Hybrid convolution-Kolmogorov-Arnod network model and classification method

Abstract

The invention discloses a hybrid convolution-Kolmogov-Arnod network model and a classification method, and relates to the technical field of image classification, the model comprises a preprocessing module, a multi-scale feature extraction module, a spatial feature extraction module, two groups of KAN branches, a fusion module and a global average pooling layer; the preprocessing module is used for extracting a preliminary feature map; the multi-scale feature extraction module is used for extracting a multi-scale feature map in the preliminary feature map; kAN branches: the initial special graph sequentially passes through two groups of KAN branches to obtain global feature representation; the spatial feature extraction module is used for strengthening key features of the multi-scale feature map to obtain a spatial feature map; the fusion module is used for fusing the global feature representation and the spatial feature map to obtain fusion features; the global average pooling layer is used for mapping and compressing the fusion features into classification features; and the full connection layer outputs a classification result according to the classification features. According to the method, global information and local information can be fully utilized, and high-precision classification is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of image classification, and more particularly to a hybrid convolution-Kolmogorov-Arnold network model and a classification method. Background Art

[0002] Image classification (IC) refers to the use of computers to automatically classify images, enabling them to automatically identify the categories of objects within an image. Hyperspectral image classification (HIC) is a crucial task in remote sensing, providing critical support for environmental monitoring, agricultural management, and urban planning. Traditional hyperspectral image classification methods rely primarily on machine learning algorithms such as support vector machines (SVMs) and principal component analysis (PCA), but these methods have limitations when dealing with the high dimensionality and spectral redundancy of hyperspectral data.

[0003] In recent years, deep learning techniques, especially convolutional neural networks (CNNs), have made significant progress in hyperspectral image classification. CNN-based HIC models include: 2D-CNNs are primarily used to extract two-dimensional spatial features; 3D-CNNs consider both spatial and spectral dimensions, enabling the joint modeling of spatial neighborhoods and spectral correlations; multi-scale CNNs capture spatial features at different scales by using convolution kernels of different scales; they are suitable for processing hyperspectral images with multi-layered structures and can better capture both fine-grained and coarse-grained features; deep CNNs extract higher-level features by increasing the depth of the network, excelling in processing complex hyperspectral image data; CNNs combined with attention mechanisms, such as SENet, dynamically adjust feature importance by introducing an attention mechanism, enhancing the model's ability to perceive key features; CNNs combined with generative adversarial networks (GANs), generate high-quality features through GANs, enhancing model robustness; and CNNs combined with Transformers, which fuse local and global features, improve the utilization of location information, and optimize computational efficiency and model capacity, achieving superior performance in tasks such as image processing.

[0004] However, the prior art still has the following problems:

[0005] (1) Insufficient utilization of global and local information: Most existing hyperspectral image classification methods rely on CNN to extract local spatial-spectral features, or rely on Transformer to model global dependencies, but few methods can fully utilize both global and local information at the same time.

[0006] (2) Insufficient fusion of global and local information: Even though some methods try to combine CNN and Transformer, the fusion method is usually relatively simple, such as simple splicing or weighted averaging, which cannot dynamically align and refine complementary features.

[0007] Therefore, how to make full use of the global and local information in hyperspectral images and improve classification accuracy is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0008] In view of this, the present invention provides a hybrid convolution-Kolmogorov-Arnold network model and classification method, which can simultaneously utilize global information and local information to improve the classification performance of the model in complex hyperspectral datasets, solve the problems of insufficient utilization and insufficient fusion of global and local information in existing models, and achieve higher-precision classification and positioning.

[0009] In order to achieve the above object, the present invention adopts the following technical solutions:

[0010] A hybrid convolutional-Kolmogorov-Arnold network model, including a preprocessing module, a multi-scale feature extraction module, a spatial feature extraction module, two sets of KAN branches, a fusion module and a global average pooling layer;

[0011] The preprocessing module extracts preliminary feature maps from the acquired hyperspectral images;

[0012] The multi-scale feature extraction module extracts hierarchical spatial features under multiple receptive fields in the preliminary feature map to obtain a multi-scale feature map;

[0013] KAN branch, the preliminary feature map passes through two groups of KAN branches in sequence to obtain the global feature representation;

[0014] The spatial feature extraction module strengthens the key features of the multi-scale feature map to obtain the spatial feature map;

[0015] Fusion module, which fuses the global feature representation and the spatial feature map to obtain the fusion feature;

[0016] The global average pooling layer applies an average operation to each feature map of the fused feature to compress the fused feature map into a single-valued classification feature;

[0017] The fully connected layer performs linear transformation operations based on the classification features and outputs the classification results.

[0018] Preferably, the preprocessing module includes a principal component analysis layer and a CreateCube layer; the principal component analysis layer performs principal component analysis on the hyperspectral image and extracts spectral features; the CreateCube layer maps the spectral features into a cube space to obtain a preliminary feature map in the form of a cube.

[0019] Preferably, the multi-scale feature extraction module includes two groups of multi-scale convolutional layers, a 2DConv layer, and a normalization layer connected in sequence; each group of multi-scale convolutional layers includes a multi-scale three-dimensional convolution block (Multi-scale 3DConvBlock) and a maximum pooling layer connected in sequence; the normalization layer includes a normalization operation and an activation function ReLU; the multi-scale three-dimensional convolution block mainly extracts feature maps of different scales through different convolution kernels; the maximum pooling layer mainly reduces the size of the extracted feature map; the 2DConv layer mainly adjusts the channel dimension; and the normalization layer mainly stabilizes training to prevent gradient explosion. Each group of multi-scale three-dimensional convolutional blocks performs multi-scale convolution and then performs maximum pooling layer pooling. After two multi-scale convolutions, the number of channels is adjusted through the 2DConv layer, and the result is output through the normalization layer.

[0020] Preferably, the multi-scale 3D convolution block includes four parallel 3D convolution branches, and a splicing layer and a 2DConv layer sequentially connected to the 3D convolution branches; the 3D convolution branch includes a 3D convolution layer and a normalization layer, and the normalization layer includes a normalization operation and an activation function ReLU; the convolution kernels of the four 3D convolution branches are 3×3×1, 3×3×3, 3×3×5, and 3×3×7, respectively; the expression of the multi-scale 3D convolution block is:

[0021] f ms =2DConv 3×3 (Concat[f 3×3×1 (X), f 3×3×3 (X), f 3×3×5 (X), f 3×3×7 (X)])

[0022] Among them, f(X) represents the convolution operation, and the convolution kernel sizes are 3×3×1, 3×3×3, 3×3×5 and 3×3×7 respectively; f ms Represents the extracted feature map; Concat represents the splicing operation; 2DConv 3×3 Represents a two-dimensional convolution operation with a convolution kernel size of 3×3.

[0023] Preferably, the spatial feature extraction module includes two sets of stacked spatial extraction blocks, each of which includes a 2DConv layer and a spatial attention layer (SpatialAttention) connected in sequence; the spatial attention layer merges multi-scale features through serial average pooling and maximum pooling operations to ensure that global and local feature representations are retained, and then generates spatial attention weights through convolution operations and activation functions; the expression of the spatial attention layer feature extraction module is:

[0024]

[0025] Among them, M s Represents the spatial attention weight matrix; σ represents the activation function of the spatial attention layer, which is used to adjust the attention mechanism and is the sigmoid activation function; [;] represents the concatenation operation; C represents the number of channels; MaxPool represents the maximum pooling operation; AvgPool represents the average pooling operation; Conv2d represents the two-dimensional convolution operation; the input feature map is refined by element multiplication according to the spatial attention weight matrix, F represents the convolution feature map output by the 2DConv layer, and F output by the second set of spatial attention layers out Represents a spatial feature map.

[0026] Preferably, the KAN branch includes a basis function calculation layer, a spline calculation layer, an adaptive grid layer and a fusion layer; the basis function calculation layer applies nonlinear SiLU activation to the preliminary feature map, and then performs a linear transformation to obtain a basis function tensor; the adaptive grid layer uses an adaptive grid mechanism to adjust the grid points according to the distribution of the preliminary feature map to obtain an adaptive network; the spline calculation layer uses the B-spline basis function to perform adaptive approximation according to the adaptive network to obtain a spline tensor; the fusion layer fuses the basis function tensor and the spline tensor to obtain an output tensor; the preliminary feature map passes through two groups of KAN branches in sequence to obtain a global feature representation, that is, the output tensor of the second group of KAN branches is used as the global feature representation; the expression of the KAN branch is:

[0027] Y=f base (X)+f spline (X);

[0028]

[0029] G adaptive =∈G uniform +(1-∈)G data

[0030] Where Y represents the output tensor; represents the input tensor to the KAN branch; f base represents the basis function tensor; f spline represents the spline tensor; φ represents the activation function SiLU; Contains learnable weights, d out and d in Represent the dimensions of the output tensor and the input tensor respectively; represents the flattened B-spline basis function evaluated at each input value of the input tensor, g represents the grid size of the adaptive grid, and s represents the spline order; represents the learnable coefficient of each B-spline basis function; B i，1 (x) represents the i-th interval after the input domain of the first layer network is segmented; B i，k(x) represents the i-th interval after the input domain of the k-th layer network is divided, which is used to represent the interval index in the network layer index and distinguish different input subintervals in different network layers; x represents a point in the definition domain, which is used to calculate the value of the B-spline function; {t i} represents the grid points for dividing the input domain, which are usually initialized uniformly between [-1,1]. i} Divide the input range; G adaptive Represents the adapted grid; G uniform Represents a uniform grid, expressing uniformly distributed grid points and combined function rules. Uniformly distributed grid points are pre-defined equally spaced segmentation points on the input domain (such as control points in B-spline). Combined function rules are connected by mathematical functions (such as B-spline basis functions, linear interpolation, etc.) to form a processing rule for input information. Uniform grid allows the network to process input features uniformly on the input domain according to these preset segmentation and function relationships; G data represents the data distribution of the target dataset for training the model; ∈ represents the control adaptation rate, which is usually 0.02.

[0031] Preferably, the fusion module includes a flattening layer, a splicing layer and a weight fusion layer; the flattening layer flattens the spatial feature map to be consistent with the dimension of the global feature representation; the flattened spatial feature map and the global feature representation are spliced ​​along the feature dimension to obtain splicing features; the weight fusion layer adjusts the weights of the splicing features to obtain fusion features; wherein,

[0032] The expression of the flattened layer is:

[0033]

[0034] Among them, B represents the batch size, and each branch produces 256-dimensional features; F CNN Represents the spatial feature map; F KAN represents the global feature representation; Flatten represents the flattening operation; F′ CNN represents the flattened spatial feature map; F′ KAN Represents the global feature representation after the flattening layer;

[0035] The expression of the splicing layer is:

[0036]

[0037] Among them, || represents the splicing operation; F concat Represents splicing features;

[0038] The expression of the weight fusion layer is:

[0039] Y logirs =W class ·Fconcat +b class

[0040] in, and Represent the weight and bias of the fully connected layer respectively, C represents the number of categories; Y logits Indicates fusion features.

[0041] A classification method based on a hybrid convolutional-Kolmogorov-Arnold network model comprises the following steps:

[0042] Collect hyperspectral image samples and construct a training sample set;

[0043] Construct a hybrid convolutional-Kolmogorov-Arnold network model, and use the training sample set to train and optimize the hybrid convolutional-Kolmogorov-Arnold network model to obtain a classification model;

[0044] The hyperspectral image to be detected is collected and input into the classification model to obtain the classification results.

[0045] Through the above technical solution, it can be seen that compared with the prior art, the present invention discloses a hybrid convolution-Kolmogorov-Arnold network model and classification method, which combines the local feature extraction capability of CNN with the high-dimensional nonlinear modeling capability of Kolmogorov-Arnold network (KAN), and simultaneously utilizes global information and local information in the model, so as to more comprehensively understand the global structure and local anomalies of hyperspectral data. At the same time, an efficient feature fusion module, namely a multi-scale feature extraction module, is designed, and the normalization operation of the normalization layer of the multi-scale three-dimensional convolution block is used to realize the correlation and weight adjustment within and between features, thereby effectively integrating global and local feature information from different scales, making full use of global information and local information, improving the classification performance of the model in complex hyperspectral data sets, and realizing high-precision classification and positioning of hyperspectral images. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0047] Figure 1 A schematic diagram of the structure of the hybrid convolutional-Kolmogorov-Arnold network model provided by the present invention;

[0048] Figure 2 A schematic diagram of the structure of the multi-scale three-dimensional convolution block provided by the present invention;

[0049] Figure 3 Schematic diagram of the structure of the spatial attention layer provided by the present invention. DETAILED DESCRIPTION

[0050] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0051] The embodiment of the present invention discloses a hybrid convolution-Kolmogorov-Arnold network model, the network architecture is as follows Figure 1 As shown, it includes a preprocessing module, a multi-scale feature extraction module, a spatial feature extraction module, two groups of KAN branches, a fusion module and a global average pooling layer;

[0052] The preprocessing module extracts preliminary feature maps from the acquired hyperspectral images;

[0053] The multi-scale feature extraction module extracts hierarchical spatial features under multiple receptive fields in the preliminary feature map to obtain a multi-scale feature map;

[0054] KAN branch, the preliminary feature map passes through two groups of KAN branches in sequence to obtain the global feature representation;

[0055] The spatial feature extraction module strengthens the key features of the multi-scale feature map to obtain the spatial feature map;

[0056] Fusion module, which fuses the global feature representation and the spatial feature map to obtain the fusion feature;

[0057] The global average pooling layer applies an average operation to each feature map of the fused feature to compress the fused feature map into a single-valued classification feature;

[0058] The fully connected layer performs linear transformation operations based on the classification features and outputs the classification results.

[0059] The hybrid convolution-Kolmogorov-Arnold network model mainly consists of three core modules: multi-scale feature extraction module (MsFE), spatial feature extraction module (SaFE) and KAN branch (KANLinear Layer); MsFE extracts hierarchical spatial features under multiple receptive fields through parallel 3×3×1, 3×3×3, 3×3×5 and 3×3×7 three-dimensional convolution kernel sets, effectively capturing cross-scale information from local texture to global structure; the KAN branch uses learnable B-spline basis functions to explicitly model high-order nonlinear interactions in the spectral dimension, accurately fitting complex spectral relationships, but lacks spatial awareness; SaFE adopts large kernel convolution (7×7) and channel attention mechanism to enhance the discriminative ability of regional feature representation. In the feature fusion stage, the dual-branch output first retains the original feature information through parallel connection, and then dynamically adjusts the feature weights through the spatial-spectral attention mechanism in the fusion module. Finally, feature compression and classification decisions are achieved through the global average pooling layer (GAP) and the fully connected layer. Specifically, the input image first passes through the principal component analysis PCA and CreateCube module of the preprocessing module to generate a preliminary feature map. Then, one branch is MsFE and SpFE receiving the feature map from the CreateCube module, and after convolution and pooling operations, multi-scale feature maps and spatial feature maps are generated. The other branch is KANLinear Layer generating global feature representation through linear transformation and adaptive grid. The two branches are fused through the weighted sum operation through the Fusion Layer, that is, the fusion layer. Finally, the final classification result is generated through the global average pooling layer (GAP) and the fully connected layer.

[0060] Furthermore, MsEF includes a multi-scale 3D convolution block (Multi-scale 3DConvBlock), a maximum pooling layer, a multi-scale 3D convolution block (Multi-scale 3DConv Block), a maximum pooling layer, a 2DConv layer and a normalization layer connected in sequence; the normalization layer includes a normalization operation and an activation function ReLU. The structure is as follows Figure 1 (b) shown.

[0061] Multi-scale 3DConv Block consists of four parallel 3D convolution branches, such as Figure 2 As shown in the figure, 3×3×1, 3×3×3, 3×3×5 and 3×3×7 convolution kernels are used to effectively capture features of different scales. The process is as follows:

[0062] f ms =2DConv3×3 (Concat[f 3×3×1 (X), f 3×3×3 (X), f 3×3×5 (X), f 3×3×7 (X)])

[0063] Among them, f(X) represents the convolution operation, and the convolution kernel sizes are 3×3×1, 3×3×3, 3×3×5 and 3×3×7 respectively; f ms Represents the extracted feature map; Concat represents the splicing operation; 2DConv 3×3 Represents a two-dimensional convolution operation with a convolution kernel size of 3×3.

[0064] Furthermore, two groups of KAN branches form a dual-channel architecture, and two consecutive KAN branches are used to model nonlinear and global features. The process of each group of KAN branches processing data is as follows: Figure 1 As shown in (d), it includes basis function calculation layer, spline calculation layer, adaptive grid layer and fusion layer, which combines traditional neural network operation with adaptive B-spline function approximation to extract powerful spectral spatial features from hyperspectral data; KANLinear Layer processes the input tensor To generate the output tensor The process is as follows:

[0065] Y=f base (X)+f spline (X)

[0066] where f base and f spline They correspond to the basis function tensor of the basis function calculation layer (Base Path) and the spline tensor of the spline calculation layer (Spline Path);

[0067] The basis function calculation layer applies nonlinear SiLU activation and then performs a linear transformation:

[0068]

[0069] Where φ represents the SiLU activation function, Contains learnable weights, d out and d in Represent the dimensions of the output tensor and the input tensor respectively;

[0070] The spline calculation layer uses B-spline basis functions for adaptive approximation:

[0071]

[0072] in represents the flattened B-spline basis function evaluated at each input value, g is the grid size of the adaptive grid, s is the spline order, Contains the learnable coefficients of each basis function; the B-spline basis function is calculated recursively according to the Cox-de Boor formula:

[0073]

[0074] where {t i} is the grid point for dividing the input domain, usually initialized uniformly between [-1,1]. i} Divide the input range; B i，1 (x) represents the i-th interval after the input domain of the first layer network is segmented; B i，k (x) represents the i-th interval after the input domain of the k-th layer network is divided, which is used to represent the interval index in the network layer index and distinguish different input subintervals in different network layers; x represents the point where the B-spline function value is to be calculated;

[0075] The adaptive grid layer uses an adaptive grid mechanism to adjust the grid points according to the distribution of the input data:

[0076] G adaptive =∈G uniform +(1-∈)G data

[0077] Among them G adaptive Represents the adapted grid; G uniform Represents a uniform grid, expressing uniformly distributed grid points and combined function rules. Uniformly distributed grid points are pre-defined equally spaced segmentation points on the input domain (such as control points in B-spline). Combined function rules are connected by mathematical functions (such as B-spline basis functions, linear interpolation, etc.) to form a processing rule for input information. Uniform grid allows the network to process input features uniformly on the input domain according to these preset segmentation and function relationships; G data The data distribution of the target dataset from which the model is trained, ∈ represents the control adaptation rate, which is usually set to 0.02.

[0078] Furthermore, SaFE includes two sets of stacked spatial extraction blocks, which include 2DConv layers and spatial attention layers connected in sequence, such as Figure 1 As shown in (c), the spatial attention layer aggregates channel information through serial average pooling and maximum pooling operations to ensure that global and local feature representations are retained, and then through convolution operations; the spatial attention layer structure is as follows Figure 3 As shown, the calculation steps are as follows:

[0079]

[0080] Among them, M s Represents the spatial attention weight matrix; σ represents the activation function of the spatial attention layer, which is used to adjust the attention mechanism and is the sigmoid activation function; [;] represents the concatenation operation; C represents the number of channels; MaxPool represents the maximum pooling operation; AvgPool represents the average pooling operation; Conv2d represents the two-dimensional convolution operation; the input feature map is refined by element multiplication according to the spatial attention weight matrix, F represents the convolution feature map output by the 2DConv layer, and F output by the second set of spatial attention layers out Represents a spatial feature map.

[0081] Furthermore, the fusion module performs fusion through a weighted sum operation. The specific process is as follows:

[0082] The features extracted by the spatial feature model are the output of CNN. The flattening layer flattens the output of the spatial feature extraction module to ensure that the dimension is consistent with the KAN branch. The calculation process is as follows:

[0083]

[0084] Among them, B represents the batch size, each branch produces 256-dimensional features; Flatten represents the flattening operation;

[0085] The concatenation layer concatenates the features from the two paths along the feature dimension, expressed as:

[0086]

[0087] The features are classified through the weight fusion layer:

[0088] Y logirs =W class ·F concat +b class

[0089] in, and They represent the weight and bias of the fully connected layer respectively, and C is the number of categories.

[0090] On the other hand, a classification method based on a hybrid convolutional-Kolmogorov-Arnold network model comprises the following steps:

[0091] Collect hyperspectral image samples and construct a training sample set;

[0092] Construct a hybrid convolutional-Kolmogorov-Arnold network model, and use the training sample set to train and optimize the hybrid convolutional-Kolmogorov-Arnold network model to obtain a classification model;

[0093] The hyperspectral image to be detected is collected and input into the classification model to obtain the classification results.

[0094] In a specific embodiment, the training model uses datasets including Indian Pines, Pavia University, WHU_Hi_Longkou, and Xuzhou. WHU_Hi_Longkou is a collection of hyperspectral images from a university research group, and Xuzhou is a hyperspectral image of a suburban area captured by a HYSPEX camera. The model is loaded and trained on an NVIDIA GeForce RTX4090 GPU, with a batch size of 64 and a learning rate of 1e. -4 , and conduct 100 rounds of training to make the model achieve the best effect.

[0095] During the training process, the model input mainly includes hyperspectral image data and training labels, while the output includes classification results and classification accuracy. Three evaluation indicators are used to evaluate the model performance, including Overall Accuracy (OA), Average Accuracy (AA) and Kappa Coefficient; the calculation formulas are:

[0096]

[0097] Among them, TP i is the true proportion of the i-th category, k is the number of categories, and N is the total number of samples;

[0098]

[0099] Among them, Accuracy i Represents the accuracy of the i-th category (the number of correctly classified i-th category samples divided by the total number of true i-th category samples N i );

[0100]

[0101] Among them, OA represents overall accuracy, P e is the expected consistency of random assortment, N i，pred Indicates the number of samples predicted to be class i, N i，true Indicates the number of samples that are actually of the i-th class.

[0102] The three indicators were compared with the existing technology to verify that the model classification method of the present invention achieved the best results in all indicators. The existing technology includes 2DCNN, 3DCNN, A2S2KResNet, MCM_CNN, SSFTT, SpectralFormer, and morphformer. The model of the present invention is represented by HybridCKAN. The results of the comparison between the existing technology and the model of the present invention in the IP dataset, PU dataset, XZ dataset, and LK dataset are shown in Tables 1-4 below:

[0103] Table 1 Classification accuracy and index evaluation results of different models in IP dataset

[0104]

[0105] Table 2 Classification accuracy and index evaluation results of different models in PU dataset

[0106]

[0107] Table 3 Classification accuracy and index evaluation results of different models in the XZ dataset

[0108]

[0109] Table 4 Classification accuracy and index evaluation results of different models in the LK dataset

[0110] TABLE VIII:COMPARISON WITH SOTA NETWORKS FOR LK DATASET

[0111]

[0112] In the table, Class represents the number of categories in each dataset. The first row represents the model name. The row corresponding to each class represents the classification accuracy of this category. OA, AA, and Kappa respectively correspond to the overall classification accuracy, the average classification accuracy, and the consistency between the predicted results and the actual results in the evaluation classification task.

[0113] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0114] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A hybrid convolutional-Kolmogorov-Arnold network model, characterized in that: It includes a preprocessing module, a multi-scale feature extraction module, a spatial feature extraction module, two groups of KAN branches, a fusion module and a global average pooling layer; The preprocessing module extracts preliminary feature maps from the acquired hyperspectral images; The multi-scale feature extraction module extracts hierarchical spatial features under multiple receptive fields in the preliminary feature map to obtain a multi-scale feature map; KAN branch, the preliminary feature map passes through two groups of KAN branches in sequence to obtain the global feature representation; The spatial feature extraction module strengthens the key features of the multi-scale feature map to obtain the spatial feature map; Fusion module, which fuses the global feature representation and the spatial feature map to obtain the fusion feature; Global average pooling layer compresses the fused feature map into classification features; The fully connected layer outputs the classification results based on the classification features.

2. The hybrid convolutional-Kolmogorov-Arnold network model according to claim 1, wherein The preprocessing module includes a principal component analysis layer and a CreateCube layer; the principal component analysis layer performs principal component analysis on the hyperspectral image and extracts spectral features; the CreateCube layer maps the spectral features into a cube space to obtain a preliminary feature map in the form of a cube.

3. The hybrid convolutional-Kolmogorov-Arnold network model according to claim 1, wherein The multi-scale feature extraction module includes two sets of multi-scale convolutional layers, a 2DConv layer, and a normalization layer connected in sequence. Each set of multi-scale convolutional layers includes a multi-scale 3D convolution block and a maximum pooling layer connected in sequence. The normalization layer includes a normalization operation and an activation function ReLU. The multi-scale 3D convolution block extracts feature maps of different scales using different convolution kernels. The maximum pooling layer reduces the size of the extracted feature map. The 2DConv layer adjusts the channel dimension. Normalization layers stabilize training and prevent gradient explosion.

4. The hybrid convolutional-Kolmogorov-Arnold network model according to claim 3, wherein: The multi-scale 3D convolution block includes four parallel 3D convolution branches, as well as a splicing layer and a 2DConv layer connected to the 3D convolution branches in sequence; the 3D convolution branch includes a 3D convolution layer and a normalization layer, and the normalization layer includes a normalization operation and an activation function ReLU; the convolution kernels of the four 3D convolution branches are 3×3×1, 3×3×3, 3×3×5 and 3×3×7 respectively.

5. The hybrid convolutional-Kolmogorov-Arnold network model according to claim 1, wherein The spatial feature extraction module consists of two sets of stacked spatial extraction blocks. The spatial extraction blocks include a 2DConv layer and a spatial attention layer connected in sequence. The spatial attention layer performs average pooling, maximum pooling, convolution operation, activation function activation and element multiplication operation in sequence.

6. The hybrid convolutional-Kolmogorov-Arnold network model according to claim 1, wherein: The KAN branch includes a basis function calculation layer, a spline calculation layer, an adaptive grid layer, and a fusion layer. The basis function calculation layer applies nonlinear SiLU activation to the preliminary feature map, and then performs a linear transformation to obtain the basis function tensor. The adaptive grid layer uses an adaptive grid mechanism to adjust the grid points according to the distribution of the preliminary feature map to obtain an adaptive network; The spline calculation layer uses the B-spline basis function to perform adaptive approximation according to the adaptive network to obtain the spline tensor; the fusion layer fuses the basis function tensor and the spline tensor to obtain the output tensor; the preliminary feature map passes through two groups of KAN branches in sequence to obtain the global feature representation, that is, the output tensor of the second group of KAN branches is used as the global feature representation.

7. The hybrid convolutional-Kolmogorov-Arnold network model according to claim 1, wherein: The fusion module includes a flattening layer, a splicing layer, and a weighted fusion layer; the flattening layer flattens the spatial feature map into the same dimension as the global feature representation; the flattened spatial feature map and the global feature representation are spliced ​​along the feature dimension to obtain the splicing feature; The weight fusion layer adjusts the weights of the spliced ​​features to obtain fused features.

8. A classification method, characterized in that Applying the hybrid convolutional-Kolmogorov-Arnold network model according to any one of claims 1 to 7 comprises the following steps: Collect hyperspectral image samples and construct a training sample set; Construct a hybrid convolutional-Kolmogorov-Arnold network model, and use the training sample set to train and optimize the hybrid convolutional-Kolmogorov-Arnold network model to obtain a classification model; The hyperspectral image to be detected is collected and input into the classification model to obtain the classification results.

Images