Hyperspectral remote sensing image classification method, system and device based on deep learning

Abstract

The application belongs to the technical field of remote sensing image processing, and particularly relates to a hyperspectral remote sensing image classification method, system and device based on deep learning, which converts three-dimensional data into one-dimensional and performs preliminary feature extraction by generating a spectrum-space token; the processed features are decomposed into low-frequency features and high-frequency features through multi-scale wavelet decomposition, and then are respectively subjected to channel attention processing and spatial attention processing to respectively enhance the low-frequency features and the high-frequency features; the enhanced low-frequency features are unfolded along the spectrum dimension, and the enhanced high-frequency features are unfolded along the space dimension; a dynamic path weight is introduced in a selective state space scanning mechanism to realize adaptive feature fusion of the low-frequency features and the high-frequency features; and finally, a classification result is obtained according to a classification score.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing technology, specifically relating to a method, system, and device for classifying hyperspectral remote sensing images based on deep learning. Background Technology

[0002] Hyperspectral remote sensing image classification is an important research direction in the field of remote sensing, with broad application prospects. Unlike ordinary RGB images, which only contain three bands (red, green, and blue), hyperspectral remote sensing images typically cover 200 bands ranging from 400 to 1000 nanometers. Hyperspectral data possesses both spatial resolution and spectral dimension, forming a three-dimensional data cube, meaning that each pixel contains rich spectral information. Due to the high-dimensionality, spectral-spatial correlation, and computational complexity of hyperspectral data, traditional monochrome, RGB, and multispectral image analysis techniques cannot be directly applied to hyperspectral remote sensing image classification. Instead, specialized adaptation methods are required to effectively extract spectral-spatial features. Summary of the Invention

[0003] This invention provides a method, system, and apparatus for classifying hyperspectral remote sensing images based on deep learning.

[0004] The technical solution of the present invention is as follows:

[0005] This invention provides a deep learning-based hyperspectral remote sensing image classification method, comprising:

[0006] S1: Acquire hyperspectral remote sensing image data, perform dimensionality transformation after 3D convolution processing, generate spectral-spatial tokens, extract features from spectral-spatial tokens through 3D convolution with different dilation rates, and then stitch and fuse the extracted features to obtain fused features.

[0007] S2: Based on the wavelet basis type, the fused features are used to perform two-dimensional wavelet decomposition to obtain the low-frequency and high-frequency features of the corresponding wavelet basis type; using learnable fusion weights, the low-frequency and high-frequency features of different wavelet basis types are weighted and summed to obtain the low-frequency fused features and high-frequency fused features.

[0008] The low-frequency fusion features and high-frequency fusion features are processed by channel attention and spatial attention, respectively, to obtain low-frequency enhancement features and high-frequency enhancement features;

[0009] S3: The low-frequency enhancement feature is scanned along the spectral dimension to obtain the first scan feature. The first scan feature is then inverted to obtain the second scan feature.

[0010] The high-frequency enhancement feature is scanned along the spatial dimension to obtain the third scan feature. The third scan feature is then inverted to obtain the fourth scan feature.

[0011] By using learnable parameters, the first scan feature, the second scan feature, the third scan feature, and the fourth scan feature are fused to obtain low-frequency scan features and high-frequency scan features.

[0012] After being processed by a dynamic gating mechanism, the high-frequency scanning features are fused with the low-frequency scanning features to obtain the target features;

[0013] S4: After global average pooling and linear processing, the target features are used to generate classification scores, which are then used for classification.

[0014] In step S1, the spectral-spatial token extracts features through three-dimensional convolution with different dilation rates. The extracted features are then sequentially concatenated and fused to obtain fused features, specifically as follows:

[0015] According to the formula: F s =σ(BN(Conv3D) s (T)),s∈{1,2,4}, extract features;

[0016] In the formula, F s For the extracted features, σ is the activation function, BN represents normalization, and Conv3D s The three-dimensional convolutions with different dilation rates are represented by T, which represents the spectral-spatial token, and s is the dilation rate.

[0017] The extracted features are concatenated along the channel dimension, and then the concatenated features are fused through convolution to obtain the fused features.

[0018] S2, based on wavelet basis type, performs two-dimensional wavelet decomposition by fusing features to obtain the low-frequency and high-frequency features of the corresponding wavelet basis type, specifically:

[0019] From the formula: cA i =X*W low This yields the low-frequency features of the corresponding wavelet basis type;

[0020] From the formula: cH i =X*W high-H , cV i =X*W high-V , cD i =X*W high-D ,and The high-frequency features of the corresponding wavelet basis type are obtained;

[0021] In the formula, cA i high_freq i These represent low-frequency and high-frequency features, respectively; X represents the fused feature; cH i cV i , cD iThese represent horizontal high-frequency features, vertical high-frequency features, and diagonal high-frequency features, respectively; W low W high-H W high-V W high-D These represent low-frequency, horizontal high-frequency, vertical high-frequency, and diagonal high-frequency filters, respectively.

[0022] After fusing features and performing two-dimensional wavelet decomposition, bilinear interpolation is also included.

[0023] In step S2, learnable fusion weights are used to perform weighted summation on low-frequency and high-frequency features of different wavelet basis types to obtain low-frequency fusion features and high-frequency fusion features, specifically:

[0024] According to the formula: accomplish;

[0025] In the formula, It is a low-frequency fusion feature. For the low-frequency characteristics of wavelet basis type i, α i For learnable fusion weights, M is the number of wavelet basis types. High-frequency fusion characteristics, These are the high-frequency features of wavelet basis type i.

[0026] In step S3, the first scanning feature, the second scanning feature, the third scanning feature, and the fourth scanning feature are fused using learnable parameters to obtain low-frequency scanning features and high-frequency scanning features, specifically:

[0027] From the formula: Obtain low-frequency scanning features;

[0028] From the formula: High-frequency sweep characteristics were obtained;

[0029] In the formula, Y low Y high These represent low-frequency scanning features and high-frequency scanning features, respectively; λ1, λ2, λ3, and λ4 represent the normalized weights; X low,spe , X high,spa , These represent the first scan feature, the second scan feature, the third scan feature, and the fourth scan feature, respectively.

[0030] This invention also provides a deep learning-based hyperspectral remote sensing image classification system, comprising:

[0031] Feature extraction module: used to acquire hyperspectral remote sensing image data, perform dimensionality transformation after 3D convolution processing, generate spectral-spatial tokens, extract features through 3D convolution with different dilation rates, and then stitch and fuse the extracted features to obtain fused features;

[0032] Feature enhancement module: Based on wavelet basis type, the fused features are decomposed into two-dimensional wavelet decomposition to obtain low-frequency and high-frequency features of the corresponding wavelet basis type; using learnable fusion weights, the low-frequency and high-frequency features of different wavelet basis types are weighted and summed to obtain low-frequency fused features and high-frequency fused features.

[0033] The low-frequency fusion features and high-frequency fusion features are processed by channel attention and spatial attention, respectively, to obtain low-frequency enhancement features and high-frequency enhancement features;

[0034] Scanning module: The low-frequency enhancement feature is scanned along the spectral dimension to obtain the first scanning feature. The first scanning feature is then inverted to obtain the second scanning feature.

[0035] The high-frequency enhancement feature is scanned along the spatial dimension to obtain the third scan feature. The third scan feature is then inverted to obtain the fourth scan feature.

[0036] By using learnable parameters, the first scan feature, the second scan feature, the third scan feature, and the fourth scan feature are fused to obtain low-frequency scan features and high-frequency scan features.

[0037] After being processed by a dynamic gating mechanism, the high-frequency scanning features are fused with the low-frequency scanning features to obtain the target features;

[0038] Classification module: After global average pooling and linear processing, the target features are used to generate classification scores, and classification is performed based on the classification scores.

[0039] The present invention also provides a hyperspectral remote sensing image classification device based on deep learning, including a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the hyperspectral remote sensing image classification method based on deep learning.

[0040] Beneficial effects

[0041] This invention presents a deep learning-based hyperspectral remote sensing image classification method. It generates spectral-spatial tokens to convert three-dimensional data into one-dimensional data for preliminary feature extraction. The features to be processed are then decomposed into low-frequency and high-frequency features through multi-scale wavelet decomposition. These features are further enhanced by channel attention and spatial attention processing, respectively. The enhanced low-frequency features are expanded along the spectral dimension, and the enhanced high-frequency features are expanded along the spatial dimension. Dynamic path weights are introduced into a selective state-space scanning mechanism to achieve adaptive feature fusion of low-frequency and high-frequency features. Finally, the classification result is obtained based on the classification score. Attached Figure Description

[0042] Figure 1 The following are classification results using different hyperspectral remote sensing image classification methods on the Pavia University dataset: (a) is the false-color image of the dataset; (b) is the ground truth label image of the dataset; (c) is the classification result of the 3D CNN method; (d) is the classification result of the SPRN method; (e) is the classification result of the CEGCN method; (f) is the classification result of the SSFTT method; (g) is the classification result of the MorphFormer method; (h) is the classification result of the SS-Mamba method; (i) is the classification result of the 3DSS-Mamba method; and (j) is the classification result of the method of this invention.

[0043] Figure 2 The following are classification results of different hyperspectral remote sensing image classification methods on the Indian Pines dataset: (a) is the false-color image of the dataset; (b) is the ground truth label image of the dataset; (c) is the classification result of the 3D CNN method; (d) is the classification result of the SPRN method; (e) is the classification result of the CEGCN method; (f) is the classification result of the SSFTT method; (g) is the classification result of the MorphFormer method; (h) is the classification result of the SS-Mamba method; (i) is the classification result of the 3DSS-Mamba method; and (j) is the classification result of the method of this invention.

[0044] Figure 3The following are classification results of different hyperspectral remote sensing image classification methods on the Houston 2013 dataset: (a) is the false-color image of the dataset; (b) is the ground truth label image of the dataset; (c) is the classification result of the 3D CNN method; (d) is the classification result of the SPRN method; (e) is the classification result of the CEGCN method; (f) is the classification result of the SSFTT method; (g) is the classification result of the MorphFormer method; (h) is the classification result of the SS-Mamba method; (i) is the classification result of the 3DSS-Mamba method; and (j) is the classification result of the method of this invention. Detailed Implementation

[0045] The following examples are intended to illustrate the present invention, and not to further limit the invention.

[0046] This invention provides a deep learning-based hyperspectral remote sensing image classification method, comprising:

[0047] S1: Acquire hyperspectral remote sensing image data, perform dimensionality transformation after 3D convolution processing, generate spectral-spatial tokens, extract features from the spectral-spatial tokens through 3D convolution with different dilation rates, and then stitch and fuse the extracted features to obtain fused features.

[0048] The specific steps are as follows:

[0049] First, the hyperspectral remote sensing image data (3D cube data) is dimensionality reduced and divided into multiple pixel-level 3D patch cubes, denoted as x, x∈R. B×B×d Where B is the size of the spatial dimension (height and width), and d is the number of bands.

[0050] Next, it undergoes three-dimensional convolution processing, which consists of a three-dimensional convolutional layer, a normalization layer, and a ReLU activation function in sequence.

[0051] Then, a dimensionality transformation is performed through a linear layer of embedding operations to generate spectral-spatial tokens. The tokenization process can be represented as:

[0052] T = Φ embed (φ 3DConv (x));

[0053] In the formula, φ embed Indicates the embedding operation; φ 3DConv Represents 3D convolution processing; T represents spectral-spatial token; and T∈R N×P×P×K N is the number of generated markers, each marker being a P×P two-dimensional spatial feature map. K is the feature dimension of each marker, with each marker having K feature values ​​at each spatial location, capturing the spectral and spatial information of that local region.

[0054] Subsequently, multi-scale features are extracted through 3D convolutions with different dilation rates to capture local and global information. Normalization is used to accelerate training and stabilize gradients, while activation functions are employed to enhance the non-linear expressive power of the features. The formula for this process is:

[0055] F s =σ(BN(Conv3D) s (T)),s∈{1,2,4};

[0056] In the formula, F s For the extracted features, σ is the activation function, BN represents normalization, and Conv3D s The three-dimensional convolutions with different dilation rates are represented by T, which represents the spectral-spatial token, and s is the dilation rate.

[0057] After the extracted features at different scales are concatenated along the channel dimension, the concatenated features are fused using a convolution with a kernel of 1 to obtain the fused features.

[0058] S2: Based on the wavelet basis type, the fused features are used to perform two-dimensional wavelet decomposition to obtain the low-frequency and high-frequency features of the corresponding wavelet basis type; using learnable fusion weights, the low-frequency and high-frequency features of different wavelet basis types are weighted and summed to obtain the low-frequency fused features and high-frequency fused features.

[0059] The low-frequency fusion features and high-frequency fusion features are processed by channel attention and spatial attention, respectively, to obtain low-frequency enhancement features and high-frequency enhancement features.

[0060] The specific steps are as follows:

[0061] First, three wavelet bases are defined: Haar, Daubechies-2, and Symlet-2. Based on the wavelet base type, two-dimensional wavelet decomposition is performed by fusing features to obtain the low-frequency and high-frequency features of the corresponding wavelet base type, specifically:

[0062] From the formula: cA i =X*W low This yields the low-frequency features of the corresponding wavelet basis type;

[0063] From the formula: cH i =X*W high-H , cV i =X*W high-V , cD i =X*W high-D ,and The high-frequency features of the corresponding wavelet basis type are obtained;

[0064] In the formula, cA ihigh_freq i These represent low-frequency and high-frequency features, respectively; X represents the fused feature; cH i cV i , cD i These represent horizontal high-frequency features, vertical high-frequency features, and diagonal high-frequency features, respectively; W low W high-H W high-V W high-D These represent low-frequency, horizontal high-frequency, vertical high-frequency, and diagonal high-frequency filters, respectively.

[0065] Next, since wavelet decomposition leads to a reduction in spatial size, after fusing features and performing two-dimensional wavelet decomposition, bilinear interpolation is also included to restore the original size. The formula for this process is:

[0066]

[0067] In the formula, Upscale represents bilinear interpolation; H and W represent the height and width before two-dimensional wavelet decomposition, respectively; cA i high_freq i These represent low-frequency characteristics and high-frequency characteristics, respectively. These represent the low-frequency and high-frequency characteristics of wavelet basis type i, respectively.

[0068] Then, different wavelet decomposition results contribute differently to the final feature extraction, therefore a learnable fusion weight α is used. i They are then summed using a weighted average. The formula for this process is:

[0069]

[0070] In the formula, It is a low-frequency fusion feature. For the low-frequency characteristics of wavelet basis type i, α i For learnable fusion weights, M is the number of wavelet basis types. High-frequency fusion characteristics, These are the high-frequency features of wavelet basis type i.

[0071] Subsequently, the low-frequency fusion features and high-frequency fusion features are processed by channel attention and spatial attention, respectively, to obtain low-frequency enhancement features and high-frequency enhancement features.

[0072] For example, for low-frequency fusion features, global average pooling and global max pooling are first used to extract global information in the channel dimension, and channel attention weights are calculated. Then, depthwise separable convolution operations are used to further enhance the features, and finally, the enhanced features are weighted and fused together using the attention weights.

[0073] For high-frequency fusion features, spatial attention information is first extracted through channel average pooling and channel max pooling to capture key spatial structural features. Subsequently, depthwise separable convolution operations are used to enhance the extracted high-frequency features, and the enhanced features are then weighted and fused using spatial attention weights.

[0074] Step S2 uses wavelet decomposition to perform multi-scale analysis on the hyperspectral remote sensing image, which can effectively capture the local and global features of the hyperspectral remote sensing image and enhance the expressive power of spectral-spatial information.

[0075] S3: The low-frequency enhancement feature is scanned along the spectral dimension to obtain the first scan feature. The first scan feature is then inverted to obtain the second scan feature.

[0076] The high-frequency enhancement feature is scanned along the spatial dimension to obtain the third scan feature. The third scan feature is then inverted to obtain the fourth scan feature.

[0077] By using learnable parameters, the first scan feature, the second scan feature, the third scan feature, and the fourth scan feature are fused to obtain low-frequency scan features and high-frequency scan features.

[0078] After being processed by a dynamic gating mechanism, the high-frequency scanning features are fused with the low-frequency scanning features to obtain the target features.

[0079] The specific steps are as follows:

[0080] First, two scanning methods are used for low-frequency enhancement features: spectral priority (along the spectral dimension) and spatial priority (along the spatial dimension), respectively. Through a reshaping operation, each feature is expanded and rearranged in different dimensions, resulting in four scanning paths. Specifically, the low-frequency enhancement feature is expanded along the spectral dimension to obtain the first scanning feature X. low,spe The first scan feature is inverted to obtain the second scan feature.

[0081] The high-frequency enhancement features are expanded along the spatial dimension to obtain the third scanning feature X. high,spa The third scan feature is inverted to obtain the fourth scan feature.

[0082] X low,spe X high,spa The above reversal operation enhances orientation perception. Four scanning features are used to obtain a complete feature representation X. scan ,

[0083] Then, by setting a learnable parameter and calculating the weights of the four scanning paths using Softmax normalization, the process of obtaining the low-frequency scanning features can be represented as follows:

[0084]

[0085] Similarly, the process of obtaining high-frequency scanning features can be represented as:

[0086]

[0087] In the formula, Y low Y high These represent low-frequency scanning features and high-frequency scanning features, respectively; λ1, λ2, λ3, and λ4 represent the normalized weights, where λ1 and λ2 correspond to the low-frequency scanning methods along the spectral dimension and their reverse, respectively, and λ3 and λ4 correspond to the high-frequency scanning methods along the spatial dimension and their reverse, respectively; X low,spe , X high,spa , These represent the first scan feature, the second scan feature, the third scan feature, and the fourth scan feature, respectively.

[0088] Since low-frequency scanning features provide global structure and ensure spectral consistency, while high-frequency scanning features provide local details and make boundaries clearer, a dynamic gating mechanism is used to calculate the attention weights of the high-frequency scanning features, ensuring that the contributions of high-frequency scanning features from different regions to the final result vary. Finally, the low-frequency scanning features are fused with the weighted high-frequency scanning features, a process that can be represented as follows:

[0089] Y final =Y low +W high ·Y high ;

[0090] In the formula, W high Calculated by a high-frequency gating mechanism, it ensures that the high-frequency sweep features of different regions can adaptively influence the final result, thereby improving the feature representation capability.

[0091] Step S3 not only employs two scanning methods for low-frequency enhancement features—spectral priority (along the spectral dimension) and spatial priority (along the spatial dimension)—but also introduces dynamic path weights into the selective state space scanning mechanism to achieve adaptive feature fusion of low-frequency and high-frequency features. Furthermore, it enables efficient information propagation, achieves deep feature fusion across bands, and improves classification accuracy.

[0092] S4: After global average pooling and linear processing, the target features are used to generate classification scores, which are then used for classification.

[0093] The specific process includes:

[0094] First, global average pooling is performed on the target features to reduce feature dimensionality while preserving global information. Next, a linear layer is used as the classification head. The pooled features are projected onto the `num_classes` dimension through the linear layer to generate classification scores. `num_classes` represents the number of categories in the classification task, such as 9 or 16 land cover types. Classification is then performed based on the classification scores.

[0095] This invention presents a deep learning-based hyperspectral remote sensing image classification method. It generates spectral-spatial tokens to convert three-dimensional data into one-dimensional data for preliminary feature extraction. The features to be processed are then decomposed into low-frequency and high-frequency features through multi-scale wavelet decomposition. These features are further enhanced by channel attention and spatial attention processing, respectively. The enhanced low-frequency features are expanded along the spectral dimension, and the enhanced high-frequency features are expanded along the spatial dimension. Dynamic path weights are introduced into a selective state-space scanning mechanism to achieve adaptive feature fusion of low-frequency and high-frequency features. Finally, the classification result is obtained based on the classification score.

[0096] Experimental results

[0097] Different hyperspectral remote sensing image classification methods were used to classify the Pavia University, Indian Pines, and Houston 2013 datasets. The classification methods included 3D CNN, SPRN, CEGCN, SSFTT, MorphFormer, SS-Mamba, 3DSS-Mamba, and the method of this invention.

[0098] Among them, 3D CNN is a classic hyperspectral remote sensing image classification method. It has significant advantages in datasets with large labeled data, but it is difficult to effectively extract spectral and spatial features in small datasets.

[0099] SPRN is a classification method based on convolutional neural networks and spectral segmentation. It has a high degree of extraction of spectral features but a low degree of extraction of spatial features, resulting in less than ideal classification results.

[0100] CEGCN combines graph convolutional networks and reinforcement learning. It constructs a spectral-spatial feature cube and integrates inter-sample correlation information using graph convolutional modules.

[0101] SSFTT combines CNN and Transformer. Convolutional layers extract low-level spectral-spatial features and convert them into semantic tokens, while the Transformer encoder models their high-level semantic information.

[0102] MorphFormer combines morphological operations with the Transformer architecture, outperforming traditional Transformer and CNN models in complex scenarios.

[0103] SS-Mamba mainly consists of a spectral-spatial token generation module and multiple superimposed spectral-spatial Mmaba blocks. Although it has lower computational complexity than the Transformer model, its overall classification performance is weaker.

[0104] 3DSS-Mamba captures the global spatial-spectral correlation of hyperspectral remote sensing images through a multi-directional scanning strategy, but it is weak in processing local features and performs poorly in classification on small datasets.

[0105] Depend on Figure 1 , Figure 2 , Figure 3 It can be seen that, compared with other classification methods, the classification method of this invention has a higher classification accuracy. In small datasets, such as... Figure 2 The performance shown is better.

[0106] This invention also provides a deep learning-based hyperspectral remote sensing image classification system, comprising:

[0107] Feature extraction module: used to acquire hyperspectral remote sensing image data, perform dimensionality transformation after 3D convolution processing, generate spectral-spatial tokens, extract features through 3D convolution with different dilation rates, and then stitch and fuse the extracted features to obtain fused features;

[0108] Feature enhancement module: Based on wavelet basis type, the fused features are decomposed into two-dimensional wavelet decomposition to obtain low-frequency and high-frequency features of the corresponding wavelet basis type; using learnable fusion weights, the low-frequency and high-frequency features of different wavelet basis types are weighted and summed to obtain low-frequency fused features and high-frequency fused features.

[0109] The low-frequency fusion features and high-frequency fusion features are processed by channel attention and spatial attention, respectively, to obtain low-frequency enhancement features and high-frequency enhancement features;

[0110] Scanning module: Low-frequency enhancement features are scanned along the spectral dimension to obtain the first scanning feature and the second scanning feature;

[0111] The high-frequency enhanced features are scanned along the spatial dimension to obtain the third and fourth scan features;

[0112] By using learnable parameters, the first scan feature, the second scan feature, the third scan feature, and the fourth scan feature are fused to obtain low-frequency scan features and high-frequency scan features.

[0113] After being processed by a dynamic gating mechanism, the high-frequency scanning features are fused with the low-frequency scanning features to obtain the target features;

[0114] Classification module: After global average pooling and linear processing, the target features are used to generate classification scores, and classification is performed based on the classification scores.

[0115] The present invention also provides a hyperspectral remote sensing image classification device based on deep learning, including a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the hyperspectral remote sensing image classification method based on deep learning.

Claims

1. A hyperspectral remote sensing image classification method based on deep learning, characterized in that, include: S1: Acquire hyperspectral remote sensing image data, perform dimensionality transformation after 3D convolution processing, generate spectral-spatial tokens, extract features from spectral-spatial tokens through 3D convolution with different dilation rates, and then stitch and fuse the extracted features to obtain fused features. S2: Based on the wavelet basis type, the features are fused to perform two-dimensional wavelet decomposition to obtain the low-frequency and high-frequency features of the corresponding wavelet basis type; By using learnable fusion weights, the low-frequency features and high-frequency features of different wavelet basis types are weighted and summed to obtain the low-frequency fusion features and high-frequency fusion features. The low-frequency fusion features and high-frequency fusion features are processed by channel attention and spatial attention, respectively, to obtain low-frequency enhancement features and high-frequency enhancement features; S3: The low-frequency enhancement feature is scanned along the spectral dimension to obtain the first scan feature. The first scan feature is then inverted to obtain the second scan feature. The high-frequency enhancement feature is scanned along the spatial dimension to obtain the third scan feature. The third scan feature is then inverted to obtain the fourth scan feature. By using learnable parameters, the first scan feature, the second scan feature, the third scan feature, and the fourth scan feature are fused to obtain low-frequency scan features and high-frequency scan features. After being processed by a dynamic gating mechanism, the high-frequency scanning features are fused with the low-frequency scanning features to obtain the target features; S4: After global average pooling and linear processing, the target features are used to generate classification scores, which are then used for classification.

2. The hyperspectral remote sensing image classification method based on deep learning according to claim 1, characterized in that, In step S1, the spectral-spatial token extracts features through three-dimensional convolution with different dilation rates. The extracted features are then sequentially concatenated and fused to obtain fused features, specifically as follows: According to the formula: F s =σ(BN(Conv3D) s (T)),s∈{1,2,4}, extract features; In the formula, F s For the extracted features, σ is the activation function, BN represents normalization, and Conv3D s The three-dimensional convolutions with different dilation rates are represented by T, which represents the spectral-spatial token, and s is the dilation rate. The extracted features are concatenated along the channel dimension, and then the concatenated features are fused through convolution to obtain the fused features.

3. The hyperspectral remote sensing image classification method based on deep learning according to claim 1, characterized in that, S2, based on wavelet basis type, performs two-dimensional wavelet decomposition by fusing features to obtain the low-frequency and high-frequency features of the corresponding wavelet basis type, specifically: From the formula: cA i =X*W low This yields the low-frequency features of the corresponding wavelet basis type; From the formula: cH i =X*W high-H , cV i =X*W high-V , cD i =X*W high-D ,and The high-frequency features of the corresponding wavelet basis type are obtained; In the formula, cA i high_freq i These represent low-frequency and high-frequency features, respectively; X represents the fused feature; cH i cV i , cD i These represent horizontal high-frequency features, vertical high-frequency features, and diagonal high-frequency features, respectively. W low W high-H W high-V W high-D These represent low-frequency, horizontal high-frequency, vertical high-frequency, and diagonal high-frequency filters, respectively.

4. The hyperspectral remote sensing image classification method based on deep learning according to claim 1, characterized in that, In step S2, learnable fusion weights are used to perform weighted summation on low-frequency and high-frequency features of different wavelet basis types to obtain low-frequency fusion features and high-frequency fusion features, specifically: According to the formula: accomplish; In the formula, It is a low-frequency fusion feature. For the low-frequency characteristics of wavelet basis type i, α i For learnable fusion weights, M is the number of wavelet basis types. High-frequency fusion characteristics, These are the high-frequency features of wavelet basis type i.

5. The hyperspectral remote sensing image classification method based on deep learning according to claim 1, characterized in that, In step S3, the first scanning feature, the second scanning feature, the third scanning feature, and the fourth scanning feature are fused using learnable parameters to obtain low-frequency scanning features and high-frequency scanning features, specifically: From the formula: Obtain low-frequency scanning features; From the formula: High-frequency sweep characteristics were obtained; In the formula, Y low Y high These represent low-frequency scanning features and high-frequency scanning features, respectively; λ1, λ2, λ3, and λ4 represent the normalized weights; X low,spe , X high,spa , These represent the first scan feature, the second scan feature, the third scan feature, and the fourth scan feature, respectively.

6. The hyperspectral remote sensing image classification method based on deep learning according to claim 3, characterized in that, After fusing features and performing two-dimensional wavelet decomposition, bilinear interpolation is also included.

7. A hyperspectral remote sensing image classification system based on deep learning, characterized in that, include: Feature extraction module: used to acquire hyperspectral remote sensing image data, perform dimensionality transformation after 3D convolution processing, generate spectral-spatial tokens, extract features through 3D convolution with different dilation rates, and then stitch and fuse the extracted features to obtain fused features; Feature enhancement module: Based on wavelet basis type, fuse features to perform two-dimensional wavelet decomposition to obtain low-frequency and high-frequency features of the corresponding wavelet basis type; By using learnable fusion weights, the low-frequency features and high-frequency features of different wavelet basis types are weighted and summed to obtain the low-frequency fusion features and high-frequency fusion features. The low-frequency fusion features and high-frequency fusion features are processed by channel attention and spatial attention, respectively, to obtain low-frequency enhancement features and high-frequency enhancement features; Scanning module: The low-frequency enhancement feature is scanned along the spectral dimension to obtain the first scanning feature. The first scanning feature is then inverted to obtain the second scanning feature. The high-frequency enhancement feature is scanned along the spatial dimension to obtain the third scan feature. The third scan feature is then inverted to obtain the fourth scan feature. By using learnable parameters, the first scan feature, the second scan feature, the third scan feature, and the fourth scan feature are fused to obtain low-frequency scan features and high-frequency scan features. After being processed by a dynamic gating mechanism, the high-frequency scanning features are fused with the low-frequency scanning features to obtain the target features; Classification module: After global average pooling and linear processing, the target features are used to generate classification scores, and classification is performed based on the classification scores.

8. A hyperspectral remote sensing image classification device based on deep learning, characterized in that, It includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the deep learning-based hyperspectral remote sensing image classification method as described in any one of claims 1-6.

Images