Hyperspectral image classification method and device and electronic equipment

Abstract

The invention provides a hyperspectral image classification method and device and electronic equipment, and the method comprises the steps: carrying out the capturing of image information of a hyperspectral image through a central difference convolution algorithm, and generating a first-order feature map; performing graph structure modeling on features in the first-order feature graph based on spectral similarity to obtain a structural feature graph; determining multi-dimensional features of the structural feature map; performing dot product operation and normalized index operation on the multi-dimensional features to obtain the importance of each feature under each dimension and the data hierarchical representation of the importance of the corresponding feature; performing nonlinear processing of a KAN layer on the structural feature map; processing the nonlinear processing result, the importance degree of each feature and the data hierarchical representation by using a multi-dimensional attention mechanism to obtain a processing result containing the global feature, the local detail feature and the feature importance degree of the hyperspectral image; and classifying the hyperspectral image based on the processing result. According to the method, the hyperspectral images can be classified efficiently and accurately.

Description

Technical Field

[0001] The present invention belongs to the technical field of image classification, and particularly relates to a hyperspectral image classification method, device and electronic device. Background Art

[0002] Hyperspectral images (HSIs) can accurately depict the spectral information of targets by acquiring the reflection or radiation characteristics of targets in hundreds of continuous spectral bands, and have become an important data source in the field of remote sensing. Different from traditional RGB imaging that can only capture information in three bands of red, green, and blue, HSIs have significant advantages in spectral resolution, which can enhance the recognition and classification ability of ground objects, thus effectively distinguishing objects with similar appearances but different spectral characteristics. Therefore, classifying using HSIs has stronger discrimination ability due to the rich information in more frequency bands. Nowadays, HSI classification has been widely applied in fields such as environmental protection, mineral classification, urban ecology, land use, precision agriculture, and medical analysis, providing important support for precise analysis and efficient decision-making.

[0003] HSIs collect high-resolution data in hundreds of narrow spectral bands, can record the fine spectral information of hundreds or even thousands of continuous bands, making each pixel contain rich spectral characteristics, far beyond the perception ability of the human eye. HSIs integrate rich spatial and spectral information to form a complex three-dimensional data structure, thus having significant advantages in ground object recognition and classification, and being able to accurately distinguish targets with similar appearances but different spectral characteristics. However, due to the high dimensionality and spectral complexity of HSI data, the computational complexity is high, and it is vulnerable to noise and the curse of dimensionality. It is difficult for traditional methods to efficiently extract features.

[0004] The emergence of deep learning has brought new possibilities for solving these problems. Although its development time in the field of remote sensing is still short, deep learning has significantly promoted the progress of hyperspectral data processing technology. Especially with the support of automatic feature extraction and attention mechanisms, it can effectively address the high-dimensional features and information redundancy problems of hyperspectral data, and solve the problem that traditional methods are difficult to capture high-level semantic information.

[0005] However, when using deep learning methods for HSI classification, the following three challenges still need to be addressed at the present stage:

[0006] 1) Effective modeling of long-term dependencies: When dealing with long-term cross-band dependencies in HSI data, the model is prone to losing key information and unable to fully capture the subtle connections between different bands, so that key features cannot be fully emphasized. 2) Low utilization rate of spectral information exacerbates sample dependence: Existing models are insufficient in extracting and utilizing spectral information in HSI data, resulting in limited expressive ability for spectral features. The low data utilization rate causes the model to strengthen its dependence on the number of training samples. 3) High-dimensional complexity of data leads to easy entrapment in local optima: Because HSI has multi-band spectral information, the high-dimensional and complex nature of its data makes the model prone to getting trapped in local optima during training, affecting classification performance. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a hyperspectral image classification method, device and electronic device that can classify hyperspectral images efficiently and accurately.

[0008] An embodiment of the present invention provides a hyperspectral image classification method, including:

[0009] Capturing image information of the obtained hyperspectral image through a central difference convolution algorithm, and generating a first-order feature map based on the captured image information;

[0010] Performing feature association on the features in the first-order feature map based on spectral similarity, and performing graph structure modeling based on the association result to obtain a structure feature map corresponding to the hyperspectral image;

[0011] Performing multi-dimensional feature extraction on the structure feature map to obtain multi-dimensional features;

[0012] Performing dot product operation and normalized exponential operation on the multi-dimensional features to obtain the importance of each feature in each dimension and the data-level representation corresponding to the feature importance;

[0013] Performing non-linear processing on the structure feature map for KAN layers to obtain a non-linear processing result;

[0014] Using a multi-dimensional attention mechanism to perform parallel processing on the non-linear processing result, the importance of each feature, and the data-level representation to obtain a processing result including the global feature, local detail feature and feature importance of the hyperspectral image;

[0015] Classifying the hyperspectral image based on the processing result.

[0016] In one embodiment, the capturing image information of the obtained hyperspectral image through a central difference convolution algorithm includes:

[0017] The obtained hyperspectral images are subjected to feature extraction at different scales through three standard convolutional models and three central difference convolutional models at three different scales, and spectral features and texture features at different scales are obtained.

[0018] In one embodiment, the method further includes:

[0019] Reshape the features in the first-order feature map into a two-dimensional node set, where each node in the two-dimensional node set corresponds to a pixel position and its feature vector in the first-order feature map;

[0020] Perform feature correlation on the features in the first-order feature map in combination with the two-dimensional node set.

[0021] In one embodiment, the feature correlation of the features in the first-order feature map based on spectral similarity includes:

[0022] Determine the spectral distance between a first node in the two-dimensional node set and second nodes around it;

[0023] Determine the second nodes corresponding to each of the spectral distances that meet the distance requirement;

[0024] Determine the second nodes with spectral distances meeting the requirement as the neighbor nodes of the first node;

[0025] Repeat the above steps to traverse each node in the two-dimensional node set to obtain the neighbor nodes of each node;

[0026] Construct a neighborhood matrix based on the neighbor nodes of each node, and the neighbor matrix records the connection relationship between nodes in the graph structure corresponding to the hyperspectral image.

[0027] In one embodiment, the multi-dimensional feature extraction of the structural feature map to obtain multi-dimensional features includes:

[0028] Adopt a three-axis attention mechanism to perform multi-channel feature extraction on the structural feature map in the horizontal, vertical, and spatial directions to obtain horizontal features, vertical features, and spatial features.

[0029] In one embodiment, the dot product operation and normalized exponential operation on the multi-dimensional features to obtain the importance of each feature in each dimension and the data-level representation of the corresponding feature importance include:

[0030] Construct a weight matrix, where each row of the weight matrix corresponds to a weight vector of a hyperspectral image category, and the weight vector is used to map the multi-dimensional features into probabilities of different categories;

[0031] Perform a dot product operation on the multi-dimensional features and the weight matrix to obtain the importance of each feature in each dimension;

[0032] Perform a normalization exponential operation on each dimension feature in the multi-dimensional features in the corresponding dimension to obtain local features with high importance in each dimension, and perform hierarchical modeling in combination with the local features.

[0033] In one embodiment, the non-linear processing of the structural feature map in the KAN layer to obtain a non-linear processing result includes:

[0034] Perform a normalization process on the structural feature map;

[0035] Calculate the spline basis function values of each feature in the structural feature map by using a target spline of a specified order through a spline basis function for the normalized structural feature map;

[0036] Perform non-linear processing on the structural feature map based on the spline basis function values of each feature to obtain a non-linear processing result.

[0037] In one embodiment, the use of a multi-dimensional attention mechanism to perform parallel processing on the non-linear processing result, the importance of each feature, and the data hierarchical representation to obtain a processing result including the global features, local detail features, and feature importance of the hyperspectral image includes:

[0038] Process the non-linear processing result and the importance of each feature based on the gaze attention mechanism to obtain the scanning information and spectral global information of the hyperspectral image;

[0039] Process the importance of each feature and the data hierarchical representation based on the attention mechanism and a scaling factor to obtain the focus information and local detail information of the spectrum in the hyperspectral image.

[0040] Another embodiment of the present invention also provides a hyperspectral image classification device, including:

[0041] A capture module for capturing image information of the obtained hyperspectral image through a central difference convolution algorithm and generating a first-order feature map based on the captured image information;

[0042] A modeling module for performing feature association on the features in the first-order feature map based on spectral similarity and performing graph structure modeling based on the association result to obtain a structural feature map corresponding to the hyperspectral image;

[0043] An extraction module for performing multi-dimensional feature extraction on the structural feature map to obtain multi-dimensional features;

[0044] An operation module, configured to perform dot product operation and normalized exponential operation on the multi-dimensional features to obtain the importance of each feature under each dimension and the data level representation corresponding to the feature importance;

[0045] A first processing module, configured to perform non-linear processing on the structural feature map for KAN layers to obtain a non-linear processing result;

[0046] A second processing module, configured to perform parallel processing on the non-linear processing result, the importance of each feature, and the data level representation by using a multi-dimensional attention mechanism to obtain a processing result including the global features, local detail features, and feature importance of the hyperspectral image;

[0047] A classification module, configured to classify the hyperspectral image based on the processing result.

[0048] Another embodiment of the present invention further provides an electronic device, including:

[0049] One or more processors;

[0050] A memory, configured to store one or more programs;

[0051] When the one or more programs are executed by the one or more processors, the one or more processors are caused to implement the hyperspectral image classification method as described in any one of the above.

[0052] The beneficial effects of the present invention include reducing information loss across bands through long-term modeling based on feature interaction and being able to focus on the expression of key features of ground objects. In addition, when extracting features, the detailed texture of ground objects in the image and the feature information in the spectrum are fully mined and efficiently utilized, thereby solving the problems of insufficient dependence modeling, difficult data acquisition, and the model training being trapped in a local optimal solution in the hyperspectral image classification task, and finally significantly improving the accuracy of hyperspectral image classification.

[0053] Other features and advantages of the present application will be described in the following specification, and some of them will become obvious from the specification or be understood by implementing the present application. The objectives and other advantages of the present application can be achieved and obtained through the structures specifically pointed out in the written specification, claims, and drawings.

[0054] The technical solutions of the present application will be further described in detail below through the drawings and embodiments. Description of the Drawings

[0055] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the accompanying drawings required for the description of the specific embodiments or the prior art. Obviously, the accompanying drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0056] Figure 1 It is a schematic flowchart of the hyperspectral image classification method in an embodiment of the present invention.

[0057] Figure 2 It is a schematic flowchart of the hyperspectral image classification method in another embodiment of the present invention.

[0058] Figure 3 It is a schematic flowchart of the hyperspectral image classification method in yet another embodiment of the present invention.

[0059] Figure 4 It is a structural block diagram of the hyperspectral image classification device in an embodiment of the present invention. Specific Embodiments

[0060] Next, specific embodiments of the present invention will be described in detail with reference to the accompanying drawings, but it is not a limitation of the present invention.

[0061] It should be understood that various modifications can be made to the embodiments disclosed herein. Therefore, the following description should not be regarded as a limitation, but only as an example of the embodiments. Those skilled in the art will think of other modifications within the scope of the present disclosure.

[0062] The accompanying drawings included in the specification and constituting a part of the specification illustrate the embodiments of the present disclosure, and together with the general description of the present disclosure given above and the detailed description of the embodiments given below, are used to explain the principles of the present disclosure.

[0063] By the following description of the preferred forms of the embodiments given as non-limiting examples with reference to the accompanying drawings, these and other features of the present invention will become apparent.

[0064] It should also be understood that although the present invention has been described with reference to some specific examples, those skilled in the art can surely implement many other equivalent forms of the present invention, which have the features as described in the claims and thus are all within the protection scope defined thereby.

[0065] When combined with the accompanying drawings, in view of the following detailed description, the above and other aspects, features and advantages of the present disclosure will become more apparent.

[0066] Specific embodiments of the present disclosure will be described hereinafter with reference to the accompanying drawings; however, it should be understood that the disclosed embodiments are merely examples of the present disclosure and can be implemented in various ways. Well-known and / or repetitive functions and structures are not described in detail to avoid obscuring the present disclosure with unnecessary or redundant details. Therefore, the specific structural and functional details disclosed herein are not intended to be limiting, but are merely used as a basis for the claims and a representative basis for teaching those skilled in the art to use the present disclosure in substantially any suitable detailed structure in a variety of ways.

[0067] This specification may use the phrases "in one embodiment", "in another embodiment", "in yet another embodiment", or "in other embodiments", all of which may refer to one or more of the same or different embodiments according to the present disclosure.

[0068] Next, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

[0069] As Figure 1 shown, the present invention includes a hyperspectral image classification method, comprising:

[0070] S1: Capturing image information of the obtained hyperspectral image through a central difference convolution algorithm, and generating a first-order feature map based on the captured image information;

[0071] S2: Associating features in the first-order feature map based on spectral similarity, and performing graph structure modeling based on the association result to obtain a structure feature map corresponding to the hyperspectral image;

[0072] S3: Extracting multi-dimensional features from the structure feature map to obtain multi-dimensional features;

[0073] S4: Performing dot product operation and normalized exponential operation on the multi-dimensional features to obtain the importance of each feature in each dimension and the data-level representation of the corresponding feature importance;

[0074] S5: Performing non-linear processing on the structure feature map for KAN layers to obtain a non-linear processing result;

[0075] S6: Using a multi-dimensional attention mechanism to perform parallel processing on the non-linear processing result, the importance of each feature, and the data-level representation to obtain a processing result including the global features, local detail features, and feature importance of the hyperspectral image;

[0076] S7: Classifying the hyperspectral image based on the processing result.

[0077] The execution subject of the method in this embodiment can directly be the processor of the electronic device, or a constructed network model capable of executing the method. The method in this embodiment reduces information loss across bands through long-term modeling based on feature interaction and can focus on the expression of key features of ground objects. Additionally, when performing feature extraction, it fully mines and efficiently utilizes the feature information of ground object detail textures and spectra in the image, thereby solving the problems of insufficient dependence modeling in hyperspectral image classification tasks, difficult data acquisition, and the model training being trapped in local optimal solutions, and finally significantly improving the accuracy in hyperspectral image classification.

[0078] Taking the execution of the method in this embodiment by the network model as an example, the network model includes three parts in total, namely the Feature Attention Mechanism Network (Glance and Gaze, abbreviated as GGA), the Feature Extraction Module (abbreviated as STM), and the Heuristic Optimization Algorithm (abbreviated as BGO). The method of this embodiment first processes the hyperspectral image by the feature extraction module. The STM in the feature extraction module will fully utilize the feature information of different scales in the hyperspectral image (HSI), strengthen the capture of detail and edge information through central difference convolution, capture more detailed and sufficient features of ground objects in the HSI, and then use the neighborhood matrix based on spectral similarity for graph structure modeling to improve the utilization rate of spectral information in the HSI. Then the feature data processed by the STM is input into the Feature Attention Mechanism Network (GGA). The GGA performs non-linear transformation on the input feature map through KAN, conducts feature cross-fusion using Tri-Axis Attention, and then uses the cooperation of Glance-attention and Gaze-attention for effective long-term modeling, so that the key information of remote cross-bands is retained, while enhancing the full expression of useful features and suppressing the expression of redundant information. Considering the high-dimensional complexity of HSI data, it is easy to be trapped in local optimal solutions during the model training stage. Therefore, the BGO better explores better solutions while ensuring the stable convergence of the network model by introducing the bud growth mechanism. The BGO is used during the network model training stage and will not be used anymore when the model is trained and put into normal use. If the method of this embodiment is directly executed by the processor, the above-mentioned process of using BGO for model convergence is not involved.

[0079] In one embodiment, the capturing of image information for the obtained hyperspectral image through the central difference convolution algorithm includes:

[0080] S8: Extract features of different scales from the obtained hyperspectral image through three standard convolution models of different scales and three central difference convolution models to obtain spectral features and texture features of different scales.

[0081] For example, the significant differences in ground objects in HSI (hyperspectral images) are mainly reflected in their texture and structural features. Therefore, in the case of limited data volume, how to fully capture useful information becomes an important and challenging task. When using CNN for classification, the image is usually divided into local regions through convolution operations, and spatial features are extracted using these local blocks. In HSI classification, however, each pixel is classified. Therefore, feature extraction based on CNN usually divides the image into fixed-size region blocks due to the single size of the convolution kernel. To capture the ground object features at different scales in HSI and make full use of spectral information to solve adjacent ground objects with similar spatial features, in this embodiment, multi-scale convolution with central difference convolution is used to construct a neighborhood matrix based on spectral information correlation for graph structure modeling. Establishing spatial long-range dependence relationships using spectral information avoids the problem of misclassification of adjacent ground objects with similar spatial features and further enhances the utilization rate of spectral features. For the input image, multi-scale parallel convolution is first used to extract feature information at different scales in the image, as shown in the formula:

[0082] Y TSM = y N [concatenate[y N1 , y N2 , y N3 , y C4 y C5 , y C6

[0083] Among them, y N1,2,3 respectively represent the standard convolution layers at different scales from the first to the third, while y C4,5,6 correspond to the central difference convolution layers of the fourth to the sixth layers respectively, as shown in the formula:

[0084]

[0085] R is represented as the local receptive field in the input feature map x, y represents the result output feature map. l0 represents the corresponding position on the input and output feature maps, and l n traverses the positions included in R. w is the weight of the convolution kernel. The hyperparameter θ controls the influence of the central position l0 on the overall output. By adjusting the value of θ, the weight of the central pixel relative to the surrounding pixels can be increased or decreased. As its value increases, the centrality of the central difference convolution operation increases, and when the value of θ is 0, it is equivalent to ordinary convolution, as shown in the above formula. Among them, the optional θ = 0.2 is the optimal hyperparameter value.

[0086] ​Compared with using only CNN for feature capture, central difference convolution is more conducive to accurately capturing subtle and important structural features in HSI such as spectral details and texture changes, thus emphasizing the structural information of edges and details, further highlighting the differential information in terms of detail features, thereby improving the adaptability to complex environments, avoiding misclassification caused by similar features at the junctions of different ground objects, and at the same time enriching the feature expression and avoiding low utilization rate of spectral information. In addition, by combining multi-scale ordinary convolution and central difference convolution in a parallel configuration, it can process feature information at different scales, enrich the feature expression, and fuse intensity and gradient information, enabling the network model to intuitively establish a harmonious balance between the extraction of global and local features during the training phase. Even when not executed using a model but only based on a processor, the method flow is the same and the above technical effects are also achieved.

[0087] Further, as Figure 2 shown, the method further includes:

[0088] S9: Reshape the features in the first-order feature map into a two-dimensional node set, where each node in the two-dimensional node set corresponds to a pixel position and its feature vector in the first-order feature map;

[0089] S10: Perform feature association on the features in the first-order feature map by combining the two-dimensional node set.

[0090] Among them, the feature association of the features in the first-order feature map based on spectral similarity includes:

[0091] S11: Determine the spectral distance between the first node and the second nodes around it in the two-dimensional node set;

[0092] S12: Determine the second nodes corresponding to each of the spectral distances that meet the distance requirement;

[0093] S13: Determine the second nodes with spectral distances meeting the requirement as the neighbor nodes of the first node;

[0094] S14: Repeat the above steps to traverse each node in the two-dimensional node set to obtain the neighbor nodes of each node;

[0095] S15: Construct a neighborhood matrix based on the neighbor nodes of each node, and the neighbor matrix records the connection relationship between nodes in the graph structure corresponding to the hyperspectral image.

[0096] Exemplarily, HSI contains the reflectance information of each pixel in multiple spectral bands. Generally, adjacent ground objects on the image are similar in spatial features such as edge shapes, while the adjacent high-dimensional spectral information on pixel points can well distinguish the differential features of different ground objects. To make full use of the spectral information at different scales, we innovatively combine multi-scale convolution with graph structure modeling. Specifically, based on the feature maps generated by parallel multi-scale convolution, we construct a neighborhood matrix based on spectral similarity and use it as the input of graph convolution for graph structure modeling. By sorting the features, we distinguish the ground objects with different spectral features to enhance the discrimination ability of adjacent ground objects. Specifically, first reshape the fused features into a two-dimensional node set, where each node corresponds to a pixel position and its feature vector in the feature map, as shown in the following formula:

[0097]

[0098] X = reshape(F, (N, C))

[0099] where, denote the fused feature map as F ∈ R H×W×C , reshape the feature map into the node feature matrix X ∈ R NxC , H is the height of the feature map, W is the width of the feature map, C is the number of feature channels, and N = H × W represents the total number of nodes. Specifically, each node i of the vector x i corresponds to a pixel position (h i , w i ) in the feature map.

[0100] For any two nodes i and j, construct a neighborhood matrix based on the spectral similarity between the node feature vectors.

[0101]

[0102] where, calculate the spectral distance d i，j between nodes i and j, and for each node i, select the top k nodes with the smallest spectral distance from it as its neighbors, and record the indices of the neighbor nodes as N(i). Then, according to the selected neighbor nodes, construct the neighborhood matrix A ∈ {0, 1} N×N , where, if A i，j = 1, it means that node j is a neighbor of node i because of its similarity to node i. In this way, the matrix A records the connection relationship between the nodes in the graph structure.

[0103] Compared with simple local convolution, graph structure modeling can make the similar spectral features on pixels with a relatively large spatial distance affect each other in the feature space, capture the long-distance spectral correlation, and improve the spectral utilization rate. Graph convolution depends on the neighborhood matrix A i,jDefine the relationships between nodes to achieve feature aggregation, thereby enhancing the richness of feature expression, as shown in the following formula:

[0104] H (l+1) = σ(AH (l) W (l) )

[0105]

[0106] where H (l) is the feature matrix of the l-th layer, W (l) is the learnable weight matrix, σ is the activation function, and O represents the graph-related output matrix. The neighborhood matrix is constructed based on spectral similarity, and sparse connection relationships are achieved by screening the most similar nodes, making feature propagation more efficient, capturing the spectral correlation between distant nodes, and enhancing the interaction of similar features. In this embodiment, elu is used as the activation function of the GCN to retain negative value information and nonlinear mapping, which not only enriches the diversity of feature expression but also alleviates the problem of gradient disappearance and improves the stability of model training.

[0107] In another embodiment, the multi-dimensional feature extraction of the structural feature map to obtain multi-dimensional features includes:

[0108] S16: Adopt a three-axis attention mechanism to perform multi-channel feature extraction on the structural feature map in the horizontal, vertical, and spatial directions to obtain horizontal features, vertical features, and spatial features.

[0109] Further, as Figure 3 shown, the dot product operation and normalization exponential operation on the multi-dimensional features to obtain the importance of each feature in each dimension and the data-level representation of the corresponding feature importance include:

[0110] S17: Construct a weight matrix, where each row of the weight matrix corresponds to a weight vector of a hyperspectral image category, and the weight vector is used to map the multi-dimensional features into probabilities of different categories;

[0111] S18: Perform a dot product operation on the multi-dimensional features and the weight matrix to obtain the importance of each feature in each dimension;

[0112] S19: Perform a normalization exponential operation on each dimension feature in the multi-dimensional features in the corresponding dimension to obtain local features with high importance in each dimension, and perform hierarchical modeling in combination with the local features.

[0113] Specifically, to address the problem of difficult long-term dependence modeling in HSI, this embodiment proposes a self-attention mechanism GGA. First, Tri-Axis Attention is used to extract multi-channel information from the input features in the horizontal, vertical, and spatial directions to enhance the ability to capture features in different dimensions. Subsequently, Glance-attention is utilized to absorb information from the global level, promote the modeling of large-scale dependence relationships, and capture key features across bands. Although Glance-attention can efficiently capture remote feature expressions, it may ignore the connections and expressions of features between some local regions. To make up for this deficiency, this embodiment designs Gaze-attention, which focuses on valuable local information to compensate for the local relationships ignored between HSI bands and between pixels. By collaborating Glance-attention and Gaze-attention in parallel, GGA can effectively learn the global information of HSI while retaining and fully expressing necessary local details, avoiding the loss of key information and the expression of irrelevant information, thereby achieving efficient modeling of long-term dependence relationships.

[0114] Tri-Axis attention: HSI contains a wide array of spectral channels, in which a considerable amount of redundant information is observed. To improve the ability of the neural network to absorb multi-dimensional features of spectral images, we propose Tri-Axis attention to integrate three different channel attention mechanisms: horizontal, vertical, and spatial. This design helps the model proficiently identify and carefully process the most important relevant features in the input data. The exact procedure is shown in the following formula, where the variables are respectively assigned Q, K, V to encapsulate the horizontal, vertical, and spatial feature information:

[0115]

[0116] Among them, W represents the weight matrix of each Softmax layer. The indices i, j, k correspond to the width, height, and depth dimensions on the horizontal axis, vertical axis, and spatial axis respectively, and p1, p2, p3 represent their respective total ranges. In the first step, the image is decomposed along these three axes to generate a feature map for element-wise multiplication with the weight matrix. Then, the exponent of these multiplication results is used as a normalization factor. Through the sum normalization process, a normalization score is calculated for each component of Q, K, V. This score quantifies the contribution of each input element to the output, thus greatly helping the model to identify the most significant features. This mechanism is crucial for highlighting relevant features and reducing redundant features, thereby more effectively extracting features to improve the accuracy of HSI classification. By applying the Softmax function along different axes of the input tensor, the model separates important features in each dimension (horizontal, vertical, or spectral space), and then uses these features to establish a more complex data hierarchical representation for classification targets.

[0117] In another embodiment, the non-linear processing of the structural feature map by the KAN layer to obtain a non-linear processing result includes:

[0118] S20: Perform a normalization process on the structural feature map;

[0119] S21: Calculate the spline basis function values of each feature in the structural feature map by using a target spline of a specified order through a spline basis function;

[0120] S22: Perform non-linear processing on the structural feature map based on the spline basis function values of each feature to obtain a non-linear processing result.

[0121] Exemplarily, the solution in this embodiment is implemented in combination with a gaze attention mechanism. For example, traditional multi-layer perceptron (MLP) architectures use fixed activation functions on neurons, which makes it difficult to deeply explore complex non-linear relationships in global information when processing high-dimensional hyperspectral image (HSI) feature maps, thus making it difficult to achieve effective long-term dependence modeling. For the Kolmogorov–Arnold network (KAN) proposed based on the Kolmogorov-Arnold representation theorem, compared with the MLP that can only rely on non-linear activation functions, the KAN network has higher learnability in weights. Specifically, instead of using linear weights, KAN replaces each weight parameter with a one-dimensional function parameterized as a spline function. This method not only improves the accuracy of the model but also enhances its interpretability. Inspired by this, in order to effectively express the global information in the HSI feature map, this embodiment performs non-linear processing of the input feature map X through the KAN layer. In this way, the attention mechanism can perform deeper interaction and fusion on HSI data in the feature space, thereby enhancing the model's understanding ability of HSI data and further improving the classification accuracy. First, the input feature X is normalized to obtain X' so that the input feature is normalized within the range:

[0122]

[0123] where represents the input feature matrix, N is the number of samples, F in is the number of input features, X max and X min represent the maximum and minimum values of each feature in the batch size respectively, and ∈ represents an infinitesimal number to avoid the denominator being 0.

[0124] Calculated by using B-splines of a specified order through spline basis functions. For each feature X', the spline basis function S(X) is evaluated at a predetermined node k:

[0125] S i，j (X' i ) = max(0, 1 - [X' i - k j )

[0126] where k j is the j-th node (knots), evenly distributed within the interval, and is the value of the j-th spline basis function of the i-th feature.

[0127] Then, complex linear and non-linear transformations are performed on X to enhance the feature expression ability, as shown in the following formula:

[0128]

[0129] Among them, is the trainable base weight matrix, and σ is the activation function. To retain non-linear features and smoothly propagate gradients, the SiLU activation function is adopted in this embodiment, as shown in the above formula. X′ i is the i-th feature column of X′, and F in refers to the number of output features. is the spline weight matrix corresponding to the i-th input feature, K is the number of nodes in the spline, and F out is the number of output features. is the matrix of spline basis functions evaluated on the input features.

[0130] It should be noted that using B-spline functions of different orders has a certain impact on the expressive power and flexibility of the module. Although low-order splines are computationally simple and reduce the risk of overfitting, they are difficult to capture complex non-linear relationships, resulting in discontinuous jumps of attention weights between different regions. High-order splines, on the other hand, can capture more complex and delicate non-linear relationships, but the complex calculation process increases the possibility of overfitting. In this embodiment, a 3rd-order B-spline function is adopted as the spline function of the KAN layer, enabling the KAN Layer to capture non-linear features fully while smoothing parameter transmission.

[0131] The Glance-attention branch takes the input feature map Y processed by the KAN layer as input, and its other formulas are shown as follows, where "Glance" is defined as the scanning information of the hyperspectral image, and "Global" is defined as the spectral global information:

[0132]

[0133] Glance = concatenate[Y, Global]

[0134] Among them, p4 represents the number of spatial channels. To reduce the number of parameters and suppress noise in the amplitude domain, the Glance-Attention branch initially uses the Softmax function to normalize the feature information in the horizontal, vertical, and spatial dimensions of the Tri-Axis attention. This process helps to scan the image comprehensively and achieve an accurate grasp of the global information. After that, the original input is merged with the self-attention weighted features, retaining the global context essence while maintaining the integrity of spatial correlation. The Glance-attention, designed to absorb global information, proficiently captures the long-distance representation features of the hyperspectral image, making it possible to conduct a comprehensive global review of the hyperspectral image, effectively performing self-attention in a vast receptive field, thereby enhancing the proficiency of the model in identifying long-distance dependencies.

[0135] Further, the parallel processing of the non-linear processing result, the importance of each feature, and the data hierarchical representation by using the multi-dimensional attention mechanism to obtain a processing result including the global feature, local detail feature, and feature importance of the hyperspectral image includes:

[0136] S23: Processing the non-linear processing result and the importance of each feature based on the gaze attention mechanism to obtain the scanning information and spectral global information of the hyperspectral image;

[0137] S24: Processing the importance of each feature and the data hierarchical representation based on the attention mechanism and the scaling factor to obtain the focusing information in the hyperspectral image and the local detail information of the spectrum.

[0138] In this embodiment, for Gaze-attention, although the Glance branch is good at capturing remote representations, it ignores the local connections across partitions and valuable local information crucial for visual tasks that rely on local cues. During the process of processing the high-dimensional spatial data of hyperspectral images, local information contains rich spatial and texture features around each pixel. To make up for the reduction of the ignored local detail information between hyperspectral bands and between pixels, this embodiment proposes Gaze-Attention, as shown in the following formula. "Gaze" represents the focusing information in the hyperspectral image, while "Local" represents the local details of the spectrum:

[0139] Local i，j，k = max(Q i，j，k , K i，j，k , V i，j，k )

[0140]

[0141] As shown in the above formula, a scaling factor can be proposed to enable the model to have scaled dot - product attention. This ensures selective access to the aspects of information necessary for the primary task. The scaling is established by normalizing the feature maps of the Q, K, V vectors and starting through exponentiation to calibrate the focusing intensity of the features. This precaution avoids the dominance of any single attribute due to scalar differences, thus ensuring focused attention on relevant regions. This scaling mechanism is a notable feature in the Gaze - attention architecture and has an important impact on the dot - product calculation between Q, K, V, especially in a high - dimensional context. Without this adjustment, the gradient of Softmax may decrease, weakening effective learning. Therefore, the details of the features processed by Tri - Axis attention in different dimensions are amplified according to the above formula, and then, by introducing a scaling factor, the enhanced results are scaled according to the Gaze formula. This can effectively ensure that the attention scores are maintained within the optimal range, avoid overly large values in the dot - product, and contribute to the efficient scaling of attention metrics. The Gaze - attention branch aims to focus on the most information - rich regions through a scaling factor based on the feature - space dimension, thus compensating for the detailed information that may be lost during the global scanning process of Glance - attention. Finally, this embodiment uses the method in the following formula to assign weights to Kance - Attention and Gaze:

[0142] GFA = concatenate[α * Glance, β * Gaze]

[0143] Through the synergistic effect of the above two attention mechanisms, GGA can effectively integrate feature information in different ranges, realizing long - term dependence modeling based on feature interaction. This method adaptively enhances important spectral and spatial features, retains necessary local details while learning the global information of HSI, thus avoiding the loss of key information to fully express key features and improve classification performance. Especially in the classification of edge information, it can effectively retain and focus on key features.

[0144] For the above BGO module, in the HSI classification task, the complexity of spectral information and the diversity of spatial features significantly affect the performance of the classification network. Selecting appropriate optimizers, such as SGD and Adam, can improve the convergence speed and classification accuracy of the model by optimizing the parameter update process. SGD accelerates learning through random sampling and momentum terms, reducing oscillation phenomena, and thus more effectively searches for the optimal solution in the high-dimensional data space. Adam uses the first-order and second-order exponential moving averages of gradients to achieve fine-grained learning rate adjustment, and normalizes and corrects biases through the EMA of gradient squares, enhancing the adaptability of parameter updates. However, the high-dimensional feature space of HSI data may lead to a high degree of overlap between classes, making gradient descent-based optimization algorithms prone to falling into local optima. In addition, the non-linear features and noise in hyperspectral data may cause unstable gradient signals, affecting the accurate estimation of the gradient direction by adaptive optimizers, and thus affecting the learning effect and generalization ability of the model. Therefore, in the HSI classification task, it is crucial to reasonably select and adjust the optimizer and its parameters. In contrast, RMSProp, as an adaptive learning rate optimization algorithm, dynamically adjusts the learning rate to adapt to the gradient distribution and update requirements of different bands or features by maintaining the moving average square value of each parameter gradient. However, it lacks an effective exploration mechanism and relies on historical gradient information, making it still have a high risk of falling into local optima under highly non-convex loss functions, especially in steep or flat saddle-point regions, thus greatly affecting the final classification performance.

[0145] In contrast, the young stems of plants usually grow towards the light source. However, during this process, the young buds will slightly swing due to external environmental disturbances, thus adjusting their growth direction. Although the changes caused by these external disturbances are completely random, this slight random swing helps plants more accurately perceive the surrounding light intensity and ultimately find the optimal light direction. Inspired by this, in order to simulate the growth trend of young buds based on light source judgment and enhance the exploration ability of growth through external disturbances to optimize the growth direction, this embodiment proposes a young bud growth optimization algorithm based on root mean square propagation (BGO). The BGO algorithm aims to introduce an exploration mechanism into the traditional RMSProp optimizer, enabling it to more effectively jump out of local optima when training the model. Specifically, the gradient can be regarded as an indication guiding the movement of model parameters towards the optimal solution direction. The algorithm dynamically adjusts the parameter update path through Gaussian perturbation by randomly extending the growth of young buds and selecting the dominant direction, simulating the fine-tuning process of plants under light, so as to achieve a more comprehensive search in the high-dimensional complex parameter space, enabling the optimizer to moderately explore new directions during the parameter update process, thereby effectively avoiding local optima to a certain extent and promoting the search for the global optimal solution.

[0146] After the hyperparameters of the initial optimizer, BGO creates corresponding slots for each model parameter and uses the formula: to calculate the partial derivative of the objective loss function j(θ t ) with respect to the model parameters to obtain the gradient g t . The update rule of RMS is shown as follows:

[0147] First, to dynamically adjust the learning rate and reduce the update step of parameters with large gradient changes, the decay rate ρ of RMS is used to control the contribution of the current gradient g t to the RMS accumulation variable R t to store the exponentially weighted moving average of the squared gradients. On this basis, the present embodiment introduces a sprout growth mechanism, as shown in the following formula:

[0148] d ∼ N(μ, σ 2 )

[0149] f d = d · g t

[0150] S t = γ · s t-1 + α · f d

[0151] For the current growth factor S t , the perturbation factor f[[ID=3⑨]] d models the perturbation effect of external interference on the current growth direction of the sprout. Specifically, the perturbation factor f d regulates the influence degree of the previous stage growth factor on the current growth factor through the growth factor decay rate γ, and on this basis, perturbs the current gradient. Among them, the perturbation coefficient α controls the influence magnitude of the perturbation factor on the growth factor, d is the introduced Gaussian perturbation term, μ is the center position of the Gaussian distribution, which determines the average offset of the perturbation, and σ is used to measure the diffusion degree of the Gaussian distribution, that is, the magnitude of the perturbation. All in all, for the growth factor of the current stage, it is jointly affected by the perturbation factor and the past growth factor, and their specific influence degrees are respectively controlled by the growth factor decay rate γ and the perturbation coefficient α in the optimizer parameters. Then, the growth state of the sprout is as shown in the formula: v t = μ · v t-1 + s t , by adding the current growth factor S t to the previous growth state v t-1 , the current growth state v t, this speed is used to represent the smooth update of parameter values during the growth of the tender shoots. The growth coefficient μ controls the contribution of the previous growth state to the current growth state. Since the growth factor is affected by the perturbation factor and has a certain degree of randomness, the update of the growth state also has a certain degree of randomness. This parameter update mechanism that includes randomness can effectively improve the exploration ability of the model, help the model jump out of local minima, and thus improve the global search ability of the optimization process.

[0152] Finally, the complete update rule of the parameters is shown in the following formula:

[0153]

[0154] where ∈ is a small value to prevent division by zero, and the gradient g t is scaled by the RMS accumulation value R t to adaptively adjust the learning rate η of each parameter. Specifically, the learning rate of the model at a larger gradient is smaller to help the model learn better, and when the model learns at a smaller gradient, it will be larger to accelerate the faster convergence of the model. The core mechanism of this algorithm is to introduce a tender shoot growth mechanism with a Gaussian perturbation factor, and explore through the growth state term v t and accelerate the gradient descent.

[0155] The traditional RMSProp optimizer does not have a corresponding mechanism to promote the exploration of the parameter space and is prone to falling into local optimal solutions, thus reducing the generalizability. Therefore, inspired by the fact that plant tender shoots have a certain random perturbation to dynamically adjust the growth direction, this embodiment chooses to introduce a growth rule with increased randomness due to Gaussian perturbation, which is more capable and efficient in exploration. This not only increases the spatiality of the solution, prevents the algorithm from prematurely converging to local optima, but also improves the performance in medium-complexity environments such as different ground objects with fuzzy edges, and increases the generalization ability of the model.

[0156] As Figure 4 shown, another embodiment of the present invention also provides a hyperspectral image classification device 100, including:

[0157] A capture module for capturing image information of the obtained hyperspectral image through a central difference convolution algorithm and generating a first-order feature map based on the captured image information;

[0158] A modeling module for performing feature association on the features in the first-order feature map based on spectral similarity and performing graph structure modeling based on the association result to obtain a structure feature map corresponding to the hyperspectral image;

[0159] An extraction module for performing multi-dimensional feature extraction on the structure feature map to obtain multi-dimensional features;

[0160] An operation module, configured to perform dot product operation and normalized exponential operation on the multi-dimensional features to obtain the importance degree of each feature under each dimension and the data level representation corresponding to the feature importance degree;

[0161] A first processing module, configured to perform non-linear processing on the structural feature map for KAN layers to obtain a non-linear processing result;

[0162] A second processing module, configured to perform parallel processing on the non-linear processing result, the importance degree of each feature, and the data level representation by using a multi-dimensional attention mechanism to obtain a processing result including the global features, local detail features, and feature importance degree of the hyperspectral image;

[0163] A classification module, configured to classify the hyperspectral image based on the processing result.

[0164] In one embodiment, the capturing of image information on the obtained hyperspectral image through the central difference convolution algorithm includes:

[0165] Performing feature extraction on the obtained hyperspectral image through three standard convolution models with different scales and three central difference convolution models to obtain spectral features and texture features at different scales.

[0166] In one embodiment, the apparatus further includes:

[0167] A reshaping module, configured to reshape the features in the first-order feature map into a two-dimensional node set, and each node in the two-dimensional node set corresponds to a pixel position and its feature vector in the first-order feature map;

[0168] An association module, configured to perform feature association on the features in the first-order feature map by combining the two-dimensional node set.

[0169] In one embodiment, the performing of feature association on the features in the first-order feature map based on spectral similarity includes:

[0170] Determining the spectral distance between a first node in the two-dimensional node set and a second node around it;

[0171] Determining the second nodes corresponding to each of the spectral distances that meet the distance requirement;

[0172] Determining the second nodes with spectral distances meeting the requirement as the neighbor nodes of the first node;

[0173] Repeating the above steps to traverse each node in the two-dimensional node set to obtain the neighbor nodes of each node;

[0174] Construct a neighborhood matrix based on the neighbor nodes of each of the said nodes, and the neighbor matrix records the connection relationships between the nodes in the graph structure corresponding to the hyperspectral image.

[0175] In one embodiment, the multi-dimensional feature extraction of the structural feature map to obtain multi-dimensional features includes:

[0176] Adopt a three-axis attention mechanism to perform multi-channel feature extraction on the structural feature map in the horizontal, vertical, and spatial directions to obtain horizontal features, vertical features, and spatial features.

[0177] In one embodiment, the dot product operation and normalization exponential operation on the multi-dimensional features to obtain the importance degree of each feature in each dimension and the data hierarchical representation of the corresponding feature importance degree include:

[0178] Construct a weight matrix, where each row of the weight matrix corresponds to a weight vector of a hyperspectral image category, and the weight vector is used to map the multi-dimensional features into probabilities of different categories;

[0179] Perform a dot product operation on the multi-dimensional features and the weight matrix to obtain the importance degree of each feature in each dimension;

[0180] Perform a normalization exponential operation on each dimension feature in the multi-dimensional features in the corresponding dimension to obtain local features with high importance in each dimension, and perform hierarchical modeling in combination with the local features.

[0181] In one embodiment, the non-linear processing of the structural feature map for KAN layers to obtain a non-linear processing result includes:

[0182] Perform a normalization process on the structural feature map;

[0183] Use a target spline of a specified order through a spline basis function to calculate the structural feature map after the normalization process to obtain the spline basis function values of each feature in the structural feature map;

[0184] Perform non-linear processing on the structural feature map based on the spline basis function values of each feature to obtain a non-linear processing result.

[0185] In one embodiment, the use of a multi-dimensional attention mechanism to perform parallel processing on the non-linear processing result, the importance degree of each feature, and the data hierarchical representation to obtain a processing result including the global feature, local detail feature, and feature importance degree of the hyperspectral image includes:

[0186] Based on the gaze attention mechanism, process the non-linear processing result and the importance degree of each feature to obtain the scanning information and spectral global information of the hyperspectral image;

[0187] Based on the attention mechanism and the scaling factor, the importance of each of the features and the data hierarchical representation are processed to obtain the focus information and the local detail information of the spectrum in the hyperspectral image.

[0188] Another embodiment of the present invention further provides an electronic device, including:

[0189] One or more processors;

[0190] A memory configured to store one or more programs;

[0191] When the one or more programs are executed by the one or more processors, the one or more processors are caused to implement the hyperspectral image classification method as described in any one of the above.

[0192] Furthermore, an embodiment of the present invention further provides a storage medium, on which a computer program is stored, and when the program is executed by a processor, the hyperspectral image classification method as described above is implemented. It should be understood that each of the solutions in this embodiment has the corresponding technical effects in the above method embodiment, and will not be elaborated here.

[0193] Furthermore, an embodiment of the present invention further provides a computer program product, the computer program product is tangibly stored on a computer-readable medium and includes computer-readable instructions, and when the computer-executable instructions are executed, at least one processor is caused to execute the hyperspectral image classification method such as in the above-described embodiment.

[0194] It should be noted that the computer storage medium of the present invention can be a computer-readable signal medium, a computer-readable storage medium, or any combination of the two. A computer-readable medium can, for example, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the above. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, a random access storage medium (RAM), a read-only storage medium (ROM), an erasable programmable read-only storage medium (EPROM or flash memory), an optical fiber, a portable compact disk read-only storage medium (CD-ROM), an optical storage medium, a magnetic storage medium, or any suitable combination of the above. In the present invention, a computer-readable storage medium can be any tangible medium that contains or stores a program, which can be used by or in conjunction with an instruction execution system, apparatus, or device. In the present invention, a computer-readable signal medium can include a data signal propagated in a baseband or as part of a carrier wave, which carries computer-readable program code. Such a propagated data signal can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program configured to be used by or in conjunction with an instruction execution system, apparatus, or device. The program code contained on a computer-readable medium can be transmitted using any appropriate medium, including but not limited to: wireless, antenna, optical cable, RF, etc., or any suitable combination of the above.

[0195] In addition, those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) that contain computer-usable program code.

[0196] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the processes Figure 1One process or multiple processes and / or boxes Figure 1 A system with the functions specified in one box or multiple boxes.

[0197] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory produce a manufactured article including an instruction system that implements the functions specified in Figure 1 One process or multiple processes and / or boxes Figure 1 One box or multiple boxes.

[0198] Those of ordinary skill in the art should understand that the discussion of any of the above embodiments is only exemplary and is not intended to imply that the scope of protection of this application is limited to these examples; under the concept of this application, the technical features in the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations in different aspects of one or more embodiments of this application as described above, and for the sake of brevity, they are not provided in detail.

Claims

1. A hyperspectral image classification method, characterized in that, Including: Capturing image information of the obtained hyperspectral image through a central difference convolution algorithm, and generating a first-order feature map based on the captured image information; Performing feature association on the features in the first-order feature map based on spectral similarity, and performing graph structure modeling based on the association result to obtain a structure feature map corresponding to the hyperspectral image; Performing multi-dimensional feature extraction on the structure feature map to obtain multi-dimensional features; Performing dot product operation and normalized exponential operation on the multi-dimensional features to obtain the importance degree of each feature in each dimension and the data-level representation corresponding to the feature importance degree; Performing non-linear processing on the structure feature map for KAN layers to obtain a non-linear processing result; Using a multi-dimensional attention mechanism to perform parallel processing on the non-linear processing result, the importance degree of each feature, and the data-level representation to obtain a processing result including the global feature, local detail feature, and feature importance degree of the hyperspectral image; Classifying the hyperspectral image based on the processing result.

2. The hyperspectral image classification method according to claim 1, wherein The capturing of image information of the obtained hyperspectral image through a central difference convolution algorithm includes: Performing feature extraction at different scales on the obtained hyperspectral image through three standard convolution models with different scales and three central difference convolution models to obtain spectral features and texture features at different scales.

3. The hyperspectral image classification method according to claim 1, wherein The method further includes: Reshaping the features in the first-order feature map into a two-dimensional node set, where each node in the two-dimensional node set corresponds to a pixel position and its feature vector in the first-order feature map; Performing feature association on the features in the first-order feature map in combination with the two-dimensional node set.

4. The hyperspectral image classification method according to claim 3, wherein The performing of feature association on the features in the first-order feature map based on spectral similarity includes: Determining the spectral distance between a first node in the two-dimensional node set and second nodes around it; Determining the second nodes corresponding to each of the spectral distances that meet the distance requirement; Determining the second nodes with spectral distances meeting the requirement as the neighbor nodes of the first node; Repeating the above steps to traverse each node in the two-dimensional node set to obtain the neighbor nodes of each node; Constructing an adjacency matrix based on the neighbor nodes of each node, where the neighbor matrix records the connection relationship between nodes in the graph structure corresponding to the hyperspectral image.

5. The hyperspectral image classification method according to claim 1, characterized in that, The performing of multi-dimensional feature extraction on the structure feature map to obtain multi-dimensional features includes: Using a three-axis attention mechanism to perform multi-channel feature extraction on the structure feature map in the horizontal, vertical, and spatial directions to obtain horizontal features, vertical features, and spatial features.

6. The hyperspectral image classification method according to claim 1, characterized in that The performing of dot product operation and normalized exponential operation on the multi-dimensional features to obtain the importance degree of each feature in each dimension and the data-level representation corresponding to the feature importance degree includes: Constructing a weight matrix, where each row of the weight matrix corresponds to a weight vector of a hyperspectral image category, and the weight vector is used to map multi-dimensional features into probabilities of different categories; Performing dot product operation on the multi-dimensional features and the weight matrix to obtain the importance degree of each feature in each dimension; Perform a normalized exponential operation on each dimension feature in the multi-dimensional feature in the corresponding dimension to obtain local features with high importance in each dimension, and perform hierarchical modeling in combination with the local features.

7. The hyperspectral image classification method according to claim 1, characterized in that The non-linear processing of the structural feature map through the KAN layer to obtain a non-linear processing result includes: Perform normalization processing on the structural feature map; Use a target spline of a specified order through a spline basis function to calculate the structural feature map after normalization processing to obtain the spline basis function value of each feature in the structural feature map; Perform non-linear processing on the structural feature map based on the spline basis function value of each feature to obtain a non-linear processing result.

8. The hyperspectral image classification method according to claim 1, wherein The use of a multi-dimensional attention mechanism to perform parallel processing on the non-linear processing result, the importance of each feature, and the data hierarchical representation to obtain a processing result including the global feature, local detail feature, and feature importance of the hyperspectral image, includes: Based on the gaze attention mechanism, process the non-linear processing result and the importance of each feature to obtain the scanning information and spectral global information of the hyperspectral image; Based on the attention mechanism and the scaling factor, process the importance of each feature and the data hierarchical representation to obtain the focus information and local detail information of the spectrum in the hyperspectral image.

9. A hyperspectral image classification device, characterized in that, Includes: A capture module for capturing image information of the obtained hyperspectral image through a central difference convolution algorithm and generating a first-order feature map based on the captured image information; A modeling module for performing feature association on the features in the first-order feature map based on spectral similarity and performing graph structure modeling based on the association result to obtain a structural feature map corresponding to the hyperspectral image; An extraction module for performing multi-dimensional feature extraction on the structural feature map to obtain multi-dimensional features; An operation module for performing dot product operation and normalized exponential operation on the multi-dimensional features to obtain the importance of each feature under each dimension and the data hierarchical representation corresponding to the feature importance; A first processing module for performing non-linear processing of the structural feature map through the KAN layer to obtain a non-linear processing result; A second processing module for using a multi-dimensional attention mechanism to perform parallel processing on the non-linear processing result, the importance of each feature, and the data hierarchical representation to obtain a processing result including the global feature, local detail feature, and feature importance of the hyperspectral image; A classification module for classifying the hyperspectral image based on the processing result.

10. An electronic device, characterized in that, Includes: One or more processors; A memory configured to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the hyperspectral image classification method according to any one of claims 1-8.

Images