A Hyperspectral Image Classification Network Model and Method Based on Deep Learning

Abstract

This invention discloses a deep learning-based hyperspectral image classification network model and method, belonging to the field of remote sensing image processing technology. It includes a stem network, a spectral feature extraction module, a hybrid spatial spectrum fusion module, and a classifier connected sequentially. The stem network is used to perform preliminary channel expansion and spectral dimension downsampling on the input raw hyperspectral data. The spectral feature extraction module extracts spectral features from the hyperspectral data. The hybrid spatial spectrum fusion module further performs deep grouping separable convolution and frequency-space attention on the spectral features extracted by the spectral feature extraction module, followed by deep fusion to obtain high-dimensional features. The classifier maps the high-dimensional features to class probabilities. This invention, using the above network model and method, improves the overall classification accuracy and average accuracy, significantly enhancing the robustness and practicality of hyperspectral image classification.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and in particular to a hyperspectral image classification network model and method based on deep learning. Background Technology

[0002] Hyperspectral images contain information from dozens to hundreds of consecutive spectral bands, providing rich spectral features and spatial texture information of ground features, and have wide applications in remote sensing, agriculture, environmental monitoring, and other fields. Hyperspectral image classification is one of the core tasks of hyperspectral remote sensing data processing, aiming to accurately classify each pixel in the image into its corresponding ground feature category.

[0003] Traditional hyperspectral image classification methods mainly rely on hand-designed feature extraction algorithms, such as shallow machine learning methods like Support Vector Machines (SVM) and Random Forests. These methods are prone to the Hughes phenomenon (when the number of training samples is limited, the accuracy of the classifier usually increases first, reaches a peak, and then decreases as the feature dimension increases) under conditions of small sample size and high-dimensional data, and they are difficult to fully explore the complex spectral-spatial joint features in hyperspectral data.

[0004] In recent years, the rapid development of deep learning technology has brought new solutions to hyperspectral image classification, but existing technologies still have the following shortcomings: Insufficient extraction of spectral-spatial features: Most methods only focus on spatial or spectral features, failing to effectively integrate complementary information from the spectrum and space; Insufficient utilization of frequency domain information: Existing methods mainly perform feature learning in the spatial domain, ignoring the advantages of frequency domain transformation in revealing global patterns and reducing computational complexity; A contradiction between computational efficiency and model complexity: Although standard convolution and self-attention mechanisms have strong expressive power, they have a large number of parameters and computational cost, which is not conducive to model deployment and real-time applications; Imperfect scale feature fusion mechanism: Different land cover features in hyperspectral images have different spatial scale features, and existing methods lack effective multi-scale adaptive feature extraction and fusion strategies; Insufficient generalization ability under small sample conditions: Hyperspectral data annotation is difficult and costly, and under limited labeled sample conditions, existing deep models are prone to overfitting and have limited generalization performance. Summary of the Invention

[0005] The purpose of this invention is to provide a hyperspectral image classification network model and method based on deep learning, which solves the technical problems of existing hyperspectral image classification models in terms of spectral redundancy suppression, frequency-space feature fusion and small sample generalization, thereby improving the robustness and practicality of hyperspectral image classification.

[0006] To achieve the above objectives, this invention provides a deep learning-based hyperspectral image classification network model, comprising a stem network, a spectral feature extraction module, a hybrid spatial spectrum fusion module, and a classifier connected in sequence. The stem network is used to perform preliminary channel expansion and spectral dimension downsampling on the input raw hyperspectral data; The spectral feature extraction module is used to extract spectral features from hyperspectral data; The hybrid spatial spectrum fusion module is used to further perform deep grouping and separable convolution and frequency-space attention on the spectral features extracted by the spectral feature extraction module, and then perform deep fusion to obtain high-dimensional features; The classifier is used to map the high-dimensional features into class probabilities.

[0007] Preferably, the stem network includes a three-dimensional convolutional layer for expanding the number of channels of the input data to a preset number of channels and performing downsampling with a stride of 2 in the spectral dimension, effectively reducing the number of spectral bands to: ; in, This represents the number of original spectral bands.

[0008] Preferably, the spectral feature extraction module includes a coarse pyramid spectral branch and a fine pyramid spectral branch connected in parallel; The coarse pyramid spectral branch consists of four parallel grouped three-dimensional convolutions, with kernel sizes of 3, 7, 11, and 15 used for each path. Each convolution kernel performs convolution operations only in the spectral dimension to capture the macroscopic trend features of the spectral curve. The fine pyramid spectral branch consists of 8 parallel grouped 3D convolutions, with kernel sizes of 3, 5, 7, 9, 11, 13, 15, and 17 for each channel. Each kernel performs convolution operation only in the spectral dimension to capture subtle changes in the spectral curve.

[0009] Preferably, each convolution kernel in the coarse pyramid spectral branch and the fine pyramid spectral branch is subjected to batch normalization and ReLU6 activation function after convolution.

[0010] Preferably, the hybrid spatial spectrum fusion module includes a grouped depthwise separable convolutional branch and a frequency-space dual-path attention branch connected in parallel, as well as a three-dimensional convolutional layer for fusing the outputs of the two branches.

[0011] Preferably, the grouped depth separable convolutional branch includes a dimension transformation layer, a grouped depth convolutional layer, a grouped convolutional layer, a grouped point convolutional layer, a residual connection and normalization layer, and a dimension recovery layer connected in sequence. The dimension transformation layer is used to merge the spectral dimension and the batch dimension in order to perform a two-dimensional convolution operation. The kernel size of the grouped depthwise convolutional layer is 3×3, and the number of groups is equal to the number of input channels; The kernel size of the grouped convolutional layer is 3×3, and the kernel size of the grouped point convolutional layer is 1×1; The dimension recovery layer is used to separate the spectral dimension from the batch dimension and restore the original dimension.

[0012] Preferably, the frequency-space dual-path attention branch includes a spatial branch and a frequency domain branch; The spatial branch maps features to the query, key, and value space through three-dimensional convolution, and models the long-distance dependency of the spatial dimension through a multi-head self-attention mechanism; The frequency domain branch performs a real-valued fast Fourier transform on the spectral dimension to convert the features to the frequency domain. The frequency domain features are grouped and processed. Attention weights on the spectral dimension are learned and generated through adaptive pooling and multilayer perceptron. These attention weights are then applied to the original spectral features. The output of the frequency-space dual-path attention branch is the result of weighted fusion of the output of the spatial branch and the frequency domain branch features weighted by the attention weights.

[0013] Preferably, the classifier performs global average pooling on the high-dimensional features across all dimensions, aggregates the feature maps of each sample into a fixed-length feature vector, flattens the feature vector, and then maps the flattened features to the category output through a fully connected layer.

[0014] A deep learning-based hyperspectral image classification method includes the following steps: S1. Acquire hyperspectral image data and input the data into the hyperspectral image classification network model; S2. Channel expansion and spectral dimension downsampling of hyperspectral image data are performed through stem network; S3. Input the features processed by the stem network into the coarse pyramid spectral branch and the fine pyramid spectral branch in parallel to extract macro trend features and subtle change features respectively. S4. The extracted spectral features are processed by the hybrid spatial spectrum fusion module to achieve deep fusion of spectral features and spatial features; S5. The fused features are subjected to global average pooling and fully connected mapping by a classifier to output the classification result.

[0015] Preferably, the mixture loss function of the hyperspectral image classification network model is: ; in, Represents a mixed loss function. The cross-entropy loss represents the weighted label smoothing. The weights represent the labeled smoothing cross-entropy loss. Represents category-weighted focus loss. The weighting coefficients represent the category-weighted focus loss; The calculation formula is as follows: ; ; in, Represents the smoothed target distribution. Representative category The weight, Represents the total number of categories. The representative model predicts the category to which the current sample belongs. The probability, Represents the smoothing parameter. Represents the original target distribution. This represents the Kronecker delta function.

[0016] Therefore, the present invention employs the above-mentioned deep learning-based hyperspectral image classification network model and method, which has the following advantages: (1) In this invention, a parallel collaborative architecture of coarse pyramid branch (CPB) and fine pyramid branch (FPB) is innovatively designed: In view of the characteristics of hyperspectral images with a large number of bands, strong correlation and redundancy, the combination of CPB and FPB achieves the complementarity of "macro trend" and "micro detail", which solves the problem that a single scale convolution kernel cannot take into account the overall spectral contour and local abrupt features. It significantly improves the model's ability to discriminate approximate spectra. Compared with standard 3D convolution, this design breaks the full connectivity between channels and greatly reduces the resource consumption of model training and inference. Compared with the traditional method of slowly expanding the receptive field by stacking small convolution kernels multiple times, the pyramid structure of this invention can directly and efficiently model the long-range dependence of spectral dimensions, thereby more accurately restoring the physical properties of ground objects.

[0017] (2) In this invention, a group separable convolution strategy is adopted to decompose the standard convolution into depthwise convolution, group convolution and group point convolution, which greatly reduces the number of parameters and is beneficial to the lightweight deployment and real-time inference of the model.

[0018] (3) In this invention, through the lightweight design of grouped separable convolution and the regularization effect of multi-level attention mechanism, the model can still maintain good generalization performance under limited labeled samples; when the proportion of training samples is low, it can still maintain high accuracy, effectively alleviating the problem of difficulty in labeling hyperspectral data.

[0019] (4) In this invention, the cross-entropy loss with label smoothing and class weights and the focus loss are innovatively integrated. The class weights directly increase the proportion of minority class samples in the total loss, ensuring that the model does not over-optimize the majority class. At the same time, the focus loss dynamically adjusts the sample weights, punishes easily classified samples, and forces the model to shift its learning focus to difficult-to-classify samples, thereby effectively mining "hard samples". This dual mechanism solves the common extreme class imbalance problem of hyperspectral data from different dimensions and avoids the model under-learning of rare classes and spectral aliasing samples.

[0020] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the structure of a hyperspectral image classification network model based on deep learning according to the present invention; Figure 2 This is a flowchart illustrating a hyperspectral image classification method based on deep learning according to the present invention. Figure 3 The experimental results of the PU dataset provided in this embodiment of the invention are shown in the figure. Figure 4 The experimental results of the IP dataset provided in this embodiment of the invention are shown in the figure. Figure 5 The figure shows the experimental results of the SA dataset provided in this embodiment of the invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Specific model specifications need to be selected and determined according to the actual specifications of the device, etc. The specific selection calculation method adopts existing technology in the art, and therefore will not be described in detail.

[0023] Example like Figure 1As shown, the present invention provides a hyperspectral image classification network model based on deep learning, comprising a stem network, a spectral feature extraction module, a hybrid spatial spectrum fusion module, and a classifier connected in sequence; The stem network is used to perform preliminary channel expansion and spectral dimension downsampling on the input raw hyperspectral data; The stem network includes a 3D convolutional layer to expand the input data from 1 channel to a preset number of channels, and performs downsampling with a stride of 2 in the spectral dimension, effectively reducing the number of spectral bands to: ; in, This represents the number of original spectral bands.

[0024] The spectral feature extraction module is used to extract spectral features from hyperspectral data; The spectral feature extraction module includes parallel-connected coarse pyramid spectral branches and fine pyramid spectral branches; The coarse pyramid spectral branch consists of four parallel grouped 3D convolutions, with kernel sizes of 3, 7, 11, and 15 for each path. Each kernel performs convolution operations only in the spectral dimension to capture the macroscopic trend features of the spectral curve. Each convolution is followed by batch normalization and a ReLU6 activation function. The input features are divided into four groups, each of which performs convolution operations independently in the spectral dimension. Large kernels (e.g., 15) can capture long-distance spectral dependence and low-frequency trends, while small kernels (e.g., 3) focus on local spectral patterns. The features output by each convolution are concatenated in the channel dimension.

[0025] By using parallel large-scale convolution kernels, coarse-grained contextual information of spectral data can be effectively obtained, and it has strong robustness to spectral noise. The grouping mechanism significantly reduces the number of parameters and computational complexity of spectral dimension convolution, alleviating the computational pressure brought about by the high dimensionality of hyperspectral data.

[0026] The fine pyramid spectral branch consists of eight parallel grouped three-dimensional convolutions, with kernel sizes of 3, 5, 7, 9, 11, 13, 15, and 17 for each. Each kernel performs convolution operations only in the spectral dimension to capture subtle changes in the spectral curve. Each convolution is followed by batch normalization (BN) and ReLU6 activation functions. By increasing the number of parallel branches and using denser kernel size selections (increment by 2), the fine pyramid spectral branch can more finely deconstruct spectral features and discover subtle differences in the spectral curve, which are often key to distinguishing similar ground features.

[0027] The denser kernel size coverage enables the model to more sensitively capture local high-frequency details and subtle changes in the spectral curve, improving the ability to distinguish spectral similarity categories. The increased number of groups further improves computational efficiency while maintaining the effectiveness of feature extraction.

[0028] The hybrid spatial spectrum fusion module is used to further perform deep grouping and separable convolution and frequency-space attention on the spectral features extracted by the spectral feature extraction module, and then perform deep fusion to obtain high-dimensional features; The hybrid spatial spectrum fusion module includes a grouped depthwise separable convolutional branch and a frequency-space dual-path attention branch connected in parallel, as well as a 3D convolutional layer for fusing the outputs of the two branches.

[0029] The grouped depthwise separable convolutional branch includes a dimension transformation layer, a grouped depthwise convolutional layer, a grouped convolutional layer, a grouped pointwise convolutional layer, a residual connection and normalization layer, and a dimension recovery layer connected in sequence. The dimension transformation layer is used to merge the spectral dimension and the batch dimension for two-dimensional convolution operations; the kernel size of the grouped depth convolutional layer is 3×3, and the number of groups is equal to the number of input channels; the kernel size of the grouped convolutional layer is 3×3, and the kernel size of the grouped point convolutional layer is 1×1; the dimension restoration layer is used to separate the spectral dimension and the batch dimension to restore the original dimension.

[0030] The structure with group depth separable convolutional branches is as follows in this embodiment: [B,C,H,W,S]→permute[BS,C,H,W] (merges the spectral dimension S with the batch B to facilitate 2D convolution operations); Grouped depthwise convolution: Conv2D(in_channels=C, out_channels=C, kernel_size=3, padding=1, groups=C); Grouped convolution: Conv2D(in_channels=C, out_channels=C_mid, kernel_size=3, groups=G); Grouped point convolution: Conv2D(in_channels=C_mid, out_channels=C, kernel_size=1, groups=G); Residual join + batch normalization; permute[B,C,H,W,S] (restores the original dimensions); Where C is the number of channels, H and W are the spatial dimensions, S is the spectral dimension, and G is the number of groups (G=8 in this embodiment).

[0031] The grouped depthwise separable convolution branches mainly extract features in the spatial dimension, avoiding the strong coupling of spatial and spectral information in 3D convolution, enabling the model to learn spatial texture features more independently.

[0032] The frequency-space dual-path attention branch is composed of ConvTE_Block, which includes a spatial branch and a frequency domain branch. The spatial branch maps features to the Query, Key, and Value spaces through 3D convolution, and models the long-distance dependencies of the spatial dimensions through a multi-head self-attention mechanism; The frequency domain branch performs a real fast fourier transform (RFFT) on the spectral dimension to transform the features to the frequency domain. The frequency domain features are grouped and processed. Attention weights on the spectral dimension are learned and generated through adaptive pooling and multilayer perceptron (MLP). These attention weights are then applied to the original spectral features to enhance key frequency components. The output of the frequency-space dual-path attention branch is the result of weighted fusion of the output of the spatial branch and the frequency domain branch features after being weighted by attention weights.

[0033] The hybrid spatial spectrum fusion module uses a three-dimensional convolutional layer (in_channels=32, out_channels=16, kernel_size=(1,1,1)) to concatenate and reduce the dimensionality of the outputs of the grouped depth-separable convolutional branch and the frequency-space dual-path attention branch.

[0034] The classifier is used to map high-dimensional features to class probabilities. The classifier performs global average pooling on the high-dimensional features across all dimensions, aggregating the feature maps of each sample into a fixed-length feature vector. After flattening the feature vector, a fully connected layer is used to map the flattened features to the class output. Global average pooling reduces model parameters, lowers the risk of overfitting, and increases the model's robustness to changes in the input space.

[0035] like Figure 2 As shown, a deep learning-based hyperspectral image classification method includes the following steps: S1. Acquire hyperspectral image data and input the data into the hyperspectral image classification network model; S2. Channel expansion and spectral dimension downsampling of hyperspectral image data are performed through stem network; S3. Input the features processed by the stem network into the coarse pyramid spectral branch and the fine pyramid spectral branch in parallel to extract macro trend features and subtle change features respectively. S4. The extracted spectral features are processed by the hybrid spatial spectrum fusion module to achieve deep fusion of spectral features and spatial features; S5. The fused features are subjected to global average pooling and fully connected mapping by a classifier to output the classification result.

[0036] The mixture loss function of the hyperspectral image classification network model is: ; The calculation formula is as follows: ; ; in, Represents the smoothed target distribution. Representative category The weight, Represents the total number of categories. The representative model predicts the category to which the current sample belongs. The probability, Represents the smoothing parameter. Represents the original target distribution. Represents the Kronecker delta function. Represents a mixed loss function. The cross-entropy loss, representing weighted label smoothing, smooths the true labels, softening the hard labels (one-hot encoding) into a probability distribution. This reduces the model's overconfidence in incorrect or noisy labels, acting as a regularization mechanism and effectively suppressing overfitting under few-shot training conditions. The weights represent the labeled smoothing cross-entropy loss. The focus loss, representing a class-weighted approach, dynamically adjusts the loss contribution of different samples by introducing a focus factor and class weights. It automatically reduces the weight of easily classified samples, forcing the model to shift its learning focus to difficult-to-classify samples, while simultaneously increasing the proportion of minority class samples in the total loss through the class weight mechanism. The weighting coefficients represent the focus loss weighted by the category.

[0037] The optimization strategy in this embodiment is as follows: The AdamW optimizer is selected, which combines the advantages of Adam's adaptive learning rate and weight decay regularization to effectively prevent overfitting; a cosine annealing learning rate scheduler is used to make the learning rate periodically decrease and increase during training, which helps the model escape local optima and converge to a better global optimum; gradient clipping is introduced to prevent gradient explosion during training; the patience parameter is set to 80, and training is stopped when the performance on the validation set does not improve within 80 consecutive epochs to avoid overfitting and save computational resources.

[0038] In this embodiment, validation experiments were conducted on three standard hyperspectral image datasets, including: The Pavia University (PU) dataset, acquired from the University of Pavia in Italy using the German Airborne Reflectance Spectroscopic Imager (ROSIS), contains a rich variety of urban features, such as asphalt roads, grass, bare soil, buildings, and metal panels. Due to its high ground sampling distance (GSD) of up to 1.3 meters, mixed pixels are prevalent in urban environments, meaning that a single pixel contains multiple features. This greatly increases the difficulty of pixel-level classification. Different man-made features (such as asphalt roads, roof tiles, and bare soil) may exhibit highly similar spectral characteristics in certain bands, posing a challenge to the model's fine-grained discrimination capabilities. The Indian Pines (IP) dataset was collected by AVIRIS sensors in the crop-growing region of northwestern Indiana, USA. This dataset has an extremely limited training sample size and a relatively small total number of samples, with many land cover categories such as oats, alfalfa, and wheat having extremely few samples (less than 100). Furthermore, the dataset suffers from severe class imbalance, with significant differences in the number of samples across different crop categories. The original spectral bands are as high as 220, exhibiting high dimensionality. The Salinas-A (SA) dataset, collected by AVIRIS sensors in the Salinas Valley, California, primarily contains a large number of vegetable crops, such as grapevines, tomatoes, and broccoli under different cultivation conditions. These crops are highly similar in macroscopic spectrum, requiring the model to have excellent ability to capture subtle textures and spectral differences. The presence of field ridges, shadows, and variations caused by different growth stages within the scene increases the classification complexity.

[0039] Experimental results on the PU dataset, where the training samples comprise 0.5% of the dataset, are as follows: Figure 3 As shown, the experimental results on the IP dataset, where the training samples account for 3% of the total dataset, are as follows: Figure 4 As shown, the experimental results on the SA dataset, where the training samples account for 0.5%, are as follows: Figure 5 As shown, the overall classification accuracy (OA), average accuracy (AA_mean), and Kappa coefficient stability are all high in the experimental results of the three datasets.

[0040] Therefore, the present invention adopts the above-mentioned deep learning-based hyperspectral image classification network model and method, which solves the technical problems of existing hyperspectral image classification models in terms of spectral redundancy suppression, frequency-space feature fusion and small sample generalization. Even with very few training samples, it can improve the overall classification accuracy and average accuracy, and significantly improve the robustness and practicality of hyperspectral image classification.

[0041] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A hyperspectral image classification network model based on deep learning, characterized in that: It includes a stem network, a spectral feature extraction module, a hybrid spatial spectrum fusion module, and a classifier connected in sequence; The stem network is used to perform preliminary channel expansion and spectral dimension downsampling on the input raw hyperspectral data; The spectral feature extraction module is used to extract spectral features from hyperspectral data; The hybrid spatial spectrum fusion module is used to further perform deep grouping and separable convolution and frequency-space attention on the spectral features extracted by the spectral feature extraction module, and then perform deep fusion to obtain high-dimensional features; The classifier is used to map the high-dimensional features into class probabilities.

2. The hyperspectral image classification network model based on deep learning according to claim 1, characterized in that: The stem network includes a three-dimensional convolutional layer to expand the number of channels of the input data to a preset number, and performs downsampling with a stride of 2 in the spectral dimension, effectively reducing the number of spectral bands to: ； in, This represents the number of original spectral bands.

3. The hyperspectral image classification network model based on deep learning according to claim 1, characterized in that: The spectral feature extraction module includes a coarse pyramid spectral branch and a fine pyramid spectral branch connected in parallel; The coarse pyramid spectral branch consists of four parallel grouped three-dimensional convolutions, with kernel sizes of 3, 7, 11, and 15 used for each path. Each convolution kernel performs convolution operations only in the spectral dimension to capture the macroscopic trend features of the spectral curve. The fine pyramid spectral branch consists of 8 parallel grouped 3D convolutions, with kernel sizes of 3, 5, 7, 9, 11, 13, 15, and 17 for each channel. Each kernel performs convolution operation only in the spectral dimension to capture subtle changes in the spectral curve.

4. The hyperspectral image classification network model based on deep learning according to claim 3, characterized in that: Each convolution kernel in the coarse pyramid spectral branch and the fine pyramid spectral branch is subjected to batch normalization and ReLU6 activation function after convolution.

5. The hyperspectral image classification network model based on deep learning according to claim 1, characterized in that: The hybrid spatial spectrum fusion module includes a parallel-connected grouped depthwise separable convolutional branch and a frequency-space dual-path attention branch, as well as a three-dimensional convolutional layer for fusing the outputs of the two branches.

6. The hyperspectral image classification network model based on deep learning according to claim 5, characterized in that: The grouped depth separable convolutional branch includes a dimension transformation layer, a grouped depth convolutional layer, a grouped convolutional layer, a grouped point convolutional layer, a residual connection and normalization layer, and a dimension recovery layer connected in sequence. The dimension transformation layer is used to merge the spectral dimension and the batch dimension in order to perform a two-dimensional convolution operation. The kernel size of the grouped depthwise convolutional layer is 3×3, and the number of groups is equal to the number of input channels; The kernel size of the grouped convolutional layer is 3×3, and the kernel size of the grouped point convolutional layer is 1×1; The dimension recovery layer is used to separate the spectral dimension from the batch dimension and restore the original dimension.

7. The hyperspectral image classification network model based on deep learning according to claim 6, characterized in that: The frequency-space dual-path attention branch includes a spatial branch and a frequency domain branch; The spatial branch maps features to the query, key, and value space through three-dimensional convolution, and models the long-distance dependency of the spatial dimension through a multi-head self-attention mechanism; The frequency domain branch performs a real-valued fast Fourier transform on the spectral dimension to convert the features to the frequency domain. The frequency domain features are grouped and processed. Attention weights on the spectral dimension are learned and generated through adaptive pooling and multilayer perceptron. These attention weights are then applied to the original spectral features. The output of the frequency-space dual-path attention branch is the result of weighted fusion of the output of the spatial branch and the frequency domain branch features weighted by the attention weights.

8. The hyperspectral image classification network model based on deep learning according to claim 1, characterized in that: The classifier performs global average pooling on high-dimensional features across all dimensions, aggregates the feature maps of each sample into a fixed-length feature vector, flattens the feature vector, and then maps the flattened features to the class output through a fully connected layer.

9. A hyperspectral image classification method based on deep learning, employing the hyperspectral image classification network model based on deep learning as described in any one of claims 1-8, characterized in that: Includes the following steps: S1. Acquire hyperspectral image data and input the data into the hyperspectral image classification network model; S2. Channel expansion and spectral dimension downsampling of hyperspectral image data are performed through stem network; S3. Input the features processed by the stem network into the coarse pyramid spectral branch and the fine pyramid spectral branch in parallel to extract macro trend features and subtle change features respectively. S4. The extracted spectral features are processed by the hybrid spatial spectrum fusion module to achieve deep fusion of spectral features and spatial features; S5. The fused features are subjected to global average pooling and fully connected mapping by a classifier to output the classification result.

10. A hyperspectral image classification method based on deep learning according to claim 9, characterized in that: The mixture loss function of the hyperspectral image classification network model is: ； in, Represents a mixed loss function. The cross-entropy loss represents the weighted label smoothing. The weights represent the labeled smoothing cross-entropy loss. Represents category-weighted focus loss. The weighting coefficients represent the category-weighted focus loss; The calculation formula is as follows: ； ； in, Represents the smoothed target distribution. Representative category The weight, Represents the total number of categories. The representative model predicts the category to which the current sample belongs. The probability, Represents the smoothing parameter. Represents the original target distribution. This represents the Kronecker delta function.

Images