Hyperspectral image classification method based on adaptive graph convolution and multi-scale feature fusion

Abstract

The purpose of this invention is to provide a hyperspectral image classification method based on adaptive graph convolution and multi-scale feature fusion, specifically implemented according to the following steps: Step 1: Collect hyperspectral images to construct a hyperspectral dataset, which is then divided into a training set, a validation set, and a test set; Step 2: Construct a dual-branch feature fusion classification network model combining CNN and AGCN; Step 3: Input the training set from Step 1 into the classification network model constructed in Step 2, set the parameters, and perform training; Step 4: Obtain the hyperspectral image land cover classification results, and evaluate the indicators, data, and image information of the classification results. This invention solves the problem in existing hyperspectral image classification methods that cannot fully extract subtle features, thus limiting classification accuracy.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing technology, specifically involving a hyperspectral image classification method based on adaptive graph convolution and multi-scale feature fusion. Background Technology

[0002] Hyperspectral imagery, a crucial product of remote sensing technology, is acquired through observations by remote sensing satellites and contains rich spectral and spatial information. Spectrally, it captures the reflection, absorption, and emission characteristics of ground features across different spectral bands, forming continuous, high-precision spectral curves that serve as unique evidence for feature identification. Spatially, it accurately presents the location, shape, size, and interrelationships of ground features, constructing a spatial distribution map of the Earth's surface. Hyperspectral imagery plays a vital role in multiple fields. In military reconnaissance, it helps discover hidden military targets and provides intelligence support; in environmental monitoring, it monitors air, water, and soil pollution, identifies pollutants, and understands the extent and trends of pollution; in vegetation surveys, it helps understand vegetation growth, species distribution, and health status; and in geological exploration, it assists in identifying rocks and minerals and inferring mineral distribution.

[0003] However, hyperspectral image applications face numerous challenges. The sheer volume of data, stemming from high-resolution imaging and numerous spectral bands, places high demands on storage and transmission equipment, increasing the complexity and time cost of data processing. Simultaneously, it possesses high dimensionality, complex data space, and significant data redundancy issues. Correlation between adjacent bands leads to information duplication, wasting storage space and interfering with subsequent processing. Traditional classification methods, such as those based on statistical models, perform poorly in processing hyperspectral images. These methods, based on simple assumptions and models, struggle to extract complex spectral and spatial structural information, and are prone to the "curse of dimensionality" when faced with high dimensionality and data redundancy, resulting in high computational costs and decreased classification accuracy. In recent years, deep learning technology has achieved significant results in image classification, automatically learning effective feature representations and improving classification accuracy. However, it still has shortcomings in hyperspectral image classification. Most methods are pixel-based, ignoring spatial relationships between pixels and overall image structural information, lacking a holistic understanding of ground features in the image. Some methods employ simple convolutional neural networks (CNNs), but due to the complex spectral features of hyperspectral images, they cannot fully extract subtle features, limiting classification accuracy. Summary of the Invention

[0004] The purpose of this invention is to provide a hyperspectral image classification method based on adaptive graph convolution and multi-scale feature fusion, which solves the problem that existing hyperspectral image classification methods cannot fully extract subtle features, thus limiting classification accuracy.

[0005] The technical solution adopted in this invention is a hyperspectral image classification method based on adaptive graph convolution and multi-scale feature fusion, which is implemented according to the following steps:

[0006] Step 1: Collect hyperspectral images to construct a hyperspectral dataset. The constructed hyperspectral dataset is divided into a training set, a validation set, and a test set.

[0007] Step 2: Construct a dual-branch feature fusion classification network model combining CNN and AGCN;

[0008] Step 3: Input the training set divided in Step 1 into the classification network model constructed in Step 2, set the parameters, and train it;

[0009] Step 4: Obtain the hyperspectral image land cover classification results, and evaluate the indicators, data, and image information of the classification results.

[0010] The invention is further characterized in that,

[0011] Step 1 is implemented in the following steps:

[0012] Using the hyperspectral dataset as input, preprocessing is performed by dividing it into training, testing, and validation sets. The training set is used for model training, allowing the model to learn the mapping relationship between the spectral-spatial features of hyperspectral images and land cover categories. The validation set is used to adjust hyperparameters during model training to prevent overfitting, and the optimal model parameters are selected by evaluating model performance on the validation set. The testing set is used to objectively evaluate the model's generalization ability and classification performance after training, obtaining the final classification accuracy index.

[0013] In step 2, the dual-branch feature fusion classification network model combining CNN and AGCN includes two branches, CNN and AGCN, as well as a feature fusion module. The CNN branch consists of three parts: data preprocessing, multi-scale residual convolutional blocks, and spatial attention mechanism. The AGCN branch also consists of three parts: data preprocessing, adaptive graph convolution, and graph attention network.

[0014] In step 2, the CNN branch first preprocesses the input data, that is, performs denoising and spectral feature extraction at the pixel level, while preserving the local spatial structure information of the image. In the first layer, the input data is first normalized, then the number of channels is transformed using a 1x1 convolutional kernel, and then it is activated by the LeakyReLU function. In the second layer, the output of the first layer is first normalized, then the feature transformation is performed using a 3x3 convolutional kernel, and then it is activated by the LeakyReLU function.

[0015] After denoising and spectral transformation, the feature maps are then processed through multi-scale convolutional layers and residual connections. Specifically, different kernel sizes are used, including 7x7, 5x5, 3x3, and 1x1 kernels. The 7x7 kernel captures a larger range of local spatial features, suitable for extracting macroscopic structural information of the image. The 5x5 and 3x3 kernels capture medium and small ranges of local features, respectively, and are able to extract detailed information of the image. The 1x1 kernel is used to adjust the channel dimension and reduce computational cost. The input feature map is convolved, and different kernel sizes are used to generate feature maps at multiple scales. Each convolutional kernel generates a feature map, and feature maps of different scales are concatenated along the channel dimension to form a multi-scale feature map. Normalization is added after each convolutional layer to standardize the input data, reduce internal covariance, and accelerate model convergence. Then, the LeakyReLU function is used to introduce nonlinearity and enhance the model's expressive power. At the same time, in each convolutional block, the input features are added to the output of the convolutional layer to form a residual connection. In the convolutional block, the input is x, and the output after the convolutional layer and activation function is F(x). The residual connection directly passes the input to the output through x+F(x).

[0016] The specific steps of the Channel Attention Module (CAM) are as follows: max pooling and average pooling are performed on the feature maps obtained by multi-scale convolution to obtain two feature maps. Then, convolution operations are used to process the pooled feature maps to generate two attention maps. The two attention maps are merged and normalized by the Sigmoid function to generate the final spatial attention map.

[0017] In step 2, the AGCN branch is used to process the sub-network of superpixel-level features. It extracts image signal features and utilizes the relationships between superpixels. First, preprocessing is performed, i.e., the input data is preprocessed using superpixel segmentation, which involves dividing the original data into superpixel segments of different sizes. Then, the processed superpixel data is used as the input to the AGCN branch. The input data to the AGCN branch is the feature matrix S obtained during superpixel segmentation, with a shape of (N, in-channels), where N is the number of superpixels and in-channels is the feature dimension of each superpixel. The Euclidean distance between all superpixel features is calculated to obtain the distance matrix, as shown in the formula:

[0018]

[0019] S i and S j These are the i-th and j-th rows of the feature matrix S, respectively. i,k and S j,k These are the k-th elements of the i-th and j-th rows of the feature matrix S, respectively;

[0020] The generated Gaussian similarity matrix is ​​the dynamic adjacency matrix D, used to represent the similarity relationship between superpixels. The formula is:

[0021]

[0022] Where σ is a learnable parameter;

[0023] The static and dynamic adjacency matrices are then fused. Matrix A is a predefined static adjacency matrix that represents the fixed connection relationship between superpixels. The static adjacency matrix A and the dynamic adjacency matrix are fused using fusion weights γ and (1-γ) to obtain the fused adjacency matrix. The node features are weighted using a node attention mechanism to enhance important features and suppress unimportant features. The attention weights are calculated, and then the input features are mapped to a new feature space through a linear layer. The fused adjacency matrix is ​​used to perform graph convolution operations. The adaptive graph convolution module goes through three processes from input to graph convolution.

[0024] The Graph Attention (GAT) mechanism first performs a linear transformation on the input node feature matrix to map the node features to a new feature space. This step is achieved through a learnable weight matrix W, which is initialized according to a Xavier uniform distribution. The resulting feature matrix Wh after the linear transformation is used for subsequent attention calculations. Each node feature vector in the transformed feature matrix Wh is repeatedly expanded to form all possible node pair combinations. That is, for each node i, its feature vector Whi is combined with the feature vectors Whj of all other nodes j to form a new feature matrix. Then, a learnable attention vector a is used to calculate the attention score between node pairs. A LeakyReLU activation function is then used to introduce non-linearity, as shown in the formula:

[0025] e ij =LeakyReLU(a T [Wh i ][Wh j ])(3)

[0026] The calculated attention score e ij Normalization is required to ensure the validity and stability of the weights. In GAT, the softmax function is used to normalize the attention score, making it sum to 1 across all neighbors of node i. Furthermore, node j is considered a valid neighbor of node i only if the value in the adjacency matrix is ​​greater than 0. The normalized attention weight a ij The transformed feature Whj is used to calculate the new feature vector of node i. In this way, the new feature of each node is a weighted average of the features of its neighboring nodes, while the attention weights dynamically measure the importance of the neighboring nodes and perform a weighted summation to obtain the new feature vector of node i.

[0027] Step 2 is as follows:

[0028] The feature fusion module employs three feature fusion strategies: weighted summation fusion, concatenation fusion, and attention mechanism fusion. The outputs of these three feature fusion strategies are then input into the adaptive feature fusion module to obtain the final output.

[0029] In the weighted summation fusion, the output scores of CNN and AGCN are classified through linear layers to obtain two classification probability score matrices. The two score matrices are weighted and summed using two learnable weight parameters (γ1 and γ2) to generate the final classification probability. The weight parameters are normalized to the [0,1] interval by the sigmoid function.

[0030] In the concatenation and fusion strategy, the features of CNN and AGCN are concatenated together along the channel dimension to form a high-dimensional feature vector. Then, a fully connected layer is used to perform a non-linear transformation on the concatenated features to generate the final classification result.

[0031] The attention mechanism fusion strategy uses an attention mechanism to dynamically adjust the contribution ratio of features from the CNN branch and the AGCN branch. Specifically, firstly, the features of the two branches are projected onto the same feature dimension through linear layers, and then the projected features are concatenated and input into a multilayer perceptron (MLP). The MLP outputs an attention weight with a value in the range of [0,1]. Finally, the features of the two branches are weighted and summed according to the attention weight to obtain the final fusion result.

[0032] Finally, an adaptive fusion strategy is used to automatically learn the optimal combination ratio of the three fusion strategies based on the different input features. Specifically, the outputs Y1, Y2, and Y3 of the three fusion strategies are concatenated along the feature dimension and then input into an adaptive fusion module. This module outputs three adaptive weights with values ​​in the range [0,1] and the sum of the three weights is 1. Finally, the outputs of the three fusion strategies are weighted and summed according to the adaptive weights to obtain the final fusion result Y.

[0033] Step 3 is implemented in the following steps:

[0034] Set the learning rate to 0.001, the Epoch to 200, and the optimizer to Adam. Randomly select 5% of the labeled pixels from the Salinas dataset as the training set and input them into the network model obtained in step 2 to finally obtain a trained classification network model.

[0035] Step 4 is implemented in the following steps:

[0036] Two image classification evaluation metrics are selected: average accuracy (AA) and Kappa coefficient. The value of AA ranges from 0 to 1, and the value of Kappa ranges from -1 to 1. The higher the value of the metric, the better the classification effect.

[0037] The beneficial effects of this invention are that, compared with traditional methods, the hyperspectral image classification method based on adaptive graph convolution and multi-scale feature fusion is more efficient in data processing, reducing the complexity and time cost of data processing, and better extracting effective information from the data, thus improving classification results. Traditional classification methods are based on simple assumptions and models, making it difficult to extract the complex spectral and spatial structural information of hyperspectral images, and are prone to the "curse of dimensionality." Most deep learning methods are based on pixel classification, ignoring the spatial relationships between pixels and the overall structural information of the image. Some simple CNN methods cannot fully extract subtle features. This invention uses a dual-branch collaborative approach to comprehensively extract the spectral and spatial features of hyperspectral images. At the same time, the multi-strategy fusion method fully utilizes the advantages of different fusion strategies. Compared with a single fusion method, it can more flexibly handle different types of hyperspectral image data, improve the model's performance and generalization ability, and achieve more accurate classification even in complex scenarios. Attached Figure Description

[0038] Figure 1 This is an overall flowchart of the hyperspectral image classification method based on adaptive graph convolution and multi-scale feature fusion of the present invention;

[0039] Figure 2 This is a network framework diagram of the hyperspectral image classification method based on adaptive graph convolution and multi-scale feature fusion of the present invention;

[0040] Figure 3 This is a framework diagram of each convolutional block of the CNN in the hyperspectral image classification method based on adaptive graph convolution and multi-scale feature fusion of the present invention;

[0041] Figure 4 This is a framework diagram of the CAM module in the hyperspectral image classification method based on adaptive graph convolution and multi-scale feature fusion of the present invention;

[0042] Figure 5 This is a comparison chart of the hyperspectral image classification method based on adaptive graph convolution and multi-scale feature fusion of the present invention on the Salinas dataset, using 5% of the training samples, with other methods. Detailed Implementation

[0043] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0044] This invention relates to a hyperspectral image classification method based on adaptive graph convolution and multi-scale feature fusion, the flowchart of which is shown below. Figure 1As shown, please follow these steps:

[0045] Step 1: Collect hyperspectral images to construct a hyperspectral dataset. The hyperspectral dataset contains spectral information from multiple continuous and narrow bands, ranging from visible light to near-infrared. Divide the constructed hyperspectral dataset into training, validation, and test sets to ensure that each category has a reasonable distribution in each subset.

[0046] Step 1 is implemented in the following steps:

[0047] Using the hyperspectral dataset as input, preprocessing is performed by dividing it into training, testing, and validation sets. The training set is used for model training, allowing the model to learn the mapping relationship between the spectral-spatial features of hyperspectral images and land cover categories. The validation set is used to adjust hyperparameters during model training to prevent overfitting, and the optimal model parameters are selected by evaluating model performance on the validation set. The testing set is used to objectively evaluate the model's generalization ability and classification performance after training, obtaining the final classification accuracy index.

[0048] Step 2, Combining Figure 2 We constructed a dual-branch feature fusion classification network model that combines CNN and AGCN;

[0049] In step 2, the dual-branch feature fusion classification network model combining CNN and AGCN includes two branches, CNN and AGCN, as well as a feature fusion module. The CNN branch consists of three parts: data preprocessing, multi-scale residual convolutional blocks, and spatial attention mechanism. The AGCN branch also consists of three parts: data preprocessing, adaptive graph convolution, and graph attention network.

[0050] Combination Figure 3 , Figure 4 In step 2, the CNN branch first preprocesses the input data, performing denoising and spectral feature extraction at the pixel level while preserving the local spatial structure information of the image. This enhances the expressive power of spectral features and suppresses noise. In the first layer, the input data is first normalized, then the number of channels is transformed using a 1x1 convolutional kernel, and finally activated by the LeakyReLU function. In the second layer, the output of the first layer is first normalized, then feature transformation is performed using a 3x3 convolutional kernel, and finally activated by the LeakyReLU function.

[0051] After denoising and spectral transformation, the feature maps are further enhanced through multi-scale convolutional layers and residual connections to improve their expressive power. Specific convolution steps involve using convolutional kernels of different sizes, including 7x7, 5x5, 3x3, and 1x1 kernels. The 7x7 kernel captures a large range of local spatial features, suitable for extracting macroscopic structural information of the image. The 5x5 and 3x3 kernels capture medium and small ranges of local features, respectively, and are able to extract detailed information of the image. The 1x1 kernel is used to adjust the channel dimension and reduce computational cost. The input feature map is convolved, and multiple scale feature maps are generated using different sized kernels, with each kernel generating one feature map. Feature maps of different scales are concatenated along the channel dimension to form a multi-scale feature map. Normalization is added after each convolutional layer to standardize the input data, reduce internal covariance, and accelerate model convergence. Then, the LeakyReLU function is used to introduce nonlinearity and enhance the model's expressive power. At the same time, in each convolutional block, the input features are added to the output of the convolutional layer to form a residual connection. In the convolutional block, the input is x, and the output after the convolutional layer and activation function is F(x). The residual connection directly passes the input to the output through x+F(x).

[0052] The main function of the Channel Attention Module (CAM) is to adaptively adjust the importance of each channel in the feature map, allowing the network to focus more on channels containing more useful information, thereby improving model performance. It dynamically adjusts the weights at each location in the multi-scale feature map output by the multi-scale CNN convolutional module and generates a spatial attention map. The specific steps of the CAM module are as follows: Max pooling and average pooling are performed on the feature maps obtained after multi-scale convolution, resulting in two feature maps. Then, convolution operations are used to process the pooled feature maps, generating two attention maps. These two attention maps are merged and normalized using the Sigmoid function to generate the final spatial attention map.

[0053] In step 2, the AGCN branch is used to process the sub-network of superpixel-level features. It extracts image signal features and utilizes the relationships between superpixels. First, preprocessing is performed, i.e., the input data is preprocessed using superpixel segmentation, which involves dividing the original data into superpixel segments of different sizes. Then, the processed superpixel data is used as the input to the AGCN branch. The input data to the AGCN branch is the feature matrix S obtained during superpixel segmentation, with a shape of (N, in-channels), where N is the number of superpixels and in-channels is the feature dimension of each superpixel. The Euclidean distance between all superpixel features is calculated to obtain the distance matrix, as shown in the formula:

[0054]

[0055] Si and S j These are the i-th and j-th rows of the feature matrix S, respectively. i,k and S j,k These are the k-th elements of the i-th and j-th rows of the feature matrix S, respectively;

[0056] The generated Gaussian similarity matrix is ​​the dynamic adjacency matrix D, used to represent the similarity relationship between superpixels. The formula is:

[0057]

[0058] Where σ is a learnable parameter;

[0059] The static and dynamic adjacency matrices are then fused. Matrix A is a predefined static adjacency matrix that represents the fixed connection relationship between superpixels. The static adjacency matrix A and the dynamic adjacency matrix are fused using fusion weights γ and (1-γ) to obtain the fused adjacency matrix. The node features are weighted using a node attention mechanism to enhance important features and suppress unimportant features. The attention weights are calculated, and then the input features are mapped to a new feature space through a linear layer. The fused adjacency matrix is ​​used to perform graph convolution operations. The adaptive graph convolution module goes through three processes from input to graph convolution.

[0060] The Fusion Graph Attention (GAT) mechanism introduces an attention mechanism to dynamically assign different weights to the neighbors of each node, thereby highlighting the influence of important neighbor nodes. Specifically, the core of the attention mechanism is to calculate the importance score between nodes, which is achieved through the following steps: First, the GAT mechanism performs a linear transformation on the input node feature matrix to map the node features to a new feature space. This step is achieved through a learnable weight matrix W, which is initialized according to a Xavier uniform distribution. The feature matrix Wh obtained after the linear transformation is used for subsequent attention calculations. The feature vector of each node in the linearly transformed feature matrix Wh is repeatedly expanded to form all possible node pair combinations. That is, for each node i, its feature vector Whi is combined with the feature vectors Whj of all other nodes j to form a new feature matrix. Then, a learnable attention vector a is used to calculate the attention score between node pairs, and a Leaky ReLU activation function is used to introduce non-linearity, as shown in the formula:

[0061] e ij =LeakyReLU(a T [Wh i ][Wh j (3)

[0062] The calculated attention score e ijNormalization is required to ensure the validity and stability of the weights. In GAT, the softmax function is used to normalize the attention score, making it sum to 1 across all neighbors of node i. Furthermore, node j is considered a valid neighbor of node i only if the value in the adjacency matrix is ​​greater than 0. The normalized attention weight a ij The transformed feature Whj is used to calculate the new feature vector of node i. In this way, the new feature of each node is a weighted average of the features of its neighboring nodes, while the attention weights dynamically measure the importance of the neighboring nodes and perform a weighted summation to obtain the new feature vector of node i.

[0063] Step 2 is as follows:

[0064] The feature fusion module employs three feature fusion strategies: weighted summation fusion, concatenation fusion, and attention mechanism fusion. The outputs of these three feature fusion strategies are then input into the adaptive feature fusion module to obtain the final output.

[0065] In the weighted summation fusion, the output scores of CNN and AGCN are classified separately through linear layers to obtain two classification probability score matrices. These two score matrices are then weighted and summed using two learnable weight parameters (γ1 and γ2) to generate the final classification probability. The weight parameters are normalized to the [0,1] interval using the sigmoid function to ensure the stability of the fusion process. This direct weighted fusion of classification scores can balance the contributions of CNN and AGCN to a certain extent and improve the accuracy of classification.

[0066] In the concatenation and fusion strategy, features from CNN and AGCN are concatenated along the channel dimension to form a high-dimensional feature vector. Then, a fully connected layer performs a non-linear transformation on the concatenated features to generate the final classification result. This concatenation operation preserves the integrity of the original features and fully utilizes the different information extracted by CNN and AGCN. Simultaneously, the fully connected layer can learn the complex relationships between the concatenated features, improving the model's expressive power.

[0067] The attention mechanism fusion strategy dynamically adjusts the contribution ratio of features from the CNN and AGCN branches using an attention mechanism. Specifically, firstly, the features from both branches are projected onto the same feature dimension through linear layers. Then, the projected features are concatenated and input into a multilayer perceptron (MLP). The MLP outputs an attention weight with a value in the range [0,1]. Finally, the features from both branches are weighted and summed according to the attention weight to obtain the final fusion result. This fusion method can adaptively adjust the importance of the features from the two branches based on the different input features, thereby improving the fusion effect.

[0068] Finally, an adaptive fusion strategy automatically learns the optimal combination ratio of the three fusion strategies based on different input features. Specifically, the outputs Y1, Y2, and Y3 of the three fusion strategies are concatenated along the feature dimension and then input into an adaptive fusion module. This module outputs three adaptive weights with values ​​in the range [0,1], and the sum of the three weights is 1. Finally, the outputs of the three fusion strategies are weighted and summed according to the adaptive weights to obtain the final fusion result Y. This fusion method can fully utilize the advantages of the three fusion strategies, dynamically adjusting their contribution ratios according to different input features, thereby improving the model's performance and generalization ability.

[0069] Step 3: Input the training set divided in Step 1 into the classification network model constructed in Step 2, set the parameters, and train it;

[0070] Step 3 is implemented in the following steps:

[0071] Set the learning rate to 0.001, the Epoch to 200, and the optimizer to Adam. Randomly select 5% of the labeled pixels from the Salinas dataset as the training set and input them into the network model obtained in step 2 to finally obtain a trained classification network model.

[0072] Step 4: Obtain the hyperspectral image land cover classification results, and evaluate the indicators, data, and image information of the classification results.

[0073] Step 4 is implemented in the following steps:

[0074] Two image classification evaluation metrics are selected: average accuracy (AA) and Kappa coefficient. The value of AA ranges from 0 to 1, and the value of Kappa ranges from -1 to 1. The higher the value of the metric, the better the classification effect.

[0075] Example 1

[0076] This invention relates to a hyperspectral image classification method based on adaptive graph convolution and multi-scale feature fusion, the flowchart of which is shown below. Figure 1 As shown, please follow these steps:

[0077] Step 1: Collect hyperspectral images to construct a hyperspectral dataset. The hyperspectral dataset contains spectral information from multiple continuous and narrow bands, ranging from visible light to near-infrared. Divide the constructed hyperspectral dataset into training, validation, and test sets to ensure that each category has a reasonable distribution in each subset.

[0078] Step 2: Construct a dual-branch feature fusion classification network model combining CNN and AGCN;

[0079] Step 3: Input the training set divided in Step 1 into the classification network model constructed in Step 2, set the parameters, and train it;

[0080] Step 4: Obtain the hyperspectral image land cover classification results, and evaluate the indicators, data, and image information of the classification results.

[0081] Example 2

[0082] This invention relates to a hyperspectral image classification method based on adaptive graph convolution and multi-scale feature fusion, the flowchart of which is shown below. Figure 1 As shown, please follow these steps:

[0083] Step 1: Collect hyperspectral images to construct a hyperspectral dataset. The hyperspectral dataset contains spectral information from multiple continuous and narrow bands, ranging from visible light to near-infrared. Divide the constructed hyperspectral dataset into training, validation, and test sets to ensure that each category has a reasonable distribution in each subset.

[0084] Step 1 is implemented in the following steps:

[0085] Using the hyperspectral dataset as input, preprocessing is performed by dividing it into training, testing, and validation sets. The training set is used for model training, allowing the model to learn the mapping relationship between the spectral-spatial features of hyperspectral images and land cover categories. The validation set is used to adjust hyperparameters during model training to prevent overfitting, and the optimal model parameters are selected by evaluating model performance on the validation set. The testing set is used to objectively evaluate the model's generalization ability and classification performance after training, obtaining the final classification accuracy index.

[0086] Step 2, Combining Figure 2 We constructed a dual-branch feature fusion classification network model that combines CNN and AGCN;

[0087] In step 2, the dual-branch feature fusion classification network model combining CNN and AGCN includes two branches, CNN and AGCN, as well as a feature fusion module. The CNN branch consists of three parts: data preprocessing, multi-scale residual convolutional blocks, and spatial attention mechanism. The AGCN branch also consists of three parts: data preprocessing, adaptive graph convolution, and graph attention network.

[0088] Step 3: Input the training set divided in Step 1 into the classification network model constructed in Step 2, set the parameters, and train it;

[0089] Step 4: Obtain the hyperspectral image land cover classification results, and evaluate the indicators, data, and image information of the classification results.

[0090] Example 3

[0091] This invention relates to a hyperspectral image classification method based on adaptive graph convolution and multi-scale feature fusion, the flowchart of which is shown below. Figure 1 As shown, please follow these steps:

[0092] Step 1: Collect hyperspectral images to construct a hyperspectral dataset. The hyperspectral dataset contains spectral information from multiple continuous and narrow bands, ranging from visible light to near-infrared. Divide the constructed hyperspectral dataset into training, validation, and test sets to ensure that each category has a reasonable distribution in each subset.

[0093] Step 1 is implemented in the following steps:

[0094] Using the hyperspectral dataset as input, preprocessing is performed by dividing it into training, testing, and validation sets. The training set is used for model training, allowing the model to learn the mapping relationship between the spectral-spatial features of hyperspectral images and land cover categories. The validation set is used to adjust hyperparameters during model training to prevent overfitting, and the optimal model parameters are selected by evaluating model performance on the validation set. The testing set is used to objectively evaluate the model's generalization ability and classification performance after training, obtaining the final classification accuracy index.

[0095] Step 2, Combining Figure 2 We constructed a dual-branch feature fusion classification network model that combines CNN and AGCN;

[0096] In step 2, the dual-branch feature fusion classification network model combining CNN and AGCN includes two branches, CNN and AGCN, as well as a feature fusion module. The CNN branch consists of three parts: data preprocessing, multi-scale residual convolutional blocks, and spatial attention mechanism. The AGCN branch also consists of three parts: data preprocessing, adaptive graph convolution, and graph attention network.

[0097] Combination Figure 3 , Figure 4 In step 2, the CNN branch first preprocesses the input data, performing denoising and spectral feature extraction at the pixel level while preserving the local spatial structure information of the image. This enhances the expressive power of spectral features and suppresses noise. In the first layer, the input data is first normalized, then the number of channels is transformed using a 1x1 convolutional kernel, and finally activated by the LeakyReLU function. In the second layer, the output of the first layer is first normalized, then feature transformation is performed using a 3x3 convolutional kernel, and finally activated by the LeakyReLU function.

[0098] After denoising and spectral transformation, the feature maps are further enhanced through multi-scale convolutional layers and residual connections to improve their expressive power. Specific convolution steps involve using convolutional kernels of different sizes, including 7x7, 5x5, 3x3, and 1x1 kernels. The 7x7 kernel captures a large range of local spatial features, suitable for extracting macroscopic structural information of the image. The 5x5 and 3x3 kernels capture medium and small ranges of local features, respectively, and are able to extract detailed information of the image. The 1x1 kernel is used to adjust the channel dimension and reduce computational cost. The input feature map is convolved, and multiple scale feature maps are generated using different sized kernels, with each kernel generating one feature map. Feature maps of different scales are concatenated along the channel dimension to form a multi-scale feature map. Normalization is added after each convolutional layer to standardize the input data, reduce internal covariance, and accelerate model convergence. Then, the LeakyReLU function is used to introduce nonlinearity and enhance the model's expressive power. At the same time, in each convolutional block, the input features are added to the output of the convolutional layer to form a residual connection. In the convolutional block, the input is x, and the output after the convolutional layer and activation function is F(x). The residual connection directly passes the input to the output through x+F(x).

[0099] The main function of the Channel Attention Module (CAM) is to adaptively adjust the importance of each channel in the feature map, allowing the network to focus more on channels containing more useful information, thereby improving model performance. It dynamically adjusts the weights at each location in the multi-scale feature map output by the multi-scale CNN convolutional module and generates a spatial attention map. The specific steps of the CAM module are as follows: Max pooling and average pooling are performed on the feature maps obtained after multi-scale convolution, resulting in two feature maps. Then, convolution operations are used to process the pooled feature maps, generating two attention maps. These two attention maps are merged and normalized using the Sigmoid function to generate the final spatial attention map.

[0100] Step 3: Input the training set divided in Step 1 into the classification network model constructed in Step 2, set the parameters, and train it;

[0101] Step 4: Obtain the hyperspectral image land cover classification results, and evaluate the indicators, data, and image information of the classification results.

[0102] Example 4

[0103] This invention relates to a hyperspectral image classification method based on adaptive graph convolution and multi-scale feature fusion, the flowchart of which is shown below. Figure 1 As shown, please follow these steps:

[0104] Step 1: Collect hyperspectral images to construct a hyperspectral dataset. The hyperspectral dataset contains spectral information from multiple continuous and narrow bands, ranging from visible light to near-infrared. Divide the constructed hyperspectral dataset into training, validation, and test sets to ensure that each category has a reasonable distribution in each subset.

[0105] Step 1 is implemented in the following steps:

[0106] Using the hyperspectral dataset as input, preprocessing is performed by dividing it into training, testing, and validation sets. The training set is used for model training, allowing the model to learn the mapping relationship between the spectral-spatial features of hyperspectral images and land cover categories. The validation set is used to adjust hyperparameters during model training to prevent overfitting, and the optimal model parameters are selected by evaluating model performance on the validation set. The testing set is used to objectively evaluate the model's generalization ability and classification performance after training, obtaining the final classification accuracy index.

[0107] Step 2, Combining Figure 2 We constructed a dual-branch feature fusion classification network model that combines CNN and AGCN;

[0108] In step 2, the dual-branch feature fusion classification network model combining CNN and AGCN includes two branches, CNN and AGCN, as well as a feature fusion module. The CNN branch consists of three parts: data preprocessing, multi-scale residual convolutional blocks, and spatial attention mechanism. The AGCN branch also consists of three parts: data preprocessing, adaptive graph convolution, and graph attention network.

[0109] In step 2, the AGCN branch is used to process the sub-network of superpixel-level features. It extracts image signal features and utilizes the relationships between superpixels. First, preprocessing is performed, i.e., the input data is preprocessed using superpixel segmentation, which involves dividing the original data into superpixel segments of different sizes. Then, the processed superpixel data is used as the input to the AGCN branch. The input data to the AGCN branch is the feature matrix S obtained during superpixel segmentation, with a shape of (N, in-channels), where N is the number of superpixels and in-channels is the feature dimension of each superpixel. The Euclidean distance between all superpixel features is calculated to obtain the distance matrix, as shown in the formula:

[0110]

[0111] S i and S j These are the i-th and j-th rows of the feature matrix S, respectively. i,k and S j,k These are the k-th elements of the i-th and j-th rows of the feature matrix S, respectively;

[0112] The generated Gaussian similarity matrix is ​​the dynamic adjacency matrix D, used to represent the similarity relationship between superpixels. The formula is:

[0113]

[0114] Where σ is a learnable parameter;

[0115] The static and dynamic adjacency matrices are then fused. Matrix A is a predefined static adjacency matrix that represents the fixed connection relationship between superpixels. The static adjacency matrix A and the dynamic adjacency matrix are fused using fusion weights γ and (1-γ) to obtain the fused adjacency matrix. The node features are weighted using a node attention mechanism to enhance important features and suppress unimportant features. The attention weights are calculated, and then the input features are mapped to a new feature space through a linear layer. The fused adjacency matrix is ​​used to perform graph convolution operations. The adaptive graph convolution module goes through three processes from input to graph convolution.

[0116] The Fusion Graph Attention (GAT) mechanism introduces an attention mechanism to dynamically assign different weights to the neighbors of each node, thereby highlighting the influence of important neighbor nodes. Specifically, the core of the attention mechanism is to calculate the importance score between nodes, which is achieved through the following steps: First, the GAT mechanism performs a linear transformation on the input node feature matrix to map the node features to a new feature space. This step is achieved through a learnable weight matrix W, which is initialized according to a Xavier uniform distribution. The feature matrix Wh obtained after the linear transformation is used for subsequent attention calculations. The feature vector of each node in the linearly transformed feature matrix Wh is repeatedly expanded to form all possible node pair combinations. That is, for each node i, its feature vector Whi is combined with the feature vectors Whj of all other nodes j to form a new feature matrix. Then, a learnable attention vector a is used to calculate the attention score between node pairs, and a Leaky ReLU activation function is used to introduce non-linearity, as shown in the formula:

[0117] e ij =LeakyReLU(a T [Wh i ][Wh j (3)

[0118] The calculated attention score e ij Normalization is required to ensure the validity and stability of the weights. In GAT, the softmax function is used to normalize the attention score, making it sum to 1 across all neighbors of node i. Furthermore, node j is considered a valid neighbor of node i only if the value in the adjacency matrix is ​​greater than 0. The normalized attention weight aij The transformed feature Whj is used to calculate the new feature vector of node i. In this way, the new feature of each node is a weighted average of the features of its neighboring nodes, while the attention weights dynamically measure the importance of the neighboring nodes and perform a weighted summation to obtain the new feature vector of node i.

[0119] Step 3: Input the training set divided in Step 1 into the classification network model constructed in Step 2, set the parameters, and train it;

[0120] Step 4: Obtain the hyperspectral image land cover classification results, and evaluate the indicators, data, and image information of the classification results.

[0121] Example 5

[0122] This invention relates to a hyperspectral image classification method based on adaptive graph convolution and multi-scale feature fusion, the flowchart of which is shown below. Figure 1 As shown, please follow these steps:

[0123] Step 1: Collect hyperspectral images to construct a hyperspectral dataset. The hyperspectral dataset contains spectral information from multiple continuous and narrow bands, ranging from visible light to near-infrared. Divide the constructed hyperspectral dataset into training, validation, and test sets to ensure that each category has a reasonable distribution in each subset.

[0124] Step 1 is implemented in the following steps:

[0125] Using the hyperspectral dataset as input, preprocessing is performed by dividing it into training, testing, and validation sets. The training set is used for model training, allowing the model to learn the mapping relationship between the spectral-spatial features of hyperspectral images and land cover categories. The validation set is used to adjust hyperparameters during model training to prevent overfitting, and the optimal model parameters are selected by evaluating model performance on the validation set. The testing set is used to objectively evaluate the model's generalization ability and classification performance after training, obtaining the final classification accuracy index.

[0126] Step 2, Combining Figure 2 We constructed a dual-branch feature fusion classification network model that combines CNN and AGCN;

[0127] In step 2, the dual-branch feature fusion classification network model combining CNN and AGCN includes two branches, CNN and AGCN, as well as a feature fusion module. The CNN branch consists of three parts: data preprocessing, multi-scale residual convolutional blocks, and spatial attention mechanism. The AGCN branch also consists of three parts: data preprocessing, adaptive graph convolution, and graph attention network.

[0128] Step 2 is as follows:

[0129] The feature fusion module employs three feature fusion strategies: weighted summation fusion, concatenation fusion, and attention mechanism fusion. The outputs of these three feature fusion strategies are then input into the adaptive feature fusion module to obtain the final output.

[0130] In the weighted summation fusion, the output scores of CNN and AGCN are classified separately through linear layers to obtain two classification probability score matrices. These two score matrices are then weighted and summed using two learnable weight parameters (γ1 and γ2) to generate the final classification probability. The weight parameters are normalized to the [0,1] interval using the sigmoid function to ensure the stability of the fusion process. This direct weighted fusion of classification scores can balance the contributions of CNN and AGCN to a certain extent and improve the accuracy of classification.

[0131] In the concatenation and fusion strategy, features from CNN and AGCN are concatenated along the channel dimension to form a high-dimensional feature vector. Then, a fully connected layer performs a non-linear transformation on the concatenated features to generate the final classification result. This concatenation operation preserves the integrity of the original features and fully utilizes the different information extracted by CNN and AGCN. Simultaneously, the fully connected layer can learn the complex relationships between the concatenated features, improving the model's expressive power.

[0132] The attention mechanism fusion strategy dynamically adjusts the contribution ratio of features from the CNN and AGCN branches using an attention mechanism. Specifically, firstly, the features from both branches are projected onto the same feature dimension through linear layers. Then, the projected features are concatenated and input into a multilayer perceptron (MLP). The MLP outputs an attention weight with a value in the range [0,1]. Finally, the features from both branches are weighted and summed according to the attention weight to obtain the final fusion result. This fusion method can adaptively adjust the importance of the features from the two branches based on the different input features, thereby improving the fusion effect.

[0133] Finally, an adaptive fusion strategy automatically learns the optimal combination ratio of the three fusion strategies based on different input features. Specifically, the outputs Y1, Y2, and Y3 of the three fusion strategies are concatenated along the feature dimension and then input into an adaptive fusion module. This module outputs three adaptive weights with values ​​in the range [0,1], and the sum of the three weights is 1. Finally, the outputs of the three fusion strategies are weighted and summed according to the adaptive weights to obtain the final fusion result Y. This fusion method can fully utilize the advantages of the three fusion strategies, dynamically adjusting their contribution ratios according to different input features, thereby improving the model's performance and generalization ability.

[0134] Step 3: Input the training set divided in Step 1 into the classification network model constructed in Step 2, set the parameters, and train it;

[0135] Step 3 is implemented in the following steps:

[0136] Set the learning rate to 0.001, the Epoch to 200, and the optimizer to Adam. Randomly select 5% of the labeled pixels from the Salinas dataset as the training set and input them into the network model obtained in step 2 to finally obtain a trained classification network model.

[0137] Step 4: Obtain the hyperspectral image land cover classification results, and evaluate the indicators, data, and image information of the classification results.

[0138] Step 4 is implemented in the following steps:

[0139] Two image classification evaluation metrics are selected: average accuracy (AA) and Kappa coefficient. The value of AA ranges from 0 to 1, and the value of Kappa ranges from -1 to 1. The higher the value of the metric, the better the classification effect.

[0140] Example 6

[0141] This experiment was conducted on a computer equipped with a Windows 10 (64-bit) system, an Intel(R) Core(TM) i7-11700, an NVIDIA GeForce RTX 3060 GPU, and 16GB of RAM.

[0142] In step 1, the simulation experiment of this invention uses the Salinas dataset, which covers a spectral range of 400 to 2500 nanometers and contains 224 bands. The data mainly consists of a large image covering various crops and uncultivated land, divided into 16 different land cover categories, such as spinach, lettuce, and corn. The data specifically includes the original spectral image, land cover classification labels, and metadata. Each pixel in the original image has 224 spectral values;

[0143] Perform step 2 to construct a dual-branch feature fusion classification network model that combines CNN and AGCN;

[0144] Perform step 3 to set training parameters and train. Throughout the training process, set the learning rate to 0.001 and the Epoch to 200, and the optimizer to Adam.

[0145] Step 4 involves using the trained network model to classify the image data in the test set and evaluating the classification results, including metrics, data, and image information.

[0146] To verify the effectiveness of the proposed method, the fusion results of the method are compared with three other methods: AMGCFN, CEGCN, and GTFN. Under the same conditions, the classification results of the proposed method are compared with those of four existing classification methods in the hyperspectral field on the Salinas hyperspectral dataset.

[0147] Table 1 compares the classification accuracy of the four methods on the Salinas dataset.

[0148]

[0149]

[0150] Table 1 compares the classification accuracy of this invention with that of AMGCFN, CEGCN, and WSSG on the Salinas dataset for different land cover categories, and also provides two comprehensive evaluation indicators: Average Accuracy (AA) and Kappa coefficient. It can be seen that in several categories, such as Class 3, Class 5, and Class 11, the classification accuracy of this invention is higher than the other three methods, demonstrating its advantage in these land cover classifications. The AA value of this invention is 97.49%, significantly higher than AMGCFN's 87.47%, CEGCN's 95.81%, and WSSG's 91.1%, indicating that this invention performs better on average in terms of classification accuracy across different land cover categories, and can accurately classify different land cover categories more evenly. Figure 5 The image shows the results of this invention compared to three other methods (AMGCFN, CEGCN, and WSSG) in hyperspectral image classification, with real-world images provided for reference. Compared to the results of this invention, AMGCFN shows differences in color distribution and feature delineation in some areas. The boundaries of some areas are not accurately defined, potentially misclassifying one feature as another, leading to a deviation between the classification results and the actual features. CEGCN's classification results also have similar issues, with insufficient precision in the delineation of some feature categories. For example, the color fill in some areas is inconsistent with the actual features, indicating errors in the identification of the corresponding features and affecting the overall classification accuracy. Similarly, WSSG's classification results show significant differences from the actual features in the image. The chaotic colors in some areas and the blurred boundaries between different features reflect deficiencies in the extraction and classification of different feature characteristics in hyperspectral image classification, resulting in unsatisfactory classification results. From the color distribution and region delineation, it can be seen that the matching degree between this invention and real-world images is relatively high. The boundaries of different land cover categories are relatively clear, and the color distribution corresponds more closely to the actual land cover as expected, indicating that the present invention can accurately identify different land cover categories in hyperspectral image classification, and the classification results are visually closer to the actual situation.

[0151] Based on the above experimental results, the algorithm of this invention outperforms other comparative algorithms in terms of performance metrics and classification effect on the Salina dataset, achieving better classification results.

Claims

1. A hyperspectral image classification method based on adaptive graph convolution and multi-scale feature fusion, characterized in that, The specific steps are as follows: Step 1: Collect hyperspectral images to construct a hyperspectral dataset. The constructed hyperspectral dataset is divided into a training set, a validation set, and a test set. Step 2: Construct a dual-branch feature fusion classification network model combining CNN and AGCN; In step 2, the AGCN branch comprises three parts: data preprocessing, adaptive graph convolution, and graph attention network. The AGCN branch is a sub-network used to process superpixel-level features, extracting graph signal features and utilizing the relationships between superpixels. First, preprocessing is performed, i.e., the input data is preprocessed using superpixel segmentation, which involves dividing the original data into superpixel segments of different sizes. Then, the processed superpixel data is used as the input to the AGCN branch. The input data to the AGCN branch is the feature matrix S obtained during superpixel segmentation, with a shape of (N, in-channels), where N is the number of superpixels and in-channels is the feature dimension of each superpixel. The Euclidean distance between all superpixel features is calculated to obtain the distance matrix, as shown in the formula: and These are the characteristic matrices S The row and number j OK, and These are the characteristic matrices S The row and number j The first line k One element; The generated Gaussian similarity matrix is ​​the dynamic adjacency matrix D, used to represent the similarity relationship between superpixels. The formula is: in These are learnable parameters; The static and dynamic adjacency matrices are then fused, where matrix A is a predefined static adjacency matrix representing the fixed connection relationship between superpixels. The static adjacency matrix A and the dynamic adjacency matrix are fused using fusion weights γ and (1 - γ) to obtain the fused adjacency matrix. The node features are weighted using a node attention mechanism to enhance important features and suppress unimportant features. The attention weights are calculated, and then the input features are mapped to a new feature space through a linear layer. The fused adjacency matrix is ​​used to perform graph convolution operations. The adaptive graph convolution module goes through three processes from input to graph convolution. The Fusion Graph Attention (GAT) mechanism first performs a linear transformation on the input node feature matrix to map the node features to a new feature space. This step is achieved through a learnable weight matrix W, which is initialized according to a Xavier uniform distribution. The feature matrix obtained after the linear transformation... The linearly transformed feature matrix is ​​used for subsequent attention calculations. The feature vector of each node is repeatedly expanded to form all possible combinations of node pairs, that is, for each node... , and its feature vector The eigenvectors of all other nodes j This is combined into a new feature matrix. Then, a learnable attention vector 'a' is used to calculate the attention score between node pairs. A LeakyReLU activation function is then used to introduce non-linearity, as shown in the formula: Calculated attention score e ij Normalization is required to ensure the validity and stability of the weights. In GAT, the softmax function is used to normalize the attention scores at each node. The sum of all its neighbors is 1. Furthermore, a node is considered a node only if the value in its adjacency matrix is ​​greater than 0. Only then are they considered nodes Effective neighbors, attention weights after normalization Used for transforming features In this way, the new features of each node are a weighted average of the features of its neighboring nodes, while the attention weights dynamically measure the importance of neighboring nodes and perform a weighted summation to obtain the node's new features. The new feature vector; Step 3: Input the training set divided in Step 1 into the classification network model constructed in Step 2, set the parameters, and train it; Step 4: Obtain the hyperspectral image land cover classification results, and evaluate the indicators, data, and image information of the classification results.

2. The hyperspectral image classification method based on adaptive graph convolution and multi-scale feature fusion according to claim 1, characterized in that, Step 1 is implemented in the following steps: Using the hyperspectral dataset as input, preprocessing is performed by dividing it into training, testing, and validation sets. The training set is used for model training, allowing the model to learn the mapping relationship between the spectral-spatial features of hyperspectral images and land cover categories. The validation set is used to adjust hyperparameters during model training to prevent overfitting, and the optimal model parameters are selected by evaluating model performance on the validation set. The testing set is used to objectively evaluate the model's generalization ability and classification performance after training, obtaining the final classification accuracy index.

3. The hyperspectral image classification method based on adaptive graph convolution and multi-scale feature fusion according to claim 2, characterized in that, The dual-branch feature fusion classification network model combining CNN and AGCN in step 2 includes two branches, CNN and AGCN, and a feature fusion module. The CNN branch consists of three parts: data preprocessing, multi-scale residual convolutional blocks, and spatial attention mechanism.

4. The hyperspectral image classification method based on adaptive graph convolution and multi-scale feature fusion according to claim 3, characterized in that, In step 2, the CNN branch first preprocesses the input data, that is, performs denoising and spectral feature extraction at the pixel level, while preserving the local spatial structure information of the image. In the first layer, the input data is first normalized, then the number of channels is transformed using a 1x1 convolution kernel, and then the LeakyReLU activation function is applied. In the second layer, the output of the first layer is first normalized, then the feature transformation is performed using a 3x3 convolution kernel, and then the LeakyReLU activation function is applied. After denoising and spectral transformation, the feature maps are then processed through multi-scale convolutional layers and residual connections. Specific convolution steps involve using different kernel sizes, including 7x7, 5x5, 3x3, and 1x1 kernels. The 7x7 kernel captures a large range of local spatial features, suitable for extracting macroscopic structural information of the image; the 5x5 and 3x3 kernels capture medium and small ranges of local features respectively, capable of extracting detailed information of the image; and the 1x1 kernel… Convolutional kernels are used to adjust the channel dimension, reducing computational cost. They perform convolution operations on the input feature map, generating feature maps at multiple scales using kernels of different sizes. Each kernel generates one feature map, and these feature maps at different scales are concatenated along the channel dimension to form a multi-scale feature map. Normalization is added after each convolutional layer to standardize the input data, reduce internal covariance, and accelerate model convergence. The LeakyReLU function is then used to introduce non-linearity, enhancing the model's expressive power. Simultaneously, within each convolutional block, the input features are added to the output of the convolutional layer, forming residual connections. In the convolutional block, the input is... The output after the convolutional layer and activation function is Residual connection through This method directly passes the input to the output; The specific steps of the Channel Attention Module (CAM) are as follows: max pooling and average pooling are performed on the feature maps obtained by multi-scale convolution to obtain two feature maps. Then, convolution operations are used to process the pooled feature maps to generate two attention maps. The two attention maps are merged and normalized by the Sigmoid function to generate the final spatial attention map.

5. The hyperspectral image classification method based on adaptive graph convolution and multi-scale feature fusion according to claim 4, characterized in that, Step 2 is described in detail below: The feature fusion module employs three feature fusion strategies: weighted summation fusion, concatenation fusion, and attention mechanism fusion. The outputs of these three feature fusion strategies are then input into the adaptive feature fusion module to obtain the final output. In the weighted summation fusion, the output scores of CNN and AGCN are classified through linear layers to obtain two classification probability score matrices. The two score matrices are weighted and summed using two learnable weight parameters γ1 and γ2 to generate the final classification probability. The weight parameters are normalized to the interval [0, 1] using the sigmoid function. In the concatenation and fusion strategy, the features of CNN and AGCN are concatenated together along the channel dimension to form a high-dimensional feature vector. Then, a fully connected layer is used to perform a non-linear transformation on the concatenated features to generate the final classification result. The attention mechanism fusion strategy uses an attention mechanism to dynamically adjust the contribution ratio of features from the CNN branch and the AGCN branch. Specifically, firstly, the features of the two branches are projected onto the same feature dimension through linear layers, and then the projected features are concatenated and input into a multilayer perceptron (MLP). The MLP outputs an attention weight with a value in the range [0, 1]. Finally, the features of the two branches are weighted and summed according to the attention weight to obtain the final fusion result. Finally, an adaptive fusion strategy is used to automatically learn the optimal combination ratio of the three fusion strategies based on different input features. Specifically, the outputs of the three fusion strategies are combined... Y 1. Y 2 and Y 3. The data is concatenated along the feature dimension and then input into an adaptive fusion module. This module outputs three adaptive weights with values ​​in the range [0, 1], and the sum of the three weights is 1. Finally, the outputs of the three fusion strategies are weighted and summed according to the adaptive weights to obtain the final fusion result. Y .

6. The hyperspectral image classification method based on adaptive graph convolution and multi-scale feature fusion according to claim 5, characterized in that, Step 3 is implemented in the following steps: Set the learning rate to 0.001, the Epoch to 200, and the optimizer to Adam. Randomly select 5% of the labeled pixels from the Salinas dataset as the training set and input them into the network model obtained in step 2 to finally obtain a trained classification network model.

7. The hyperspectral image classification method based on adaptive graph convolution and multi-scale feature fusion according to claim 6, characterized in that, Step 4 is implemented in the following steps: Two image classification evaluation metrics are selected: average accuracy (AA) and Kappa coefficient. The value of AA ranges from 0 to 1, and the value of Kappa ranges from -1 to 1. The higher the value of the metric, the better the classification effect.

Images