Agricultural hyperspectral image classification method based on hypergraph and Mama

Abstract

The invention discloses an agricultural hyperspectral image classification method based on a hypergraph and Mama, and the method comprises the steps: S1, constructing a spectrum-space hypergraph according to an agricultural hyperspectral image; s2, a Mama framework used for hyperspectral classification is constructed; and S3, inputting the spectrum-space hypergraph into a Mama framework for hyperspectral classification, and carrying out classification, the method effectively fuses space and spectrum information, improves the precision of hyperspectral image classification, and solves the problem of information capture in a traditional method.

Description

Technical Field

[0001] The present application relates to the field of image processing and classification, and in particular to an agricultural hyperspectral image classification method based on hypergraph and Mamba. Background Art

[0002] With the development of remote sensing technology, agricultural hyperspectral imagery (HSI) can be widely acquired. Unlike traditional visual systems that only use RGB channels to collect images, agricultural hyperspectral imagery covers a wider spectral range from the visible spectrum to the near-infrared, mid-infrared, and even far-infrared, typically with dozens to hundreds of continuous bands. Due to its rich spectral information, HSI can help accurately identify subtle differences in the agricultural field, such as crop species, soil type, and pests and diseases, which are often difficult to distinguish in RGB images. For this reason, agricultural hyperspectral imagery provides more data support for precision agriculture, enabling tasks such as precision fertilization, pest and disease monitoring, and crop growth prediction.

[0003] Hyperspectral image classification, as an important task in agricultural remote sensing applications, aims to predict the category label of each pixel in the image. In the past few years, thanks to the development of deep learning technology, hyperspectral image classification has made significant progress. Deep learning (DL)-based methods mainly include graph convolutional networks (GCN), convolutional neural networks (CNN), and Transformer methods. However, HSI classification methods based on CNN and Transformer still have some inherent limitations. CNN-based models are limited by local receptive fields and have difficulty in effectively modeling long-range dependencies. Although the Transformer-based model performs well in global modeling, the computational complexity of its self-attention mechanism is quadratic, resulting in a heavy computational burden when processing high-resolution agricultural hyperspectral images, affecting the extraction of fine-grained spatial features at the pixel level.

[0004] In recent years, structured state-space models (S4) and Mamba, as efficient deep learning layers, have achieved superior performance in analyzing long-sequence data. Mamba improves upon traditional S4 by incorporating a selection mechanism that dynamically selects relevant information based on input dependencies. Combined with hardware-aware implementation, this significantly improves computational efficiency. Due to its powerful long-range modeling capabilities and linear computational complexity, Mamba has been widely used in fields such as language understanding, medical image analysis, and computer vision. However, traditional Mamba methods fail to consider the spectral continuity of hyperspectral images, making them difficult to directly apply to agricultural hyperspectral image classification.

[0005] In addition, most existing hyperspectral image classification methods ignore the close correlation between spatial and spectral features. Hyperspectral images not only contain rich spectral information, but also have significant spatial structure. The relationship between the two is crucial for agricultural image classification. However, most existing methods process spatial and spectral information separately and fail to effectively fuse them, resulting in insufficient classification accuracy and robustness. The Mamba algorithm, as a strong sequential model, performs well in some areas. However, in the classification of agricultural hyperspectral images, since it is mainly designed for sequential data, it is easy to encounter difficulties in capturing local spatial and spectral features. In particular, when dealing with local changes and spatial dependencies in crop areas, it may not be able to fully utilize the complex information in hyperspectral images, thus affecting the classification effect. Summary of the Invention

[0006] In response to the above-mentioned deficiencies in the prior art, the present invention provides an agricultural hyperspectral image classification method based on hypergraph and Mamba, which solves the problems that the existing Mamba-based method has strong sequentiality, is prone to insufficient information capture in capturing spatial and spectral features when applied to hyperspectral images, and has limited modeling capabilities for long-distance dependencies.

[0007] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: a method for agricultural hyperspectral image classification based on hypergraph and Mamba, comprising:

[0008] S1. Construct a spectral-spatial hypermap based on agricultural hyperspectral images.

[0009] S2, building the Mamba architecture for hyperspectral classification;

[0010] S3. Input the spectral-spatial hypergraph into the Mamba architecture for hyperspectral classification for classification.

[0011] Furthermore: S1 includes:

[0012] S101. In the spectral feature space of the agricultural hyperspectral image, each pixel is regarded as a central vertex. K pixels with the closest feature distance to the central vertex are selected using the KNN algorithm, and the pixels are connected to form a spectral hyperedge.

[0013] S102, in the spectral feature space of the agricultural hyperspectral image, connecting the central pixel and its neighbors in spatial distance to form a spatial hyperedge;

[0014] S103 , performing weighted fusion of the spectral hyperedge and the spatial hyperedge and performing graph embedding and dimensionality reduction to obtain a spectral-spatial hypergraph.

[0015] Further: The weight w of the spectral hyperedge Fea The expression is:

[0016]

[0017] Among them, v i represents the i-th vertex, Represents a supernode The average distance between all vertices in , Represents a supernode The average distance between all vertices in x i and x j Represent the i-th pixel and the j-th pixel respectively, and n represents the total number of pixels;

[0018] Spatial hyperedge weight w Spa The expression is:

[0019] Further: In S103, the Laplacian matrix L of the spectral-spatial hypergraph is expressed as:

[0020] L=(1-α)L Spe +αL Spa

[0021] Among them, α is the equilibrium parameter, L Spe and L Spa are the Laplace matrices of the spectral hyperedge and spatial hyperedge, respectively, where L Spe =D Spe -W Spe , L Spe =D Spe -W Spe , D Spe and D Spa are the degree matrices of the spectral and spatial hyperedges, W Spe and W Spa are the weight matrices of spectral and spatial hyperedges, respectively.

[0022] Furthermore: In S2, the Mamba architecture for hyperspectral classification includes:

[0023] Embedding layer, used to project the feature vector of the spectral-spatial hypergraph into the embedding space;

[0024] The encoder backbone is used to extract discriminative spatial spectral features for classification;

[0025] The segmentation head is used to obtain the final classification result through the convolutional layer.

[0026] Further: The encoder backbone includes:

[0027] Spatial Mamba module, used to extract spatial features;

[0028] Spectral Mamba module, used to capture spectral features;

[0029] The spatial-spectral fusion module is used to integrate spatial features and spectral features.

[0030] Furthermore, the forward transfer process of the spatial Mamba module is expressed as:

[0031] HF spa =Flatten(H l )

[0032] HR spa =SiLU(GN(Mamba(HF spa )))

[0033]

[0034] Among them, Flatten() represents the Flatten layer, SiLU() represents the SiLU activation function, Reshape() represents the Reshape layer, H l ∈R B×H×W×D The fine-grained pixel embedding obtained by the embedding layer is divided into denotes the pixel-level input embedding and SpaMB output features, R denotes the dimension, B, H, W, D denotes the batch size, image height, image width and embedding dimension, respectively, HF spa ∈R B×L1×D and HR spa ∈R B×L1×D They represent the flattened input and the learned residual space features respectively, L1=H×W, and Mamba(.) represents the standard Mamba block.

[0035] Furthermore: the spectral Mamba module divides the spectral features into spectral groups, models the relationships between different spectral groups, and updates the spectral features based on the relationships between the spectral groups;

[0036] Spectral features extracted by the Spectral Mamba module The expression is:

[0037] HG spe =Split Spe ctralGroup(H l )

[0038] HF spe =Flatten(HG spe )

[0039] HR spe =SiLU(Gn(Mamba(HF spa )))

[0040]

[0041] Among them, HF spe ∈R B×H×W×G×M , HFpe∈R N×G×M , HR spe ∈R N×G×M , They represent the divided spectral group features, flattened spectral group features, residual spectral group features and output spectral group features respectively, G represents the number of groups into which the semantic vector is split, M is the pixel embedding dimension, and N = B × H × W represents the intermediate parameters.

[0042] Furthermore, the fusion process of the spatial-spectral fusion module is expressed as:

[0043]

[0044] Among them, w spa and w spe represent the spatial and spectral fusion weights respectively.

[0045] Furthermore: Before S3, graph embedding technology is used to reduce the dimensionality of the spectral-spatial hypergraph. The objective function of the embedding process is:

[0046]

[0047] Y T =P T X

[0048] Among them, X represents the sample matrix, P represents the projection matrix, and P T represents the transpose of the projection matrix P, L represents the Laplace matrix, Tr(·) represents the trace operation of the matrix, and Y represents the multidimensional array after the dimensionality reduction of X.

[0049] The beneficial effects of the present invention are:

[0050] 1. By constructing two different hyperedges to simulate the spectral and spatial relationships between pixels, and further fusing these two types of information to generate a spectral-spatial hypergraph;

[0051] 2. Use graph embedding methods to reduce the dimensionality of the generated hypergraph, reducing computational complexity;

[0052] 3. A Mamba architecture for hyperspectral classification was constructed. This framework modeled long-range dependencies, comprehensively utilized spatial and spectral information, and supported end-to-end training. It accurately extracted spatial and spectral information through the spatial Mamba module and the spectral Mamba module, and adaptively fused these two types of information through the spatial-spectral fusion module to further enhance feature expression. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 Flowchart of the agricultural hyperspectral image classification method based on hypergraph and Mamba.

[0054] Figure 2 Schematic diagram of the process of constructing the spectral-spatial hypergraph.

[0055] Figure 3 Schematic diagram of the Mamba architecture for hyperspectral classification. DETAILED DESCRIPTION

[0056] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0057] like Figure 1 As shown, in one embodiment of the present invention, a method for agricultural hyperspectral image classification based on a hypergraph and Mamba is provided, comprising:

[0058] S1. Construct a spectral-spatial hypermap based on agricultural hyperspectral images.

[0059] S2, building the Mamba architecture for hyperspectral classification;

[0060] S3. Input the spectral-spatial hypergraph into the Mamba architecture for hyperspectral classification for classification.

[0061] A hypergraph is a generalized graph model. In a hypergraph, edges no longer simply connect two vertices, but can connect any number of vertices. The hypergraph structure can be well applied in image processing.

[0062] Previous Mamba series models first split images into patches, then flatten the patches into sequences and feed them into the model one by one. This approach places very strict requirements on the order of the input sequences and may not be suitable for non-autoregressive vision models. Vision tasks typically require converting images into one-dimensional sequences and performing recursive computations. This limits the receptive field of each image token to the previous sequence, introduces additional latency, and thus fails to effectively capture long-range dependencies.

[0063] In order to better obtain adjacent structural features and capture long-range spectral dependency feature information, we design a joint hypergraph G to achieve this goal. Let G =<V,E,W> , where V is the finite set of vertices of the hypergraph, each vertex represents a pixel in the agricultural hyperspectral image, E is the set of hyperedges of the hypergraph, and W is the weight set of the hyperedges.

[0064] like Figure 2 As shown, S1 includes:

[0065] S101. In the spectral feature space of the agricultural hyperspectral image, each pixel is regarded as a central vertex. K pixels with the closest feature distance to the central vertex are selected using the KNN algorithm, and the pixels are connected to form a spectral hyperedge.

[0066] Thus, there are K+1 vertices in each hyperedge, and the weight of the spectral hyperedge is w Fea The expression is:

[0067]

[0068] Among them, v i represents the i-th vertex, Represents a supernode The average distance between all vertices in x i and x j represent the i-th pixel and the j-th pixel respectively, Represents a supernode The average distance between all vertices in x i and x j Denote the i-th pixel and the j-th pixel respectively, and n denotes the total number of pixels; thus, hyperedges with smaller spacing between pixel pairs will be assigned higher weights;

[0069] S102, in the spectral feature space of the agricultural hyperspectral image, connecting the central pixel and its neighbors in spatial distance to form a spatial hyperedge;

[0070] Assume that we find k neighbors of the central pixel in terms of spatial distance, and form a hyperedge with the k spatial neighbors and the central pixel. Then there are k+1 vertices in the hyperedge.

[0071] Spatial hyperedge weight w Spa The expression is: This means that each hyperedge is constructed for a single pixel and has a fixed effect on the hypergraph structure;

[0072] S103, performing weighted fusion of the spectral hyperedge and the spatial hyperedge and performing graph embedding and dimensionality reduction to obtain a spectral-spatial hypergraph;

[0073] The Laplacian matrix L of the spectral-spatial hypergraph is expressed as:

[0074] L=(1-α)L Spe +αL Spa

[0075] Among them, α is the equilibrium parameter, D Spe and L Spa are the Laplace matrices of the spectral hyperedge and spatial hyperedge, respectively, where D Spe =D Spe -W Spe , L Spa =D Spa -W Spa , D Spe and D Spa are the degree matrices of the spectral and spatial hyperedges, W Spe and W Spa are the weight matrices of spectral and spatial hyperedge respectively, and the expressions are:

[0076]

[0077] in, and is the weight of the kth spectral hyperedge and spatial hyperedge.

[0078] In particular, since hyperspectral data is a typical high-dimensional data, which often brings challenges such as the curse of dimensionality, to address these problems, we use graph embedding technology to reduce the dimensionality of the spectral-spatial hypergraph. Considering the nonlinear characteristics of hyperspectral data, the embedding process is defined by the following objective function:

[0079]

[0080] Among them, X represents the sample matrix, P represents the projection matrix, and P T represents the transpose of the projection matrix P, L represents the Laplacian matrix, which is used to encode the local geometric structure, and Tr(·) represents the trace operation of the matrix.

[0081] In this optimization objective, the trace operation converts the quadratic form P T XLX T P is converted into a scalar value. By minimizing this value, the projected data can maintain the local neighborhood structure of the original high-dimensional data in the low-dimensional space. This method effectively preserves the local structure of the data, ensuring that points that are close in high-dimensional space remain close in low-dimensional space.

[0082] In particular, there are the following relations:

[0083] Y T =P T X

[0084] Among them, Y represents the multidimensional array after X dimension reduction, and then the dimension-reduced Y is reshaped into I∈R through the Reshape() operation. H×W×C , I is expressed as a low-dimensional spectral-spatial hypergraph:

[0085] I=Reshape(Y)

[0086] Hyperspectral image classification is a pixel-level classification task. This means that the Mamba architecture for hyperspectral classification needs to learn features that can finely distinguish subtle differences between pixels. Due to inherent limitations such as the locality of CNNs and the quadratic computational complexity of Transformers, existing agricultural hyperspectral image classification methods have limited ability to model long-range dependencies. Due to the quadratic computational complexity of Transformers, Transformer-based models often use image patches as basic representation units when processing images, which hinders the model's ability to extract finer pixel-level representations for classification. This motivates us to develop a new framework to model long-range dependencies with linear complexity, thereby enabling finer feature extraction at the pixel level for classification. Therefore, we introduce Mamba as a basic unit to model long-range dependencies, which is crucial for obtaining discriminative pixel-level spatial features. However, Mamba ignores the fact that hyperspectral (HS) images have approximately continuous spectral information.

[0087] Therefore, if Figure 3 As shown, the Mamba architecture for hyperspectral classification consists of three main components:

[0088] Embedding layer, used to project the feature vector of the spectral-spatial hypergraph into the embedding space;

[0089] Encoder backbone (Encoder), used to extract discriminative spatial spectral features for classification;

[0090] Segmentation Head: used to obtain the final classification result through the convolutional layer.

[0091] In particular, unlike existing block-based embedding methods, the embedding layer proposed in this application extracts the embedding of each pixel, which enables our model to obtain more fine-grained pixel embedding, making it more suitable for dense prediction tasks such as hyperspectral classification. Specifically, from the low-dimensional spectral-spatial hypergraph I∈R H×W×C Fine-grained pixel embedding E′∈R can be obtained H ×W×D :

[0092] E′=Embedding(I)=SiLU(GN(Conv(I)))

[0093] where Conv, GN, and SiLU represent the convolutional layer with a kernel size of 1×1, the group norm layer, and the SiLU activation function, respectively. H, W, and C represent the height, width, and number of spectral channels of the input low-dimensional spectral-spatial hypergraph, respectively. D is the embedding dimension.

[0094] The encoder backbone is used to extract discriminative spatial-spectral features for classification. Specifically, the encoder backbone consists of three parts: the spatial Mamba module for extracting spatial features, the spectral Mamba module for capturing spectral features, and the spatial-spectral fusion module for integrating spatial and spectral features. The encoder backbone process is defined as follows:

[0095] H=Encoder(E′)

[0096] Where Encoder and H represent the encoder backbone and the extracted hidden features, respectively.

[0097] The segmentation head uses a convolutional layer with a kernel size of 1×1 to obtain the final result l, that is, l=SegHead(H).

[0098] Spatial Mamba Block:

[0099] Hyperspectral classification is a pixel-level classification task. This means that the representation used for classification needs to meet two conditions. First, the representation should be refined and should reflect the differences between pixels. Therefore, unlike existing patch partitioning methods, we extract embeddings at the pixel level. Second, the representation should be discriminative for classification. Therefore, we require the proposed module to have strong long-range modeling capabilities. We adopt the Mamba layer as the basic unit to construct the spatial feature extractor, which can model long-range dependencies with linear computational complexity.

[0100] The detailed structure of the space Mamba module is as follows Figure 3 As shown, the forward pass process is expressed as:

[0101] HF spa =Flatten(H l )

[0102] HR spa =SiLU(GN(Mamba(HF spa )))

[0103]

[0104] Among them, Flatten() represents the Flatten layer, SiLU() represents the SiLU activation function, Reshape() represents the Reshape layer, Hl ∈R B×H×W×D Divided by E′, denotes the pixel-level input embedding and the output features of the spatial Mamba block, R denotes the dimension, B, H, W, D denote the batch size, image height, image width and embedding dimension, respectively, HF spa ∈R B×L1×D and HR spa ∈R B×L1×D denote the flattened input and the learned residual space features, L1 = H × W, Mamba(.) denotes the standard Mamba block. In the spatial Mamba module, the design of the group norm (GN) and residual connection facilitates learning.

[0105] Spectral Mamba Block:

[0106] Unlike traditional visual systems that collect images through RGB channels, hyperspectral images can cover a larger spectral range and higher spectral resolution. The spectral Mamba block divides spectral features into spectral groups, models the relationship between different spectral groups, and updates the spectral features based on the relationship between spectral groups, thereby modeling the relationship between spectra and extracting discriminative features. The spectral features extracted by the spectral Mamba module The expression is:

[0107] HG spe =SplitSpectralGroup(H l )

[0108] HF spe =Flatten(HG spe )

[0109] HR spe =SiLU(GN(Mamba(HF spe )))

[0110]

[0111] Among them, HG spe ∈R B×H×W×G×M , HF spe ∈R N×G×M , HR spe ∈R N×G×M , They represent the divided spectral group features, flattened spectral group features, residual spectral group features and output spectral group features respectively, G represents the number of groups into which the semantic vector is split, M is the pixel embedding dimension, and N = B × H × W represents the intermediate parameters.

[0112] Spatial-Spectral Fusion Module:

[0113] Spatial and spectral features are crucial for the classification of agricultural hyperspectral images. Integrating spatial and spectral information is beneficial to classification, which prompted this application to design a spatial-spectral fusion module, whose structure is as follows Figure 3 As shown in Figure 2. In particular, considering that hyperspectral classification usually has fewer labeled samples, we introduce the idea of residual learning to alleviate the overfitting phenomenon that may occur during training. The spatial-spectral fusion module adaptively estimates the importance of space and spectrum to guide fusion. The fusion process is expressed as:

[0114]

[0115] Among them, w spa and w spe Represent the fusion weights of space and spectrum respectively, and randomly initialize w spa and w spe , these weights are updated through back-propagation to determine the final fusion weights.

[0116] During the training phase, the training loss function Loss is given as:

[0117] Loss=CrossEntropy(Graph,Y tr )

[0118] Among them, Y tr are training labels, and Graph represents a low-dimensional spectral-spatial hypergraph.

[0119] In the inference phase, given a test graph I te , we can get the final prediction results

[0120]

[0121] Among them, MambaHSI(I te ) are the logits output by the Mamba architecture for hyperspectral classification.

[0122] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A hyperspectral agricultural image classification method based on hypergraph and Mamba, characterized by: include: S1. Construct a spectral-spatial hypermap based on agricultural hyperspectral images. S2, building the Mamba architecture for hyperspectral classification; S3. Input the spectral-spatial hypergraph into the Mamba architecture for hyperspectral classification for classification.

2. The agricultural hyperspectral image classification method based on hypergraph and Mamba according to claim 1 is characterized in that: S1 includes: S101. In the spectral feature space of the agricultural hyperspectral image, each pixel is regarded as a central vertex. K pixels with the closest feature distance to the central vertex are selected using the KNN algorithm, and the pixels are connected to form a spectral hyperedge. S102, in the spectral feature space of the agricultural hyperspectral image, connecting the central pixel and its neighbors in spatial distance to form a spatial hyperedge; S103 , performing weighted fusion of the spectral hyperedge and the spatial hyperedge and performing graph embedding and dimensionality reduction to obtain a spectral-spatial hypergraph.

3. The agricultural hyperspectral image classification method based on hypergraph and Mamba according to claim 2 is characterized in that: The weight w of the spectral hyperedge Fea The expression is: Among them, v i represents the i-th vertex, Represents a supernode The average distance between all vertices in , Represents a supernode The average distance between all vertices in x i and x j Represent the i-th pixel and the j-th pixel respectively, and n represents the total number of pixels; Spatial hyperedge weight w Spa The expression is:

4. The agricultural hyperspectral image classification method based on hypergraph and Mamba according to claim 3 is characterized in that: In S103, the Laplacian matrix L of the spectral-spatial hypergraph is expressed as: L=(1-α)L Spe +αL Spa Among them, α is the equilibrium parameter, L Spe and L Spa are the Laplace matrices of the spectral hyperedge and spatial hyperedge, respectively, where L Spe =D Spe -W Spe , L Spa =D Spa -W Spa , D Spa and D Spa are the degree matrices of the spectral and spatial hyperedges, W Spe and W Spa are the weight matrices of spectral and spatial hyperedges, respectively.

5. The agricultural hyperspectral image classification method based on hypergraph and Mamba according to claim 1, characterized in that: In S2, the Mamba architecture for hyperspectral classification includes: Embedding layer, used to project the feature vector of the spectral-spatial hypergraph into the embedding space; The encoder backbone is used to extract discriminative spatial spectral features for classification; The segmentation head is used to obtain the final classification result through the convolutional layer.

6. The agricultural hyperspectral image classification method based on hypergraph and Mamba according to claim 5, characterized in that: The encoder backbone consists of: Spatial Mamba module, used to extract spatial features; Spectral Mamba module, used to capture spectral features; The spatial-spectral fusion module is used to integrate spatial features and spectral features.

7. The agricultural hyperspectral image classification method based on hypergraph and Mamba according to claim 6, characterized in that: The forward transfer process of the spatial Mamba module is expressed as: HF spa =Flatten(H l ) HR spa =SiLU(GN(Mamba(HF spa ))) Among them, Flatten() represents the Flatten layer, SiLU() represents the SiLU activation function, Reshape() represents the Reshape layer, H l ∈R B×H×W×D The fine-grained pixel embedding obtained by the embedding layer is divided into denotes the pixel-level input embedding and SpaMB output features, R denotes the dimension, B, H, W, D denote the batch size, image height, image width and embedding dimension, respectively, HG spa ∈R B×L1×D and HR spa ∈R B×L1×D They represent the flattened input and the learned residual space features respectively, L1=H×W, and Mamba(.) represents the standard Mamba block.

8. The agricultural hyperspectral image classification method based on hypergraph and Mamba according to claim 7, characterized in that: The spectral Mamba module divides spectral features into spectral groups, models the relationships between different spectral groups, and updates spectral features based on the relationships between spectral groups; Spectral features extracted by the Spectral Mamba module The expression is: HG spe =SplitSpectralGroup(H l ) HF spe =Flatten(HG spe ) HR spe =SiLU(GN(Mamba(HF spe ))) Among them, HG spe ∈R B×H×W×G×M , HF spe ∈R N×G×M , HR spe ∈R N×G×M , They represent the divided spectral group features, flattened spectral group features, residual spectral group features and output spectral group features respectively, G represents the number of groups into which the semantic vector is split, M is the pixel embedding dimension, and N = B × H × W represents the intermediate parameters.

9. The agricultural hyperspectral image classification method based on hypergraph and Mamba according to claim 8, characterized in that: The fusion process of the spatial-spectral fusion module is expressed as: Among them, w spa and w spe represent the spatial and spectral fusion weights respectively.

10. The agricultural hyperspectral image classification method based on hypergraph and Mamba according to claim 3, characterized in that: Before S3, graph embedding technology was used to reduce the dimensionality of the spectral-spatial hypergraph. The objective function of the embedding process was: Y T =P T X Among them, X represents the sample matrix, P represents the projection matrix, and P T represents the transpose of the projection matrix P, L represents the Laplace matrix, Tr(·) represents the trace operation of the matrix, and Y represents the multidimensional array after the dimensionality reduction of X.

Images