A fast hyperspectral image classification method based on dynamic graph convolutional network

Through dynamic graph convolutional neural network combined with multiple branches of neural networks, the problem of long training time and inference time in hyperspectral remote sensing image classification is solved, efficient image classification and large-picture processing are achieved, and classification accuracy and inference speed are improved. .

CN114937173BActive Publication Date: 2024-11-26CHINA UNIV OF GEOSCIENCES (WUHAN)
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210535223.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-05-17
Publication Date
2024-11-26
Estimated Expiration
2042-05-17

AI Technical Summary

Technical Problem

In the existing hyperspectral remote sensing image classification methods, the training time and inference time of the deep learning model are long, and the graph convolutional network consumes too much memory on large images and cannot be effectively applied to remote sensing images.

Method used

Using a method based on dynamic graph convolutional neural network, combining neural networks with multiple branches for hyperspectral image classification, including ordinary convolutional neural networks, convolutional neural networks with discarded blocks, and dynamic graph convolutional networks, through feature development Flat, global average pooling and node encoder extract features and iteratively trained using cross-entropy loss functions.

Benefits of technology

It realizes efficient hyperspectral image classification, greatly improves the inference speed, can process large images, and has strong generalization capabilities for modeling.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114937173B_ABST
    Figure CN114937173B_ABST
Patent Text Reader

Abstract

The present invention provides a method for rapid classification of hyperspectral images based on a dynamic graph convolutional network, comprising: S1: obtaining a hyperspectral remote sensing image #imgabs0# and its corresponding label Y, wherein H, W, and B represent the length, width, and height of the hyperspectral remote sensing image, respectively; S2: performing data preprocessing on the hyperspectral remote sensing image X, and dividing the preprocessed image into pixel blocks to obtain a true label corresponding to each pixel block; S3: inputting the pixel blocks processed in step S2 into a neural network composed of multiple branches; S4: after splicing the network outputs of multiple branches, converting them into category probability information through a Softmax layer, selecting the maximum category probability as the predicted label of the pixel block, and then comparing it with the true label, calculating the loss, and obtaining a fusion classification model after multiple iterations. The beneficial effects of the present invention are: the present invention fully considers the multiple representations of samples, comprehensively considers each level of features, and improves the classification accuracy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and in particular to a method for fast classification of hyperspectral images based on a dynamic graph convolutional network. Background Art

[0002] Hyperspectral remote sensing image processing plays an important role in material information detection and is an important topic in the field of remote sensing. Hyperspectral remote sensing images can provide continuous radiation spectral bands, carry rich ground information, and can be used to process different application fields, such as house change detection, crop assessment, geological and mineral resource surveys, etc. (He, Li et al., 2018; Ahmad; Shabbir et al., 2022). In the field of hyperspectral remote sensing image processing, classification is one of the main tasks. The so-called hyperspectral image classification is to determine the category of each hyperspectral image pixel. Existing classification methods are mainly based on deep learning methods, that is, using models such as convolutional neural networks and recurrent neural networks for end-to-end image classification.

[0003] As a deep learning model, convolutional neural network (Liang, Li, 2016; Hamida, Benoit et al. 2018) has been successfully applied to hyperspectral remote sensing image classification. The method based on convolutional neural network can process high-dimensional data and has good robustness to sample noise. However, the convolutional network based on deep feature extraction has a sharp increase in training time and inference time as the number of layers increases and a large number of trainable parameters increase. Graph convolutional network (Ye, He et al. 2020) can make full use of the topological structure information in hyperspectral images and achieve the same accuracy as multi-layer convolutional neural network with fewer layers. However, in general, graph convolutional neural network calculates the adjacency matrix on the entire graph in a semi-supervised manner, which consumes a lot of memory resources and cannot be applied to large graphs such as remote sensing images.

[0004] With the help of the hyperspectral image pixel block division strategy, based on the graph convolutional network, the hyperspectral image classification can be quickly performed in a supervised learning manner. However, due to the diversity of pixel blocks caused by pixel-level block division, the classification results will be affected to a certain extent. Considering that shallow convolutional networks can also extract useful features, the present invention combines the graph convolutional network and convolutional neural network features, introduces an effective method based on dynamic graph convolutional neural network, and aims at the problems of the existing methods, combined with label smoothing and discard block regularization strategies, to better complete the hyperspectral image classification task. Summary of the invention

[0005] In order to solve the above problems, the present invention provides a high-precision and high-efficiency hyperspectral image fast classification method based on dynamic graph convolutional neural network. The hyperspectral image is a data cube and can be represented by a tensor. The vector of each point in the tensor is the pixel radiation value corresponding to each band. The method mainly includes the following steps:

[0006] S1: Get the hyperspectral remote sensing image X and its corresponding label Y; Labels for ground reference images Where H, W, B and C represent the length, width, height and total number of categories of the hyperspectral image, respectively;

[0007] S2: Perform data preprocessing on the hyperspectral remote sensing image X, that is, use the principal component analysis method to reduce the dimension of the hyperspectral image, and divide each pixel into pixel blocks to obtain the true label corresponding to each pixel block;

[0008] S3: Input the pixel block obtained in step S2 into a neural network consisting of multiple branches;

[0009] S4: After splicing the network outputs of multiple branches, they are converted into category probability information through the Softmax layer. The largest category probability is selected as the predicted label of the pixel block, which is then compared with the true label and the loss is calculated. After multiple iterations, a fusion classification model is obtained. The hyperspectral remote sensing image is input into the fusion classification model to finally obtain the classification result.

[0010] Step S2 is specifically as follows:

[0011] Reduce the dimensionality of a hyperspectral image to a specified dimension:

[0012] X=QΛQ -1

[0013] Among them, Q is the matrix composed of the eigenvectors of this matrix X, Q -1 represents the inverse matrix of Q. Λ is a diagonal matrix. Each diagonal element is an eigenvalue, and the eigenvalues ​​are arranged in order of small. Select the largest eigenvalues ​​and perform feature dimensionality reduction.

[0014] The neural network composed of the multiple branches is a neural network composed of two convolutional neural network branches with discard blocks and one dynamic graph convolutional network branch. Step S3 is specifically as follows:

[0015] (1) First, ordinary convolution is used to extract primary features, and then the extracted primary features are passed to two 2D convolutional neural networks with discard blocks for feature extraction, which can be expressed as:

[0016]

[0017] Where * refers to the convolution operator. The matrix and They are the feature maps of the current layer and the next layer respectively. and denote the filter and bias of the (l+1)th convolutional layer respectively. σ(·) denotes the LeakyReLU activation function. The function of the dropout block is to randomly convert part of the values ​​of the extracted features to 0.

[0018] According to the above principle, after passing through the first branch and the second branch of the neural network, the classification features O1 and O2 are obtained, as follows:

[0019] In the first branch, feature flattening operation is used to obtain classification feature O1:

[0020] O1=Flatten(h1)

[0021] Where h1 is the convolution feature obtained by the first branch, and Flatten(·) is the flattening operation.

[0022] In the second branch, the feature global average pooling operation is used to obtain the classification feature O2:

[0023] O1=GAP(h2)

[0024] Where h2 is the convolution feature obtained by the second branch, and GAP(·) is the global average pooling operation.

[0025] (2) The node encoder is used to process the primary features obtained after the 2D convolutional neural network is processed, and the node features are passed into the third branch of the neural network. First, the convolutional neural network calculates the activation map of a specific category based on the feature map of the hidden layer. Then use them to convert the feature map h of the hidden layer after transformation into a node sequence representation

[0026]

[0027] in It represents the transposition of the feature map h after one layer of convolution. Represents the deformed feature vector.

[0028] Graph Convolutional Layer:

[0029] The node embedding features are calculated using the hyperspectral pixel node encoding feature vector:

[0030] V'=σ(A f VW f )

[0031] The adjacency matrix Records the relationship between each node, A f and W f is initialized randomly.

[0032] σ(·) is the LeakyReLU activation function, and V’ is the node embedding feature after being encoded by a layer of graph convolutional network.

[0033]

[0034] Where η is the slope coefficient.

[0035] Compute a dynamic adjacency matrix using hyperspectral pixels:

[0036]

[0037]

[0038] where α ji To obtain attention information, which represents the influence of the i-th dimension on the j-th dimension in the hidden layer, the feature vectors E, F and G are obtained by transforming the node feature V through the fully connected network. K is the total number of dimensions of the node encoding feature vector V, and β is a learnable parameter. Represents a dynamically updated adjacency matrix.

[0039] Dynamic graph convolution layer:

[0040] Calculate the second-layer node embedding features based on the dynamic adjacency matrix:

[0041] V"=σ(A s V'W s )

[0042] Where W s represents a learnable parameter. σ(·) represents the LeakyReLU activation function, and V” is the node embedding feature after being encoded by a layer of graph convolutional network.

[0043] (3) Use the fusion classification model shown below to perform feature fusion:

[0044] O=O1||O2||O3

[0045] The || operator represents the concatenation of two features along the spectral dimension, O1, O2 and O3 are the classification features generated after two convolutional neural network branches with discard blocks and dynamic graph convolutional network branches, respectively, and O3 = V".

[0046] The entire network is iteratively trained using the cross entropy loss function and the back propagation function to obtain a trained fusion classification model. The cross entropy loss function is used to calculate the loss:

[0047]

[0048] Among them, U is the number of samples in the mini-batch, C is the total number of categories in the dataset, and y c and are the ground truth reference image and predicted label of class c, respectively.

[0049] The beneficial effects brought by the technical solution provided by the present invention are:

[0050] 1. A deep learning framework based on dynamic graph convolutional neural network is proposed to realize HSI classification by integrating dynamic graph convolutional network and convolutional neural network. The framework uses three branches to process HSI in parallel to achieve fast information extraction.

[0051] 2. We propose a novel dynamic graph convolutional network module that can adaptively capture content-aware category dependencies through supervised learning and extend graph convolutional networks to large graphs.

[0052] 3. The proposed network framework outperforms the most advanced convolutional neural networks and significantly surpasses current deep learning and machine learning models in inference speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0054] Figure 1 It is a flow chart of a method for rapid classification of hyperspectral images based on a dynamic graph convolutional network in an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.

[0056] An embodiment of the present invention provides a method for rapid classification of hyperspectral images based on a dynamic graph convolutional network. The method first obtains two classification features by flattening and pooling the results of the convolution layer, which makes full use of the spatial-spectral information contained in the hyperspectral data cube. Then the dynamic graph convolution module is applied to extract the intrinsic structural information of each pixel block. Finally, the method can obtain the classification results of the hyperspectral image based on these spatial, spectral and structural features. By using three branches, the method can simultaneously process multiple features of the hyperspectral image in a supervised learning manner and quickly complete the hyperspectral image classification. In addition, the method also applies regularization techniques such as discarding blocks and label smoothing to further improve the generalization ability of the model.

[0057] The specific implementation of the present invention is written in PYTHON language and implemented using the classic deep learning framework PYTORCH, with PYTHON remote sensing image reading and writing functions as the implementation basis. Call the data processing library NUMPY, SCIPY and SPECTRAL library, input the remote sensing image file name to be read, and the remote sensing image is read into a tensor of size H×W×B. Each element in the tensor is the pixel radiation value corresponding to each band, where H is the length of the remote sensing image, W is the width of the remote sensing image, and B is the number of remote sensing image bands. PYTHON remote sensing image reading and writing functions are well-known technologies in the technical field and will not be described in detail here.

[0058] like Figure 1 As shown in the figure, the hyperspectral image classification based on dynamic graph convolutional neural network is performed as follows:

[0059] (1) For the hyperspectral image X, since it consists of hundreds of very narrow continuous spectra and contains rich information to be mined, we first use a two-dimensional convolutional neural network to process the original hyperspectral image and extract primary features, that is, using:

[0060]

[0061] Where * refers to the convolution operator, the matrix and are the feature maps of the current layer and the next layer respectively. and They represent the filter and bias vector of the (l+1)th convolutional layer, and σ(·) represents the LeakyReLU activation function. Through the above operations, the hidden features are obtained for the first time. The first hidden feature is obtained from the original feature, that is, the input data X. Based on the knowledge of deep learning, h1 and h2 are obtained. The following is a schematic expression:

[0062] h1=X*w

[0063] h2=h1*w=X*w*w

[0064] h3=h2*w=X*w*w*w ......

[0066] Where X is the input feature and w is the learnable parameter.

[0067] (2) In order to further explore the rich features contained in the hyperspectral image, the features extracted by the initial two convolutional layers are respectively passed to the three branches, as described in (1), and the three branches are processed simultaneously. The first and second branches are composed of convolutional layers, in which the dropout block regularization strategy is inserted in the convolution module. The function of the dropout block is to randomly convert part of the values ​​of the extracted features to 0.

[0068] (3) In the first branch, feature flattening operation is used to obtain classification feature O1:

[0069] O1=Flatten(h1)

[0070] Where h1 is the convolution feature obtained by the first branch, and Flatten(·) is the flattening operation.

[0071] (4) In the second branch, the feature global average pooling operation is used to obtain the classification feature O2:

[0072] O1=GAP(h2)

[0073] Where h2 is the convolution feature obtained by the second branch, and GAP(·) is the global average pooling operation.

[0074] (5) Node Encoder: The third branch consists of two layers. First, the convolutional neural network calculates the activation map of a specific category based on the feature map of the hidden layer. Then use them to transform the feature maps of the hidden layer Convert to node sequence representation

[0075]

[0076] in It represents the transposition of the feature map h after one layer of convolution. It represents the feature vector after the hidden feature h is deformed. The data organization methods of the two are different. It can be understood that h is the horizontal data and h' is the vertical data. For example, if h is H*W*D, the deformed h' is (H*W)*D, where the first two dimensions are expanded into one column.

[0077] (6) Using the vertex representation V, the first layer performs conventional graph convolution operations. The goal of the graph convolutional layer is to use the adjacency matrix and learnable parameters W f To get its value V'. In specific form, this process can be expressed as:

[0078] V'=σ(A f VW f )

[0079] The adjacency matrix The relationship between each node is recorded. f and W f is randomly initialized and learned by gradients during training. σ(·) is the LeakyReLU activation function.

[0080] (7) For the second layer, introduce the adjacency matrix As To update node V'. is obtained by encoding V, which is different from the first layer. f is fixed after training. s It can be dynamically updated as the input features change.

[0081]

[0082]

[0083] The feature vectors E, F and G are generated by the hidden layer feature h' through the fully connected network, and β is a learnable parameter. Therefore, each pixel block will produce a different A s , which greatly enhances the expressiveness of the model and reduces the risk of overfitting. Specifically, the V” output can be defined as:

[0084] V"=σ(A s V'W s )

[0085] Among them A s and W s They represent the dynamically updated adjacency matrix and learnable parameters respectively, and σ(·) represents the LeakyReLU activation function.

[0086] (8) In order to fully explore the information contained in the hyperspectral image, three branches are used to extract the hyperspectral image features from different angles. Figure 1 As shown, the outputs of the three branches are O1, O2 and O3 respectively. In order to unify the representation, the feature V" obtained by the graph convolution layer is represented by O3. The features obtained from these three different angles are fused and then classified. The final classification feature O output can be expressed as:

[0087] O=O1||O2||O3

[0088] where the || operator represents the concatenation of two features along the spectral dimension.

[0089] (9) In order to optimize the proposed FDGC, the cross entropy loss function is used:

[0090]

[0091] Where U is the number of samples in the mini-batch, C is the total number of categories in the dataset, and y c and They are the ground truth reference image and predicted label of class c respectively. Under this classification framework, the best model is obtained through continuous iteration.

[0092] The Indian pine and Salinas hyperspectral datasets were used as experimental datasets to verify the effectiveness of the model. The Indian pine was imaged by the Airborne Visible Infrared Imaging Spectrometer (AVIRIS) in 1992 in Indiana, USA, and then the image was cut into 145×145 pieces and labeled for hyperspectral image classification testing. The AVIRIS imaging spectrometer has an imaging wavelength range of 0.4-2.5μm, and it continuously images the ground objects in 220 continuous bands. However, since the 104-108, 150-163 and 220 bands cannot be reflected by water, the 200 bands remaining after removing these 20 bands are generally used as the research object. The data has a total of 21,025 pixels, but only 10,249 of them are ground object pixels, and the remaining 10,776 pixels are background pixels. In actual classification, these pixels need to be eliminated. Since the intercepted area is all crops, there are a total of 16 categories, so different ground objects have relatively similar spectral curves, and the distribution of samples in these 16 categories is extremely uneven. Like the Indian pine image, the Salinas data is also taken by the AVIRIS imaging spectrometer. It is an image of the Salinas Valley in California, USA. Unlike the Indian pine dataset, its spatial resolution reaches 3.7m. The image originally also has 224 bands. Similarly, the image generally used is the 204 bands left after eliminating the 108-112, 154-167, and 224 bands that cannot be reflected by water. The size of the image is 512×217, so it contains a total of 111104 pixels, of which 56975 pixels are background pixels, and 54129 pixels can be applied to classification. These pixels are divided into 16 categories in total, including fallow land (Fallow), celery (Celery), etc. The present invention uses the classic support vector machine (method 1), 3D convolutional neural network (method 2), deep convolutional network (method 3), spatial-spectral joint residual network (method 4), deep feature fusion network (method 5), hybrid spatial-spectral network (method 6), spectral partition residual network (method 7) and the present invention method for classification comparison to prove the effectiveness of the present invention.

[0093] Classification evaluation index: Using quantitative evaluation method, a total of 1260 pixels were selected as training samples and 8989 pixels were selected as test samples in data set 1, and a total of 1600 pixels were selected as training samples and 52529 pixels were selected as test samples in data set 2. The evaluation index uses the following two indicators:

[0094] 1) Kappa coefficient:

[0095] The kappa coefficient is an authoritative evaluation indicator for evaluating classification problems. The larger the kappa coefficient, the higher the accuracy. In this experiment, the highest kappa coefficient that can be obtained by methods 1 to 3 was selected to evaluate the detection capabilities of methods 1 to 7.

[0096] The sample acquisition confusion matrix is ​​shown in Table 1:

[0097] Table 1 Confusion matrix

[0098]

[0099] In Table 1, TTO indicates the number of cases whose labels are category 1 and are predicted to be category 1; TF indicates the number of cases whose actual category is category 2 but are predicted to be category 1. FT indicates the number of cases whose actual category is category 1 but are predicted to be category 2. TTT indicates the number of cases whose labels are category 2 and are predicted to be category 2; NCO is the sum of TTO and FT, NCT is the sum of TF and TTT, NRO is the sum of TTO and TF, NRT is the sum of FT and TTT, and N is the total number of samples.

[0100] Obtain the confusion matrix based on the sample and calculate the Kappa coefficient. The calculation formula is:

[0101]

[0102] 2) Overall accuracy:

[0103] Overall Accuracy (OA) is an evaluation index used to evaluate classification problems. The higher the overall accuracy, the higher the detection accuracy. The calculation method of OA is also based on the confusion matrix shown in Table 1. The calculation formula of overall accuracy OA is:

[0104]

[0105] The two evaluation indicators of kappa coefficient and overall accuracy were used to compare the classification ability of the method of the present invention with the comparison experiments of 1 to 7. The evaluation indicators are shown in Table 2:

[0106] Table 2 Comparative test results 1

[0107]

[0108]

[0109] As can be seen from Table 2, the method of the present invention can obtain higher overall accuracy and kappa value, indicating that the method of the present invention has stronger classification ability. The method of the present invention and the comparative method have achieved the best results in terms of overall classification accuracy and kappa value. It can be concluded that the method of the present invention has higher classification accuracy than traditional and recent deep learning classification methods. The present invention fully considers the multiple representations of samples and comprehensively considers all levels of features to improve classification accuracy.

[0110] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for fast classification of hyperspectral images based on dynamic graph convolutional networks, characterized by: The following steps are involved: S1: Acquisition of hyperspectral remote sensing images and its corresponding label Y, where H, W, and B represent the length, width, and height of the hyperspectral remote sensing image, respectively; S2: Perform data preprocessing on the hyperspectral remote sensing image X, and divide the preprocessed image into pixel blocks to obtain a number of pixel blocks and a true label corresponding to each pixel block; S3: Input the pixel block processed in step S2 into a neural network consisting of multiple branches; The neural network composed of the multiple branches is a neural network composed of two convolutional neural network branches with discard blocks and one dynamic graph convolutional network branch. The specific implementation process of step S3 is: (1) First, ordinary convolution is used to extract primary features, and then the extracted primary features are passed to two 2D convolutional neural networks with discard blocks for feature extraction, which can be expressed as: Where * refers to the convolution operator, the matrix and are the feature maps of the current layer and the next layer respectively. and They represent the weight and bias of the (l+1)th convolutional layer, σ(·) represents the LeakyReLU activation function; the function of the dropout block is to randomly convert part of the values ​​of the extracted features to 0; According to the above principle, after passing through the first branch and the second branch of the neural network, classification features O1 and O2 are obtained; (2) The primary features obtained after processing by the 2D convolutional neural network are passed to the third branch of the neural network. First, the neural network calculates the activation map of a specific category based on the feature map of the hidden layer. Then, the feature map h of the transformed hidden layer is converted into a node encoding feature vector according to the activation map of the specific category in It represents the transposition of the feature map h after one layer of convolution. represents the deformed feature vector; (3) Calculate node embedding features using hyperspectral pixel node encoding feature vectors: V'=σ(A f VW f ) The adjacency matrix Records the relationship between each node, W f represents the randomly initialized weight; σ(·) is the LeakyReLU activation function, and V' is the node embedding feature after being encoded by a layer of graph convolutional network; in η is the slope coefficient; (4) Calculate the dynamic adjacency matrix using hyperspectral pixels: where α ji To obtain attention information, which represents the influence of the i-th dimension on the j-th dimension in the hidden layer, the feature vectors E, F and G are obtained by transforming the node feature V through the fully connected network. K is the total number of dimensions of the node encoding feature vector V, and β is a learnable parameter. Represents a dynamically updated adjacency matrix; (5) Calculate the second-layer node embedding features based on the dynamic adjacency matrix: V”6σ(A s V'W s ) Where W s represents the learnable parameter, σ(·) represents the LeakyReLU activation function, and V” is the node embedding feature after being encoded by a layer of graph convolutional network; (6) Perform feature concatenation on the three obtained features to obtain the final classification feature: O=O1||O2||O3 The || operator represents the concatenation of two features along the spectral dimension. O1, O2 and O3 are the classification features generated by two convolutional neural network branches with discard blocks and dynamic graph convolutional network branches, respectively. O3 = V”; S4: After splicing the network outputs of multiple branches, they are converted into category probability information through the Softmax layer. The largest category probability is selected as the predicted label of the pixel block, which is then compared with the true label and the loss is calculated. After multiple iterations, a fusion classification model is obtained. The hyperspectral remote sensing image is input into the fusion classification model to finally obtain the classification result.

2. The method for rapid classification of hyperspectral images based on dynamic graph convolutional networks according to claim 1, characterized in that: In step S2, the data preprocessing refers to the dimensionality reduction processing of the hyperspectral remote sensing image using the principal component analysis method: X=QΛQ -1 Among them, Q is the matrix composed of the eigenvectors of the hyperspectral remote sensing image matrix X, Q -1 Represents the inverse matrix of Q, Λ is a diagonal matrix, each diagonal element is an eigenvalue, the eigenvalues ​​are arranged from small to large, and the largest eigenvalues ​​are selected to perform feature dimensionality reduction.

3. The method for rapid classification of hyperspectral images based on dynamic graph convolutional networks according to claim 1, characterized in that: In the first branch, feature flattening operation is used to obtain classification feature O1: O1=Flatten(h1) Where h1 is the convolution feature obtained by the first branch, and Flatten(·) is the flattening operation.

4. The method for rapid classification of hyperspectral images based on dynamic graph convolutional networks according to claim 3, characterized in that: In the second branch, the feature global average pooling operation is used to obtain the classification feature O2: O1=GAP(h2) Where h2 is the convolution feature obtained by the second branch, and GAP(·) is the global average pooling operation.

5. The method for rapid classification of hyperspectral images based on dynamic graph convolutional networks according to claim 1, characterized in that: The entire network is iteratively trained using the cross entropy loss function and back propagation function to obtain a trained fusion classification model.

6. The method for rapid classification of hyperspectral images based on dynamic graph convolutional networks according to claim 5, characterized in that: The loss is calculated using the cross entropy loss function: Where U is the number of samples in the mini-batch, C is the total number of categories in the dataset, and y c and are the ground truth reference image and predicted label of the cth class respectively, and L is the loss value obtained after one round of training.