such as deep learning and feature embedding, as well as more detailed and accurate analysis of large-scale hyperspectral image data.

Summary of the invention

The purpose of the present invention is to overcome the limitations of existing methods, improve the accuracy and robustness of unsupervised classification of hyperspectral images, and provide an unsupervised hyperspectral image classification method based on total variation regularized guided graph convolution.

In order to achieve the above object, the technical solution of the present invention is as follows:

An unsupervised hyperspectral image classification method based on total variation regularized guided graph convolution includes the following steps:

11) Hyperspectral image denoising preprocessing based on relative total variation: obtain the hyperspectral remote sensing image of the area to be classified, and perform relative total variation denoising on the acquired hyperspectral image;

12) Construct a joint model of the total variation regularized guided graph convolution module and the spatial spectrum autoencoder module: An unsupervised hyperspectral image classification model of total variation regularized guided graph convolution includes three parts: the first is the spatial spectrum autoencoder module for local context extraction and feature dimension compression; the second is the total variation guided graph convolution module for global context extraction and feature fusion; the third is the K-means algorithm module used only in the test process to obtain classification results;

13) Train the joint model of the total variation regularized guided graph convolution module and the spatial spectral autoencoder module: divide the hyperspectral image into image blocks, input them into the joint model to first extract the spatial and spectral context, then use the graph neural network to perform global context modeling, and finally complete the model training through the back propagation algorithm;

14) Obtaining unsupervised classification results of hyperspectral images: Apply the model to the entire hyperspectral image to obtain unsupervised hyperspectral image classification results.

The acquisition and preprocessing of the hyperspectral image comprises the following steps:

21) Obtain hyperspectral remote sensing satellite images of the area to be classified;

22) Normalize the pixel values of hyperspectral remote sensing satellite images;

23) Perform relative total variation denoising on hyperspectral remote sensing satellite images

24) Each pixel of the hyperspectral remote sensing satellite image is cropped into a 7×7 image block with the border area filled with 0;

25) Export the processed image to .GIF format;

26) Divide all image blocks into training set and validation set in a ratio of 7:3.

The joint model of constructing the total variation regularized guided graph convolution module and the spatial spectral autoencoder module comprises the following steps:

31) Set up an unsupervised hyperspectral image classification model, including a spatial spectral autoencoder module for feature dimension compression and a total variation guided graph convolution module. The features output by the encoder in the spatial spectral autoencoder are used as node features for each sample to construct a graph structure. After the node features are embedded in the graph, they are input into the graph convolution module to complete feature extraction. Finally, in the test phase, after each sample has completed feature extraction, the K-means algorithm is applied to obtain the final classification result.

32) Setting the network structure of the spatial spectral autoencoder module, which includes an encoder and a corresponding decoder;

The encoder network structure includes:

The first 2D convolution: the convolution kernel size is 3×3, the stride is 2, and the padding is 0;

The second 2D convolution: the convolution kernel size is 3×3, the stride is 2, and the padding is 0;

The third 2D convolution: the convolution kernel size is 1×1, the stride is 1, and the padding is 0;

The decoder network structure includes:

The first 1D convolution: the convolution kernel size is 1×3, the stride is 1, and the padding is 1;

The second 1D convolution: the convolution kernel size is 1×3, the stride is 1, and the padding is 1;

The third 2D convolution: the convolution kernel size is 1×1, the stride is 1, and the padding is 0;

33) Set the total variation guided graph convolutional network structure. The specific steps are:

331) Set the graph embedding module structure:

set up is the field of real numbers, The size of is the number of graph nodes, The size of is the encoder output channel dimension size of the empty spectral autoencoder;

The input of the graph embedding module is , the adjacency matrix is calculated by the following formula Value:

,

in represents the smoothing coefficient, which is set to 0.5; Represents the L2 norm operation of the matrix;

332) Set the graph convolution module network structure:

The graph convolution module accepts an undirected graph ,in represents a vertex set, represents the edge set, and the output of the graph embedding module is used as the input of the first graph convolution unit. Represented as module input , is represented as an adjacency matrix ;

Output of the first graph convolutional unit It can be expressed as the following formula:

,

in is the activation function, is the offset; and is the identity matrix; are the learnable parameters of the network; and Calculated by the following formula:

,

The total variation guided graph convolution includes two graph convolution units, so the output of the entire total variation guided graph convolution is It can be expressed by the following formula:

,

in , The meaning and , The meanings of the other symbols are the same as those of the corresponding symbols above;

After a forward propagation in the graph convolution module, the graph total variation regularization loss It can be calculated by the following formula:

,

in is the total variation loss weight, which is set to 0.01 by default; It is a slicing operation on a tensor; the meanings of the other symbols are the same as those of the corresponding symbols above;

34) Set the K-means algorithm parameters. The specific steps are:

341) Randomly select K cluster centers, denoted as , and the input of the K-means algorithm is the output of the total variation guided graph convolution module denoted as ;

342) During the iteration process, whether the classification is completed can be judged by whether the loss function converges; the loss function is expressed by the following formula:

,

The total variation guided graph convolution module outputs It remains unchanged during the iteration process. Indicates that in the iteration process The cluster category to which the sample belongs;

343) Order is the number of iteration steps, where is the maximum number of iterations, set to 1000;

344) For each sample In each iteration, it is assigned to the nearest cluster center, which is expressed by the following formula:

,

345) For each cluster center In each iteration, the value of the cluster center is updated according to the samples belonging to the cluster center, which can be expressed by the following formula:

.

The joint model of training the total variation regularized guided graph convolution module and the spatial spectral autoencoder module comprises the following steps:

41) During the training process, only the Spatial Spectral Autoencoder module and the Total Variation Guided Graph Convolution module participate in gradient propagation and parameter update; after the network completes a forward propagation, the total loss is calculated;

42) Determine the gradient direction of parameter update through back propagation of total loss, and then update the parameters to complete an iteration of network training;

43) When the network training rounds reach the preset number of training rounds, stop the network training and save the model parameters.

The acquisition of the unsupervised classification result of the hyperspectral image comprises the following steps:

51) For a processed hyperspectral image to be classified, all its divided image blocks are input into the model;

52) Define the spatial spectrum autoencoder module and the total variation guided graph convolution module, load the trained network parameters and freeze the network parameter update;

53) In the graph embedding module, the number of nodes is fixed and equal to the batch size of the forward propagation during training. In the test process of the network, the forward propagation is performed with the same batch size.

If the total number of image blocks cannot divide the batch size, forward propagation is performed on some samples again, and the features extracted twice are finally averaged;

54) After obtaining the features of all samples in the area to be classified, use K-means to iteratively cluster the sample features;

55) Obtain K classification results for all samples in the area to be classified.

Beneficial Effects

The present invention relates to an unsupervised hyperspectral image classification method based on total variation regularized guided graph convolution.

Compared with the prior art, firstly, by introducing the relative total variation denoising operation, the noise in the image is effectively suppressed, and the classification accuracy is improved. Secondly, the hyperspectral image is jointly modeled by the graph convolutional network and the spatial spectrum autoencoder, which better captures the spatial and spectral information of the image and improves the effect of feature extraction. Finally, the method of the present invention is an unsupervised learning method, which avoids the cumbersome process of relying on a large number of labeled samples and is more flexible and applicable to different scenarios and data sets. Therefore, the proposal of the present invention fills the gap in the current unsupervised hyperspectral image classification method, and has high practical value and broad application prospects.

Due to the high noise, complex scenes and redundant spectral information of hyperspectral images, the hyperspectral image classification method proposed in this paper, which is based on total variation regularized guided graph convolution, can maintain the smoothness and consistency of the image space during the learning process, is insensitive to high noise and high dimensionality, and does not require any label annotation, and has higher classification accuracy, robustness and versatility. This method is expected to provide more reliable hyperspectral image classification solutions for agriculture, environmental monitoring, geological exploration and other fields, and promote the development and application of hyperspectral image processing technology.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of an unsupervised hyperspectral image classification method using total variation regularized guided graph convolution;

FIG2 is a general structure diagram of an unsupervised hyperspectral image classification method based on total variation regularized guided graph convolution;

FIG3 is a framework diagram of an autoencoder module according to the present invention;

FIG4 is a framework diagram of a total variation guided graph convolution module involved in the present invention.

Detailed ways

In order to have a further understanding and recognition of the structural features and the effects achieved by the present invention, a preferred embodiment and accompanying drawings are used for detailed description as follows:

As shown in FIG1 , the unsupervised hyperspectral image classification method of the present invention using total variation regularized guided graph convolution comprises the following steps:

The first step is to preprocess the hyperspectral image based on relative total variation:

Obtain the hyperspectral remote sensing image of the area to be classified, perform normalization preprocessing on the hyperspectral remote sensing satellite image, so that the model can converge quickly and stably, and improve the classification accuracy; perform relative total variation denoising on the hyperspectral remote sensing satellite image; crop each pixel point of the hyperspectral remote sensing satellite image into image blocks of the same size and divide them into training sets and validation sets, so that each sample can obtain local spatial context information during the feature extraction process. The specific steps are as follows;

(1) Perform relative total variation denoising on the hyperspectral image of the classification area. The specific steps are as follows:

(1-1) Based on the pixel points in the image As the center, calculate the total change of the horizontal window of the pixel point And the total change of the window in the vertical direction , and The calculation process is as follows:

,

,

in Represents the set of pixels in the window, and the window size is 7×7; Represents a pixel within the window; represents the weight coefficient; represents the denoised image; and Represent the denoised image. Relative to Horizontal and vertical gradients; Represents the absolute value operation;

The horizontal gradient and vertical gradient The calculation formulas are as follows:

,

,

in Represents the pixel value at a specific location on the denoised image; and Representing Relative to Horizontal and vertical offsets;

Weight coefficient The calculation formula is as follows:

,

in Represents the exponential operation; and Representing Relative to Horizontal and vertical offsets; represents the smoothing coefficient, set to 0.5;

(1-2) Based on the pixel points in the image As the center, calculate the window intrinsic change in the horizontal direction of the pixel point , the intrinsic change of the window in the vertical direction , and The calculation process is as follows:

,

,

The meaning of each symbol is the same as that of the total window change;

(1-3) Output the denoised image. The process can be expressed by the following formula:

,

in is the original hyperspectral image, is the denoised image; is the smoothing coefficient, set to 0.01; To prevent the denominator from being a constant of 0, it is set to 0.001; the meanings of the other symbols are the same as those of the corresponding symbols above;

(2) Normalizing the pixel values of hyperspectral remote sensing satellite images;

(3) Each pixel of the hyperspectral remote sensing satellite image is cropped into a 7×7 image block with the boundary area filled with 0;

(4) Export the processed image into .GIF format;

(5) Divide all image blocks into training set and validation set in a ratio of 7:3.

The second step is to build a joint model of the total variation regularized guided graph convolution module and the spatial spectral autoencoder module:

The unsupervised hyperspectral image classification model is set to include a spatial spectrum autoencoder module network structure for feature dimension compression and a total variation guided graph convolution module network structure, wherein the features output by the encoder in the spatial spectrum autoencoder module are used as node features for constructing a graph structure for each sample, and the node features are input into the total variation guided graph convolution module after completing graph embedding to complete feature extraction. Finally, in the test phase, after each sample has completed feature extraction, the K-means algorithm is applied to obtain the final classification result; in addition, as shown in FIG2, the constructed spatial spectrum autoencoder module can effectively extract the spatial local context information of the sample and compress the spectral dimension of the sample, significantly reducing the influence of noise and preventing the phenomenon of different spectra of the same object caused by the dimensional disaster, which leads to a decrease in classification accuracy; and as shown in FIG3, the constructed total variation guided graph convolution module can construct a global contextual connection between all samples, and the introduction of the total variation regularization term further eliminates the influence of noise, making the inter-class features sparse and the intra-class features aggregated, and effectively extracting the final features of each sample;

The specific steps are as follows:

(1) Setting the network structure of the spatial spectral autoencoder module, which includes an encoder and a corresponding decoder;

The encoder network structure includes:

The first 2D convolution: the convolution kernel size is 3×3, the stride is 2, and the padding is 0;

The second 2D convolution: the convolution kernel size is 3×3, the stride is 2, and the padding is 0;

The third 2D convolution: the convolution kernel size is 1×1, the stride is 1, and the padding is 0;

The decoder network structure includes:

The first 1D convolution: the convolution kernel size is 1×3, the stride is 1, and the padding is 1;

The second 1D convolution: the convolution kernel size is 1×3, the stride is 1, and the padding is 1;

The third 2D convolution: the convolution kernel size is 1×1, the stride is 1, and the padding is 0;

(2) Set the total variation guided graph convolutional network structure. The specific steps are:

(2-1) Set the graph embedding module structure:

The graph embedding module accepts the features of a batch of samples, and calculates the adjacency matrix through the similarity of the features between the samples to complete the graph topology structure embedding; the features of each node depend on the output result of the encoder in the spatial spectral autoencoder module;

set up is the field of real numbers, The size of is the number of graph nodes, The size of is the encoder output channel dimension size of the empty spectral autoencoder;

The input of the graph embedding module is , the adjacency matrix is calculated by the following formula Value:

,

in represents the smoothing coefficient, which is set to 0.5; Represents the L2 norm operation of the matrix;

(2-2) Set the graph convolution module network structure:

The graph convolution module accepts an undirected graph ,in represents a vertex set, represents the edge set, and the output of the graph embedding module is used as the input of the first graph convolution unit. Represented as module input , is represented as an adjacency matrix ;

Output of the first graph convolutional unit It can be expressed as the following formula:

,

in is the activation function, is the offset; and is the identity matrix; are the learnable parameters of the network; and Calculated by the following formula:

,

The total variation guided graph convolution includes two graph convolution units, so the output of the entire total variation guided graph convolution is It can be expressed by the following formula:

,

in , The meaning and , The meanings of the other symbols are the same as those of the corresponding symbols above;

After a forward propagation in the graph convolution module, the graph total variation regularization loss It can be calculated by the following formula:

,

in is the total variation loss weight, which is set to 0.01 by default; It is a slicing operation on a tensor; the meanings of the other symbols are the same as those of the corresponding symbols above;

(3) Set K-means algorithm parameters:

Divide the given multiple sample features into K clusters, and continuously update the center point of each cluster while dividing; the value of K is set to the number of classification categories;

(3-1) Randomly select K cluster centers, denoted as , and the input of the K-means algorithm is the output of the total variation guided graph convolution module denoted as ;

(3-2) During the iteration process, whether the classification is completed can be judged by whether the loss function converges; the loss function is expressed by the following formula:

,

in It remains unchanged during the iteration process. Indicates that in the iteration process The cluster category to which the sample belongs;

(3-3) Command is the number of iteration steps, where is the maximum number of iterations, set to 1000;

(3-4) For each sample In each iteration, it is assigned to the nearest cluster center, which is expressed by the following formula:

,

(3-5) For each cluster center In each iteration, the value of the cluster center is updated according to the samples belonging to the cluster center, which can be expressed by the following formula:

.

The third step is to train the joint model of the total variation regularized guided graph convolution module and the spatial spectral autoencoder module:

Obtain segmented hyperspectral remote sensing image blocks and input them into the spatial spectrum autoencoder module and the total variation guided graph convolution module for model training;

The specific steps are as follows:

(1) Randomly sample sample points from the processed hyperspectral image and divide it into image blocks of size 7×7×C, and input them into the network as training set and validation set; where C is the number of spectral bands of the hyperspectral image;

(2) A batch of image blocks are input into the spatial spectral autoencoder module. The tensor dimension of the encoder output becomes 1×1×B, which serves as the input of the decoder of the spatial spectral autoencoder module and the graph embedding module of the total variation guided graph convolution module; where B is the number of compressed bands of the encoder;

(3) The decoder receives the output of the encoder as input, upsamples the tensor of dimension 1×1×B to 1×1×C, and calculates the L1 loss with the center sampling point of the sample image block;

(4) The feature vector of size 1×1×B is first subjected to two 1D convolutions with a step size of 1 and a convolution kernel size greater than 1 to extract the local context information of the spectral dimension of the feature vector, and then a 2D convolution with a convolution kernel size of 1×1 and a number of convolution kernels C is performed to achieve spectral dimension upsampling to complete spectral reconstruction;

(5) The reconstructed spectrum and the original sample center point spectrum vector are used to update the parameters of the empty spectrum autoencoder module by calculating the L1 loss. The spectrum reconstruction loss It can be expressed by the following formula:

,

in The size of is the same as the number of graph nodes; Represents the L1 norm operation of the matrix;

(6) The graph embedding module accepts the output of the encoder as input, which compresses the features After the topology graph is constructed, it is input into the graph convolutional network for feature extraction, and the output result is used to calculate the total variation regularization loss of the graph. ;

(7) During the training process, only the spatial spectral autoencoder module and the total variation guided graph convolution module participate in gradient propagation and parameter update; after the network completes a forward propagation, the total loss It is expressed by the following formula:

,

The training of the spatial spectral autoencoder module and the total variation guided graph convolution module can be performed simultaneously, where the spectral reconstruction loss Only used to update the parameters of the empty spectrum autoencoder module, graph total variation regularization loss It is only used to update the parameters of the total variation guided graph convolution module, which can be expressed by the following formulas:

,

in Represents the parameters of the self-space spectrum encoder module;

,

in Represents the total variation guided graph convolution module parameters;

(8) Through total loss Back propagation determines the gradient direction of parameter update, and then updates the parameters to complete an iteration of network training;

(9) When the network training rounds reach the preset number of training rounds, stop the network training and save the model parameters.

The fourth step is to obtain the unsupervised classification results of hyperspectral images:

Obtain the hyperspectral remote sensing satellite image of the area to be classified, input all the divided images into a trained total variation regularized guided graph convolution unsupervised hyperspectral image classification model for forward propagation and obtain the classification map;

The specific steps are as follows:

(1) For a processed hyperspectral image of the area to be classified, all its sample points are divided into image blocks of size 7×7×C;

(2) Define the spatial spectral autoencoder module and the total variation guided graph convolution module, load the trained network parameters and freeze the network parameter updates;

(3) The number of nodes in the graph embedding module is fixed and equal to the batch size of the forward propagation during training. In the network testing process, the forward propagation is performed with the same batch size.

If the total number of image blocks cannot divide the batch size, forward propagation is performed on some samples again, and finally the features extracted twice are averaged;

(4) After obtaining the features of all samples in the area to be classified, use K-means to iteratively cluster the sample features;

(5) Obtain K classification results for all samples in the area to be classified.

The proposed spatial spectrum autoencoder module and total variation guided graph convolution module significantly reduce the impact of noise and prevent the phenomenon of different spectra of the same object caused by the dimensionality disaster, which leads to a decrease in classification accuracy. The global context information is integrated for more effective feature extraction. Compared with the existing unsupervised hyperspectral classification technology, the present invention solves the problem of difficult classification caused by high noise, complex scenes and dimensional redundancy in hyperspectral remote sensing images.

The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions only describe the principles of the present invention. The present invention may be subject to various changes and improvements without departing from the spirit and scope of the present invention. These changes and improvements fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the attached claims and their equivalents.

Claims (5)

1. An unsupervised hyperspectral image classification method based on total variation regularized guided graph convolution, characterized by comprising the following steps: 11) Hyperspectral image denoising preprocessing based on relative total variation: obtain the hyperspectral remote sensing image of the area to be classified, and perform relative total variation denoising on the acquired hyperspectral image; 12) Construct a joint model of the total variation regularized guided graph convolution module and the spatial spectrum autoencoder module: An unsupervised hyperspectral image classification model of total variation regularized guided graph convolution includes three parts: the first is the spatial spectrum autoencoder module for local context extraction and feature dimension compression; the second is the total variation guided graph convolution module for global context extraction and feature fusion; the third is the K-means algorithm module used only in the test process to obtain classification results; 13) Train the joint model of the total variation regularized guided graph convolution module and the spatial spectral autoencoder module: divide the hyperspectral image into image blocks, input them into the joint model to first extract the spatial and spectral context, then use the graph neural network to perform global context modeling, and finally complete the model training through the back propagation algorithm; 14) Obtaining unsupervised classification results of hyperspectral images: Apply the model to the entire hyperspectral image to obtain unsupervised hyperspectral image classification results. 2. The unsupervised hyperspectral image classification method based on total variation regularized guided graph convolution according to claim 1 is characterized in that the acquisition of the hyperspectral image and the denoising preprocessing based on relative total variation include the following steps: 21) Perform relative total variation denoising on the hyperspectral image of the classification area. The specific steps are: 211) Take the pixels in the image As the center, calculate the total change of the horizontal window of the pixel point/> And the total change of the window in the vertical direction/> ,/> and/> The calculation process is as follows: , , in Represents the pixel set in the window, the window size is 7×7;/> Represents a pixel within the window; /> Represents the weight coefficient; /> Represents the denoised image; /> and/> Respectively represent the denoised image /> Relative to /> Horizontal and vertical gradients; /> Represents the absolute value operation; The horizontal gradient and vertical gradient/> The calculation formulas are as follows: , , in Represents the pixel value at a specific position on the denoised image; /> and/> Respectively represent/> Relative to /> Horizontal and vertical offsets; Weight coefficient The calculation formula is as follows: , in Represents the index operation; /> and/> Respectively represent/> Relative to /> Horizontal and vertical offsets; /> represents the smoothing coefficient, set to 0.5; 212) Based on the pixels in the image As the center, calculate the window inherent change amount of the pixel in the horizontal direction/> , the window's inherent change in the vertical direction/> ,/> and/> The calculation process is as follows: , , The meaning of each symbol is the same as that of the total window change; 213) Output the denoised image, the process can be expressed by the following formula: , in is the original hyperspectral image, /> is the denoised image; /> is the smoothing coefficient, set to 0.01; /> To prevent the denominator from being a constant of 0, it is set to 0.001; the meanings of the other symbols are the same as those of the corresponding symbols mentioned above. 3. The unsupervised hyperspectral image classification method of total variation regularized guided graph convolution according to claim 1 is characterized in that the joint model of constructing the total variation regularized guided graph convolution module and the spatial spectrum autoencoder module comprises the following steps: 31) The unsupervised hyperspectral image classification model is set to include a spatial spectrum autoencoder module and a total variation guided graph convolution module, where the features output by the encoder in the spatial spectrum autoencoder are used as node features for each sample to construct a graph structure. After the node features are embedded in the graph, they are input into the graph convolution module to complete feature extraction. Finally, in the test phase, after each sample has completed feature extraction, the K-means algorithm is applied to obtain the final classification result; 32) Set the network structure of the spatial spectral autoencoder, which includes an encoder and a corresponding decoder; The encoder network structure includes: The first 2D convolution: the convolution kernel size is 3×3, the stride is 2, and the padding is 0; The second 2D convolution: the convolution kernel size is 3×3, the stride is 2, and the padding is 0; The third 2D convolution: the convolution kernel size is 1×1, the stride is 1, and the padding is 0; The decoder network structure includes: The first 1D convolution: the convolution kernel size is 1×3, the stride is 1, and the padding is 1; The second 1D convolution: the convolution kernel size is 1×3, the stride is 1, and the padding is 1; The third 2D convolution: the convolution kernel size is 1×1, the stride is 1, and the padding is 0; 33) Set the total variation guided graph convolutional network structure. The specific steps are: 331) Set the graph embedding module structure: The graph embedding module accepts the features of a batch of samples, and calculates the adjacency matrix through the similarity of the features between the samples to complete the graph topology structure embedding; the features of each node depend on the output result of the encoder in the spatial spectral autoencoder module; set up is the field of real numbers, /> The size of is the number of graph nodes, /> The size of is the encoder output channel dimension size of the empty spectral autoencoder; The input of the graph embedding module is , the adjacency matrix is calculated by the following formula Value: , in Represents the smoothing coefficient, set to 0.5; /> Represents the L2 norm operation of the matrix; 332) Set the graph convolution module network structure: The graph convolution module accepts an undirected graph , where/> represents a vertex set, /> represents the edge set; the output of the graph embedding module is used as the input of the first graph convolution unit, then/> Represented as module input/> ,/> is represented as an adjacency matrix/> ; Output of the first graph convolutional unit It can be expressed as the following formula: , in is the activation function, /> is the offset; /> And/> is the identity matrix; /> is the network learnable parameter; /> And/> Calculated by the following formula: , The total variation guided graph convolution includes two graph convolution units, so the output of the entire total variation guided graph convolution module is It can be expressed by the following formula: , in is the offset, /> is the network learnable parameter; the meanings of the other symbols are the same as those of the corresponding symbols mentioned above; After a forward propagation in the graph convolution module, the graph total variation regularization loss It can be calculated by the following formula: , in is the total variation loss weight, set to 0.01;/> It is a slicing operation on a tensor; the meanings of the other symbols are the same as those of the corresponding symbols above; 34) Set the K-means algorithm parameters. The specific steps are: The core goal of K-means is to divide a given data set into K clusters and continuously update the center point of each cluster while dividing; the value of K is set to the number of classification categories; 341) Randomly select K cluster centers, denoted as , and the input of the K-means algorithm is the output of the total variation guided graph convolution module denoted as/> ; 342) K-means is an iterative algorithm. During the iteration process, the classification can be judged by whether the loss function converges. The loss function is expressed by the following formula: , The total variation guided graph convolution module outputs It remains unchanged during the iteration process. Indicates that in the iteration process /> The cluster category to which the sample belongs; 343) Order is the number of iteration steps, where /> is the maximum number of iterations, set to 1000; 344) For each sample In each iteration, it is assigned to the nearest cluster center, which is expressed by the following formula: , 345) For each cluster center In each iteration, the value of the cluster center is updated according to the samples belonging to the cluster center, which can be expressed by the following formula: . 4. The unsupervised hyperspectral image classification method of total variation regularized guided graph convolution according to claim 1 is characterized in that the joint model of training the total variation regularized guided graph convolution module and the spatial spectrum autoencoder module comprises the following steps: 41) Randomly sample sample points from the processed hyperspectral image and divide it into image blocks of size 7×7×C, and input them into the network according to the training set and the validation set; where C is the number of spectral bands of the hyperspectral image; 42) Input a batch of image blocks into the spatial spectrum autoencoder module, and each image block of size 7×7×C is converted into a feature vector of size 1×1×B through the encoder; the image block of size 7×7×C first undergoes two 2D convolutions with a step size greater than 1 to extract the local context information of the spatial dimension of the central sample and implement the downsampling process; the changes in length and width of the input through each 2D convolution module can be expressed by the following formulas: , in Extract the local context information of the spatial dimension of the center sample for the product and implement it,/> is the vertical size of the output; /> 、/> 、/> They represent the vertical convolution kernel size, vertical padding size, and vertical convolution step size of the 2D convolution operation respectively; , in is the horizontal size of the input, /> is the horizontal size of the output; /> 、/> 、/> They represent the lateral convolution kernel size, lateral padding size, and lateral convolution step size of the 2D convolution operation respectively; The tensor dimension of the encoder output becomes 1×1×B, which serves as the input of the decoder of the spatial spectral autoencoder module and the graph embedding module; where B is the number of bands after the encoder compression; 421) The decoder receives the output of the encoder as input, upsamples the tensor of dimension size 1×1×B to 1×1×C, and calculates the L1 loss with the center sampling point of the sample image block; 4211) The feature vector of size 1×1×B is first subjected to two 1D convolutions with a step size equal to 1 and a convolution kernel size greater than 1 to extract the local context information of the spectral dimension of the feature vector, and then the spectral dimension is upsampled by a 2D convolution with a convolution kernel size of 1×1 and a number of convolution kernels C to complete the spectral reconstruction; 4212) The reconstructed spectrum and the original sample center point spectrum vector are updated by calculating the L1 loss of the empty spectrum autoencoder module parameters, and the spectrum reconstruction loss It can be expressed by the following formula: , in The size is the same as the number of graph nodes; /> Represents the L1 norm operation of the matrix; 422) The graph embedding module takes the output of the encoder as input, which compresses the features After the topology graph is constructed, it is input into the graph convolutional network for feature extraction, and the output result is used to calculate the total variation regularization term loss of the graph. ; 43) During the training process, only the Spatial Spectral Autoencoder module and the Total Variation Guided Graph Convolution module participate in gradient propagation and parameter update. After the network completes a forward propagation, the total loss It is expressed by the following formula: , The training of the spatial spectral autoencoder module and the total variation guided graph convolution module can be performed simultaneously, where the spectral reconstruction loss Only used to update the parameters of the empty spectrum autoencoder module, graph total variation regularization loss/> It is only used to update the parameters of the total variation guided graph convolution module, which can be expressed by the following formulas: , in Represents the parameters of the self-space spectrum encoder module; , in Represents the total variation guided graph convolution module parameters; 44) By total loss Back propagation determines the gradient direction of parameter update, and then updates the parameters to complete an iteration of network training; 45) When the network training rounds reach the preset number of training rounds, stop the network training and save the model parameters. 5. The unsupervised hyperspectral image classification method based on total variation regularized guided graph convolution according to claim 1, characterized in that obtaining the unsupervised classification result of the hyperspectral image comprises the following steps: 51) For a processed hyperspectral image to be classified, all its sample points are divided into image blocks of size 7×7×C; 52) Define the spatial spectrum autoencoder module and the total variation guided graph convolution module, load the trained network parameters and freeze the network parameter update; 53) In the graph embedding module, the number of nodes is fixed and equal to the batch size of the forward propagation during training, and the forward propagation is performed with the same batch size during the network testing process; If the total number of image blocks cannot divide the batch size, forward propagation is performed on some samples again, and the features extracted twice are finally averaged; 54) After obtaining the features of all samples to be classified, use K-means to iteratively cluster the sample features; 55) Obtain K classification results for all samples to be classified.