Hyperspectral Image Band Selection Method Applicable to Complex Geographic Scenes

Abstract

This invention relates to a hyperspectral image band selection method applicable to complex terrain scenes, comprising: inputting a raw 3D hyperspectral image; extracting local spatial features and local spectral features through a heterogeneous dual-channel 3D convolutional network with a cross-attention mechanism; adding spatial location encoding and spectral location encoding to the local spatial features and local spectral features respectively; inputting the encoded features into a dual-channel Transformer module to generate band attention weights; constructing a band evaluation criterion based on the redundancy between bands attended by multi-head attention in the Transformer module and the band attention weights; and selecting the required bands after sorting the bands. The beneficial effects of this invention are: it fully considers the global-local and spatial-spectral features between bands, providing reliable support for subsequent applications. Therefore, the method proposed in this invention has significant practical application value.

Description

Technical Field

[0001] This invention belongs to the field of hyperspectral imaging technology, and in particular relates to a hyperspectral image band selection method applicable to complex terrain scenes. Background Technology

[0002] Hyperspectral images consist of hundreds of continuous narrow bands, containing rich spatial and spectral information. They possess continuous and smooth spectral curves for ground features, with each feature exhibiting distinct spectral characteristics in each band. Hyperspectral data is far superior to multispectral and panchromatic data in ground feature identification. It is widely used in agriculture and land management, urban planning, and environmental assessment. However, the hundreds or thousands of highly correlated adjacent bands lead to severe "information redundancy," diluting effective information. Furthermore, this ultra-high dimensionality can trigger the "curse of dimensionality": with limited training samples, the data is extremely sparse in high-dimensional space, making it difficult for classifiers to learn robust discrimination boundaries. This makes them prone to "overfitting" noise, significantly increasing classification difficulty and reducing accuracy. This phenomenon is particularly pronounced in the "Hughes phenomenon," where classification accuracy initially increases and then decreases with the number of bands. The root cause lies in the sharp contradiction between the exponential growth of model complexity (required sample size) with dimensionality and the scarcity of training samples in reality.

[0003] Feature extraction and band selection are effective dimensionality reduction methods. Feature extraction transforms the original high-dimensional bands using linear or nonlinear methods to achieve dimensionality reduction, but this method can cause hyperspectral images to lose their original physical meaning. Band selection, on the other hand, selects several representative bands from the original bands using certain criteria or methods, which better preserves the physical meaning of the bands and makes them more interpretable. Therefore, band selection is a more commonly used hyperspectral dimensionality reduction method.

[0004] While traditional band selection methods are effective in certain applications, they rely heavily on linear assumptions or shallow features, making it difficult to effectively capture the complex nonlinear relationships between spectral bands. They also lack sensitivity to subtle but crucial spectral shape changes and are susceptible to noise. In recent years, deep learning-based hyperspectral image band selection algorithms have been extensively studied. These methods train deep learning networks to select the optimal band combinations, directly extracting nonlinear features from the raw data and optimizing potential correlations between bands, thereby improving the accuracy and efficiency of band selection. Zhan et al. pioneered the application of Convolutional Neural Networks (CNNs) to band selection. This method trains a base model using the original HSI dataset, then uses this model to randomly generate various band combinations for validation testing, determining the optimal subset based on classification accuracy. This groundbreaking work is the first to validate the feasibility of deep learning in hyperspectral band selection. To address the inefficiency of randomly selected band combinations, Zhan et al. further developed the DDCNN (DistanceDensity CNN) algorithm, which innovatively integrates distance density analysis and subspace partitioning techniques. By quantifying the similarity between bands, the spectral space is deconstructed into several sub-regions. Feature bands are selected within each sub-region to form candidate sets, and then a pre-trained CNN model is used to evaluate the classification performance of each subset. Similarly, Chamisi et al. proposed an intelligent search strategy based on the Fast Optimized Discrete Particle Swarm Optimization (FODPSO) algorithm, combined with a CNN classifier to construct a dynamic evaluation system. Tschannerl et al. proposed the Segmented Autoencoders Band Selection (S-AEBS) algorithm, which innovatively introduces a sparse autoencoder structure while maintaining the spectral partitioning framework, identifying key bands by analyzing the weight distribution of network layers. Based on the idea of ​​subspace partitioning, Zeng et al. proposed the Deep Subspace Clustering (DSC) method, achieving a technological breakthrough. This algorithm combines a two-dimensional convolutional autoencoder with a self-expression mechanism, effectively capturing nonlinear correlation features in spectral data. Notably, the Deep Latent Spectral Representation Learning (DLSRL) model developed by Xie et al. enhances spectral consistency. Feng et al. introduced an embedded approach, integrating land cover classification and band selection, and using classification accuracy to optimize network parameters; however, this method suffers from slow network parameter optimization. To address this issue, Feng et al. proposed a hyperspectral image band selection method based on a ternary weighted CNN.To address the issues of low classification accuracy and efficiency in embedded methods, Feng et al. introduced the concept of graph convolution kernel graph attention into band selection in 2021, proposing a dual-graph convolutional network based on band attention and sparse constraints (BSD-GCN).

[0005] In recent years, Transformer networks have seen rapid development in the field of hyperspectral image band selection. Early research mainly embedded self-attention mechanisms as auxiliary modules into CNN architectures, using quantized channel attention weights to assess band importance. With the introduction of pure attention mechanisms such as SpectralFormer, researchers began to construct end-to-end band selection frameworks. For example, BS-TransNet establishes a global association between bands and scene semantics through cross-attention layers, significantly improving the modeling ability of long-range spectral dependencies. In the optimization phase, LightBandFormer reduced computational complexity through a band grouping strategy, while PhysFormer introduced an atmospheric transport model as a physical constraint for the first time, demonstrating the effectiveness of fusing prior knowledge with attention mechanisms. Current research is gradually shifting towards a multimodal collaborative paradigm. For example, BandGPT achieves cross-sensor band inference through generative pre-training, but its generalization ability in complex terrain scenes is still limited by the scarcity of labeled data and insufficient modeling of physical laws. At the same time, most existing methods only consider single band features, neglecting the issue of the equal importance of spatial and spectral features, as well as local and global features. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a hyperspectral image band selection method suitable for complex terrain scenes.

[0007] Firstly, a method for selecting hyperspectral image bands in complex terrain scenes is provided, including:

[0008] Step 1: Input the original three-dimensional hyperspectral image X∈R with height H, width W, and number of bands B. H×W×B Local spatial features and local spectral features are extracted by a heterogeneous dual-channel 3D convolutional network that integrates cross-attention mechanism;

[0009] Step 2: Add spatial location coding and spectral location coding to the local spatial features and local spectral features respectively;

[0010] Step 3: Input the encoded features into the dual-channel Transformer module to generate band attention weights a∈R B×1 The dual-channel Transformer module includes a spatial feature extraction module and a spectral feature extraction module.

[0011] Step 4: Based on the redundancy between bands and the band attention weights of the multi-head attention in the Transformer module, construct a band evaluation standard, sort the bands, and select the required k bands.

[0012] Preferably, in step 1, the heterogeneous dual-channel 3D convolutional network consists of three parts: a spatial feature path, a spectral feature path, and a cross-modal interaction module. Its overall processing flow is as follows:

[0013] F spa ,F spe =H(X)

[0014] F fusion =C a (F spa ,F spe )

[0015] Where F spa F represents the output of the spatial feature path. spa H(·) represents the output of the spectral feature path, H(·) represents heterogeneous dual-path feature extraction, and F represents the output of the spectral feature path. fusion Representing the interaction between spatial and spectral paths, C a (·) represents the cross-modal attention interaction module.

[0016] Preferably, in step 1, the spatial feature path uses a local spectral-spatial joint convolution kernel to extract texture features, as shown below:

[0017]

[0018] in For 3D convolution operations, BN(·) is the normalization module, and ReLU(·) is the non-linear activation function;

[0019] The spectral feature path enhances the discriminative power of the spectrum through full-band convolution and channel attention, as shown below:

[0020]

[0021] in Using K spa ∈R C×1×1 Full-band convolution kernel, G atten (·) represents channel attention gating.

[0022] Preferably, the cross-modal interaction module is used to introduce a cross-attention mechanism after dual-path feature extraction to establish a dynamic correlation between spatial and spectral features, as follows:

[0023] Q spa =F spa W q K spe =F spe W k V spe =F spe W v

[0024]

[0025] F fusion =LayerNorm((F spa +CrossAtten(Q spa K spe V spe ))

[0026] Among them W q W k W v ∈R d×d Let be the learnable parameter matrix, d be the feature dimension, CrossAtten(·) denotes the cross-attention module, Softmax(·) be the normalization function, and F be the normalization function. fusion This is the output of the spatial feature extraction branch after the interaction of spatial and spectral features.

[0027] Preferably, step 3 includes:

[0028] Step 3.1: Input the local spatial features with added location coding information into the Transformer-based spatial feature extraction module, and extract the global spatial dependencies through a self-attention mechanism; input the local spectral features with added location coding information into the Transformer-based spectral feature extraction module, and extract cross-band spectral correlations through a cross-attention mechanism.

[0029] Step 3.2: After cross-fusion of the outputs from the spatial feature extraction module and the spectral feature extraction module, the result is downsampled through a pooling layer to generate band attention weights a∈R. B×1 , is represented as:

[0030]

[0031] Where Y = Y spa +Y spe Y spa Y represents the output of the spatial feature extraction module. spa This represents the output of the spectral feature extraction module. These are the trainable parameters of the TBFE module, T TBFE This indicates the output of the entire module.

[0032] Preferably, in step 3.1, the dual-channel Transformer module analyzes the representativeness and redundancy of the bands while extracting global features.

[0033] Preferably, in step 4, the candidate band x is averaged across all samples. t Attention weights are used to calculate candidate band x t The representative measure is expressed as:

[0034]

[0035] Where p(x) t ) is x t The representativeness measure, where n is the number of all training samples, p(x) t The larger the value of ), the stronger x is. t It is more representative of the original HSI.

[0036] In a second aspect, a hyperspectral image band selection system suitable for complex terrain scenes is provided for performing any of the methods described in the first aspect, including:

[0037] The first input module is used to input the original hyperspectral image and extract local spatial features and local spectral features through a heterogeneous dual-channel three-dimensional convolutional network that integrates cross-attention mechanism.

[0038] The encoding module is used to add spatial location encoding and spectral location encoding to the local spatial features and local spectral features, respectively;

[0039] The second input module is used to input the encoded features into the dual-channel Transformer module to generate band attention weights a∈R. B×1 The dual-channel Transformer module includes a spatial feature extraction module and a spectral feature extraction module.

[0040] The module is used to construct band evaluation criteria based on the redundancy between bands attended by the multi-head attention in the Transformer module and the band attention weights, and to select the required k bands after sorting the bands.

[0041] Thirdly, a computer storage medium is provided, wherein a computer program is stored therein; when the computer program is run on a computer, the computer causes the computer to perform any of the methods described in the first aspect.

[0042] Fourthly, an electronic device is provided, comprising:

[0043] Memory, used to store computer programs;

[0044] A processor for executing the computer program to implement the method as described in any of the first aspects.

[0045] The beneficial effects of this invention are as follows: First, after inputting the original hyperspectral image, this invention utilizes a heterogeneous dual-channel 3D convolutional network with a cross-attention mechanism to extract local spatial and spectral features. To address the permutation invariance problem of Transformer, positional encoding is added to both channels, injecting prior spatial and spectral positional knowledge into the model, facilitating the model's differentiation of pixels and bands at different locations. To extract global features from the image, a dual-channel Transformer feature extraction network is constructed. The spatial feature extraction branch models the global spatial dependencies between pixels through a self-attention mechanism, deeply mining the spatial context information of each band and evaluating the consistency and stability of its spatial representation. The spectral feature extraction branch uses a cross-attention mechanism to fuse spatial features from different bands, establishing cross-band spectral associations, capturing highly discriminative spectral feature combinations and their complementarity. The outputs of the two branches are cross-fused and downsampled through a pooling layer to generate a band attention weight matrix. Finally, a band selection network is constructed, and a band evaluation criterion is designed based on the correlation between bands and the information richness of the bands to select the required k bands. In summary, the method proposed in this invention fully considers the global-local and spatial-spectral characteristics between bands, providing reliable support for subsequent applications. Therefore, the method proposed in this invention has significant practical application value. Attached Figure Description

[0046] Figure 1 This is a flowchart of the hyperspectral image band selection method for complex terrain scenes provided by the present invention;

[0047] Figure 2 This is a schematic diagram of the structure of a heterogeneous dual-channel three-dimensional convolutional network;

[0048] Figure 3 This is a schematic diagram of the structure of a dual-channel Transformer feature extraction network;

[0049] Figure 4 This is a schematic diagram of the hyperspectral data of Hangzhou Bay provided by the present invention;

[0050] Figure 5 This is a schematic diagram of the hyperspectral data of Yancheng provided by the present invention. Detailed Implementation

[0051] The present invention will be further described below with reference to embodiments. The description of the embodiments below is only for the purpose of helping to understand the present invention. It should be noted that those skilled in the art can make several modifications to the present invention without departing from the principle of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

[0052] Example 1:

[0053] To address the problems of existing technologies, Embodiment 1 of this application provides a method for selecting hyperspectral image bands in complex terrain scenes. This method fully considers the global-local and spatial-spectral characteristics of bands, selecting the most representative subset of bands. It is applicable to hyperspectral image band selection in complex terrain scenes and has strong practicality.

[0054] Specifically, such as Figure 1 As shown, the method includes:

[0055] Step 1: Input the original three-dimensional hyperspectral image X∈R with height H, width W, and number of bands B. H×W×B Local spatial and spectral features are extracted by a heterogeneous dual-channel 3D convolutional network that incorporates a cross-attention mechanism.

[0056] The purpose of this step is to extract local features from hyperspectral images while performing preliminary processing to reduce the workload of subsequent networks.

[0057] In step 1, the heterogeneous dual-channel 3D convolutional network consists of three parts: a spatial feature path, a spectral feature path, and a cross-modal interaction module. It aims to extract local texture information from hyperspectral data while focusing on the relationships between bands. The overall processing flow is shown below:

[0058] F spa ,F spe =H(X)

[0059] F fusion =C a (F spa ,F spe )

[0060] Where F spe F represents the output of the spatial feature path. spe H(·) represents the output of the spectral feature path, H(·) represents heterogeneous dual-path feature extraction, and F represents the output of the spectral feature path. fusion Representing the interaction between spatial and spectral paths, C a (·) represents the cross-modal attention interaction module.

[0061] In step 1, the spatial feature path uses a local spectral-spatial joint convolution kernel to extract texture features, as shown below:

[0062]

[0063] in This is a 3D convolution operation; the kernel parameter is K. spa ∈R 9×3×3 In this system, a large-stride kernel (stride=2) is used to achieve spectral downsampling in the spectral dimension, and a 3×3 kernel is used to capture local neighborhood features in the spatial dimension. BN(·) is a normalization module, and ReLU(·) is a nonlinear activation function.

[0064] The spectral feature path enhances the discriminative power of the spectrum through full-band convolution and channel attention, as shown below:

[0065]

[0066] in Using K spa ∈R C×1×1 Full-band convolution kernel, G atten (·) represents channel attention gating:

[0067] G (atten) (F)=σ(MLP(GAP(F)))⊙F

[0068] The cross-modal interaction module is used to introduce a cross-attention mechanism after dual-path feature extraction to establish a dynamic correlation between spatial and spectral features, as shown below:

[0069] Q spa =F spa W q K spe =F spe W k V spe =F spe W v

[0070]

[0071] F fusion =LayerNormF spa +CrossAtten(Q spa K spe V spe ))

[0072] Among them W q W k W v ∈R d×d Let be the learnable parameter matrix, and d be the feature dimension. Similarly, a reverse attention path from spectral to spatial can be established to form a bidirectional feature interaction. CrossAtten(·) represents the cross-attention module, Softmax(·) is the normalization function, and F...fusion This is the output of the spatial feature extraction branch after the interaction of spatial and spectral features.

[0073] Step 2: Before inputting the dual-channel Transformer module, add spatial location encoding and spectral location encoding to the local spatial features and local spectral features respectively, so as to facilitate the model's understanding of the sequence order.

[0074] Specifically, spatial branching is represented as follows:

[0075] Assign two-dimensional coordinates to each pixel location:

[0076] h∈[0,H-1]

[0077] w∈[0,W-1]

[0078] Where H represents the height of the image and W represents the width of the image.

[0079] Then calculate the row and column position encoding vectors respectively:

[0080]

[0081] Where D is the location encoding dimension. Next, the row and column codes are added together to obtain the final spatial location code:

[0082] E spa (h,w)=E row (h)+E col (w)

[0083] Finally, the positional encoding is superimposed pixel by pixel onto the spatial feature map to obtain the encoded features:

[0084] F′ spa =F spa +E spa (h,w)

[0085] in E is a local tensor extracted through 3D convolution. spa (h,w) is the two-dimensional position code of the spatial coordinates (h,w).

[0086] The spectral branches are first numbered sequentially for the C bands:

[0087]

[0088] Then perform sine and cosine encoding:

[0089]

[0090] Where D is the spectral feature dimension. Next, the position codes are superimposed band by band onto the spectral feature sequence:

[0091]

[0092] Among them, global spectral tensor For spectral path input, Band number One-dimensional positional encoding.

[0093] The purpose of this step is to solve the permutation invariance problem of Transformer, inject spatial and spectral location prior knowledge into the model, so that the model can distinguish pixels and bands at different locations.

[0094] Step 3: Input the encoded features into the dual-channel Transformer module to generate band attention weights a∈R B ×1 The dual-channel Transformer module includes a spatial feature extraction module (Spatial-TBFE) and a spectral feature extraction module (Spectral-TBFE).

[0095] The purpose of this step is to utilize the collaborative work of the two branches to jointly construct a deep and joint feature representation of hyperspectral data, providing a discriminative basis for subsequent accurate assessment of band importance and band selection.

[0096] Step 4: Based on the redundancy between bands and the band attention weights of the multi-head attention in the Transformer module, construct a band evaluation standard, sort the bands, and select the required k bands.

[0097] The purpose of this step is to simultaneously consider the informativeness and redundancy of the bands and select the most representative subset of bands.

[0098] Example 2:

[0099] Based on Example 1, Example 2 of this application provides a more specific method for selecting hyperspectral image bands for complex terrain scenes, including:

[0100] Step 1: Input the original hyperspectral image and extract local spatial features and local spectral features through a heterogeneous dual-channel 3D convolutional network with a cross-attention mechanism.

[0101] Step 2: Add spatial location coding and spectral location coding to the local spatial features and local spectral features, respectively.

[0102] Step 3: Input the encoded features into the dual-channel Transformer module to generate band attention weights a∈R B ×1The dual-channel Transformer module includes a spatial feature extraction module and a spectral feature extraction module.

[0103] Step 3 includes:

[0104] Step 3.1: Input the local spatial features with added location encoding information into the Transformer-based spatial feature extraction module. The spatial feature extraction branch extracts global spatial dependencies through a self-attention mechanism, deeply mines the spatial context information of each band, and evaluates the consistency and stability of its spatial representation. Input the local spectral features with added location encoding information into the Transformer-based spectral feature extraction module. The spectral feature extraction branch uses a cross-attention mechanism to fuse spatial features from different bands, establishes cross-band spectral correlations, captures highly discriminative spectral feature combinations and their complementarity, and then cross-fuses the outputs of the two branches and downsamples them through a pooling layer to generate band attention weights a∈R. B×1 .

[0105] In step 3.1, the dual-channel Transformer module extracts global features while analyzing the representativeness and redundancy of the bands.

[0106] Step 3.2: After cross-fusion of the outputs from the spatial feature extraction module and the spectral feature extraction module, the result is downsampled through a pooling layer to generate band attention weights a∈R. B×1 , is represented as:

[0107]

[0108] Where Y = Y spa +Y spe Y spa Y represents the output of the spatial feature extraction module. spe This represents the output of the spectral feature extraction module. These are the trainable parameters of the TBFE module, T TBFE This indicates the output of the entire module.

[0109] Specifically, this module consists of N identical feature extraction modules. Each branch mainly comprises a multi-head self-attention layer, a normalization layer, and a multilayer perceptron (MLP) layer. All branches use residual connections to stabilize feature propagation. For the query (Q), key (K), and value (V) matrices of the input sequence, the multi-head attention first generates multiple independent query, key, and value weight matrices through linear projection. Each set of weights corresponds to an attention head. Each head independently computes scaled dot product attention, focusing on the correlation between different positions in the sequence, and outputs a weighted value vector. The outputs of all heads are merged through a concatenation operation, and then subjected to a linear transformation (W). O The final result is obtained. This process can be formally represented as:

[0110] MutiHead(Q,K,V)=Concat(head1,…,head h W O

[0111] The calculation of the i-th head

[0112] After feature extraction via multi-head attention, layer normalization is used to stabilize the feature distribution and calibrate it to avoid gradient explosion. Next, an MLP is used to enhance the model's expressive power and improve its sensitivity to subtle spectral differences. Finally, the output is normalized again, and the outputs of the two branches are fused and downsampled through a pooling layer to generate band attention weights a∈R. B×1 .

[0113] Step 4: Based on the redundancy between bands and the band attention weights of the multi-head attention in the Transformer module, construct a band evaluation standard, sort the bands, and select the required k bands.

[0114] In step 4, the candidate band x is averaged over all samples. t The attention weights of (the t-th band in X) are used to calculate the candidate band x. t The representative measure is expressed as:

[0115]

[0116] Where p(x) t ) is x t The representativeness measure, where n is the number of all training samples, p(x) t The larger the value of ), the stronger x is. t It is more representative of the original HSI.

[0117] To account for inter-band correlation, a band correlation matrix is ​​used to measure inter-band redundancy. First, the inter-band correlation matrix is ​​calculated independently for each head:

[0118]

[0119] Where, d k This represents the projection dimensions of Q and K, used for scaling the gradient.

[0120] Next, the score matrix of each head is softmax normalized to obtain the correlation matrix A. (i) :

[0121]

[0122] in This represents the dependence of band j on band k at the i-th head.

[0123] Then, the correlation matrices of the h heads are averaged:

[0124]

[0125] Since redundancy is mutual, symmetry processing is then performed:

[0126]

[0127] Where C jk =C kj C jk ∈[0,1] represents the redundancy score of band j and band k. The larger the value, the higher the redundancy between them.

[0128] Therefore, when searching for the (q+1)th band, the candidate band x t The redundancy matrix is ​​calculated as follows:

[0129]

[0130] Where Q represents the selected band subset, C t,k Indicates candidate band x t With band x k The correlation matrix between them.

[0131] Finally, in order to select a subset of bands with high information richness and low band correlation, a new band evaluation criterion is proposed:

[0132] s(x t )=p(x t )-ω×d(x t )

[0133] Where ω is the balance coefficient.

[0134] When using a search-based method to find the (q+1)th selected band, the goal is to find the highest-scoring band in the candidate band subset as the selected band.

[0135] x best =max[s(x t )]

[0136] The proposed method was compared with three other band selection methods and the classification results using a self-constructed hyperspectral dataset (images from the ZY1-02D ​​satellite). Specific results are shown in Tables 1-6. The results show that on the Hangzhou Bay dataset, the proposed method achieves an overall accuracy (OA) of 91.537%, an average accuracy (AA) of 88.273%, and a Kappa coefficient of 0.88488; on the Yancheng dataset, the proposed method achieves an overall accuracy (OA) of 96.671%, an average accuracy (AA) of 94.003%, and a Kappa coefficient of 0.95247. These results demonstrate that on these two hyperspectral datasets, the proposed method significantly outperforms other comparative methods, even surpassing the full-band method, indicating that the selected band subset has strong representativeness and robustness, further validating its effectiveness in complex remote sensing image classification tasks.

[0137] Table 1. OA of Hangzhou Bay under different band selection methods

[0138]

[0139] Table 2 AA of Hangzhou Bay under different band selection methods

[0140]

[0141] Table 3 Kappa values ​​for Hangzhou Bay under different band selection methods

[0142]

[0143]

[0144] Table 4. OA of Yancheng under different band selection methods

[0145]

[0146] Table 5. AA of Yancheng under different band selection methods

[0147]

[0148] Table 6. Kappa values ​​for Yancheng under different band selection methods

[0149]

[0150] It should be noted that the parts in this embodiment that are the same as or similar to those in Embodiment 1 can be referred to each other, and will not be repeated in this application.

[0151] Example 3:

[0152] Based on Example 2, Example 3 of this application provides a hyperspectral image band selection system suitable for complex terrain scenes, including:

[0153] The first input module is used to input the original hyperspectral image and extract local spatial features and local spectral features through a heterogeneous dual-channel three-dimensional convolutional network that integrates cross-attention mechanism.

[0154] The encoding module is used to add spatial location encoding and spectral location encoding to the local spatial features and local spectral features, respectively;

[0155] The second input module is used to input the encoded features into the dual-channel Transformer module to generate band attention weights a∈R. B×1 The dual-channel Transformer module includes a spatial feature extraction module and a spectral feature extraction module.

[0156] The module is used to construct band evaluation criteria based on the redundancy between bands attended by the multi-head attention in the Transformer module and the band attention weights, and to select the required k bands after sorting the bands.

[0157] It should be noted that the system provided in this embodiment is the corresponding system of the method provided in embodiment 2. Therefore, the parts that are the same as or similar to those in embodiment 2 in this embodiment can be referred to each other, and will not be described again in this application.

Claims

1. A method for selecting hyperspectral image bands suitable for complex terrain scenes, characterized in that, include: Step 1: Input height as Width The number of bands is 3D raw hyperspectral image Local spatial features and local spectral features are extracted by a heterogeneous dual-channel 3D convolutional network that integrates cross-attention mechanism; In step 1, the heterogeneous dual-channel 3D convolutional network consists of three parts: spatial feature path, spectral feature path, and cross-modal interaction module. Its overall processing flow is as follows: in The output represents the spatial feature path. The output represents the spectral characteristic path. This indicates heterogeneous dual-path feature extraction. This represents the interaction between spatial and spectral paths. For cross-modal interaction modules; In step 1, the spatial feature path uses a local spectral-spatial joint convolution kernel to extract texture features, as shown below: in This is a 3D convolution operation. For the normalization module, It is a non-linear activation function; The spectral feature path enhances the discriminative power of the spectrum through full-band convolution and channel attention, as shown below: in use Full-band convolution kernel, Channel attention gating; In step 1, the cross-modal interaction module is used to introduce a cross-attention mechanism after dual-path feature extraction to establish a dynamic correlation between spatial and spectral features, as shown below: in For learnable parameter matrix, For feature dimension, This indicates a cross-attention module. The output of the spatial feature extraction branch after the interaction of spatial and spectral features; Step 2: Add spatial location coding and spectral location coding to the local spatial features and local spectral features respectively; Step 3: Input the encoded features into the dual-channel Transformer module to generate band attention weights. The dual-channel Transformer module includes a spatial feature extraction module and a spectral feature extraction module. Step 4: Based on the redundancy between bands attended to by multi-head attention in the Transformer module and the band attention weights, construct a band evaluation criterion, sort the bands, and select the required ones. Each band.

2. The hyperspectral image band selection method applicable to complex terrain scenes according to claim 1, characterized in that, Step 3 includes: Step 3.1: Input the local spatial features with added location coding information into the Transformer-based spatial feature extraction module, and extract the global spatial dependencies through a self-attention mechanism; input the local spectral features with added location coding information into the Transformer-based spectral feature extraction module, and extract cross-band spectral correlations through a cross-attention mechanism; Step 3.2: After cross-fusion of the outputs from the spatial feature extraction module and the spectral feature extraction module, the outputs are downsampled through a pooling layer to generate band attention weights. , is represented as: in , This represents the output of the spatial feature extraction module. This represents the output of the spectral feature extraction module. These are the trainable parameters of the TBFE module. This indicates the output of the entire module.

3. The hyperspectral image band selection method applicable to complex terrain scenes according to claim 2, characterized in that, In step 3.1, the dual-channel Transformer module extracts global features while analyzing the representativeness and redundancy of the bands.

4. The hyperspectral image band selection method for complex terrain scenes according to claim 3, characterized in that, In step 4, candidate bands are averaged across all samples. Attention weights are used to calculate candidate bands The representative measure is expressed as: in yes Representative measure It is the number of all training samples. The larger the value, the more it indicates It is more representative of the original HSI.

5. A hyperspectral image band selection system suitable for complex terrain scenes, characterized in that, For performing the method according to any one of claims 1 to 4, comprising: The first input module is used to input the original hyperspectral image and extract local spatial features and local spectral features through a heterogeneous dual-channel three-dimensional convolutional network that integrates cross-attention mechanism. The encoding module is used to add spatial location encoding and spectral location encoding to the local spatial features and local spectral features, respectively; The second input module is used to input the encoded features into the dual-channel Transformer module to generate band attention weights. The dual-channel Transformer module includes a spatial feature extraction module and a spectral feature extraction module. The module is used to construct band evaluation criteria based on the redundancy between bands attended to by the multi-head attention in the Transformer module and the band attention weights, and then sorts the bands to select the required ones. Each band.

6. A computer storage medium, characterized in that, The computer storage medium stores a computer program; when the computer program is run on the computer, it causes the computer to perform the method described in any one of claims 1 to 4.

7. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the method as described in any one of claims 1 to 4.

Images