A Classification Method and System for Hyperspectral Images

Abstract

The present invention discloses a classification method and system for hyperspectral images, relating to the technical field of hyperspectral image classification, including: obtaining multi-scale features by feature extraction from a hyperspectral image dataset; using ADMM to construct a joint optimization objective function as a joint optimization model; setting the dictionary matrix as a trainable parameter to implicitly learn a dynamic dictionary in an end-to-end manner; inputting the multi-scale features into the joint optimization model to iteratively solve for each variable; adopting a modal decomposition strategy to perform multi-dimensional deconstruction on the auxiliary variable Z, and optimizing the solution of the auxiliary variable through a dual-branch prior network based on the Mamba state space model; the joint optimization model outputs a sparse coefficient matrix and implicit dictionary features; splicing the sparse coefficient features and the dictionary features to obtain fused features; inputting the fused features into a classifier to obtain the classification result of the hyperspectral image. The present invention can improve the accuracy of the classification result of hyperspectral data.

Description

Technical Field

[0001] The present invention relates to the technical field of hyperspectral image classification, and particularly relates to a classification method and system for hyperspectral images. Background Art

[0002] Hyperspectral remote sensing images have the characteristic of "one image with multiple spectra", and can record the spatial distribution and spectral response characteristics of ground objects simultaneously.

[0003] The existing technologies mainly adopt two types of methods: one is a mathematical optimization framework based on the traditional Alternating Direction Method of Multipliers (ADMM), and the other is a pure data-driven deep learning model. However, the traditional ADMM method relies on artificially setting fixed thresholds and fixed iterative rules, and lacks adaptability when facing the complex band correlation of hyperspectral data; at the same time, its dictionary is usually pre-designed through mathematical rules (such as the dictionary learning algorithm K-Singular Value Decomposition (KSVD) based on sparse representation), which has nothing to do with the goal of the classification task. While the pure deep learning method can be trained end-to-end, stacking 3D convolutional layers leads to a sharp increase in computational overhead, and the "black box characteristic" of the model results in a lack of interpretability in the decision-making process. In addition, the attention mechanism introduced to improve the long-range dependence modeling ability will significantly increase the computational complexity due to the expansion of the spectral dimension and is difficult to be practically applied.

[0004] In summary, the existing technologies cannot adapt to the complex structural characteristics of hyperspectral data and capture the long-range dependence characteristics in hyperspectral data, resulting in inaccurate classification results of hyperspectral data. Summary of the Invention

[0005] The embodiments of the present invention provide a classification method and system for hyperspectral images, which can solve the problem of inaccurate classification results of hyperspectral data in the existing technologies.

[0006] The embodiments of the present invention provide a classification method for hyperspectral images, including the following steps: obtaining a hyperspectral image data set containing multiple ground object categories, and sequentially using a 3D convolutional network and a 2D convolutional network for feature extraction to obtain multi-scale features;

[0007] Using the Alternating Direction Method of Multipliers (ADMM) to construct a joint optimization objective function for hyperspectral image classification for variable iterative solution, as a joint optimization model;

[0008] Optimize the auxiliary variables in the joint optimization model using a dual-branch processing mechanism based on the state-space model Mamba; wherein, the dual-branch processing mechanism based on the state-space model Mamba includes: decomposing the auxiliary variables into multiple independent modes through modal decomposition, extracting the global features and local features of the multiple independent modes using the main branch and the context branch respectively, and using layer normalization and linear projection to combine the global features and local features obtained from the two branches to obtain optimized auxiliary variables; the main branch uses multiple linear transformations, each linear transformation with layer normalization and the SiLU activation function, and maintains feature consistency by adjusting the dimensions of the linear layer; the context branch performs spatial processing using small convolutional kernels and adds attention fusion;

[0009] Input the multi-scale features into the joint optimization model for variable iterative solution to obtain sparse coefficient features and dictionary features; concatenate the sparse coefficient features and the dictionary features to obtain fused features; input the fused features into the classifier to obtain the classification result of the hyperspectral image.

[0010] Furthermore, the joint optimization objective function L has the formula:

[0011] ;

[0012] wherein, is the multi-scale feature, is the dictionary matrix, is the sparse coefficient matrix, is the auxiliary variable, is the implicit prior regularization term for the hyperspectral image, is the Lagrange multiplier, and are both trade-off parameters.

[0013] Furthermore, the step of optimizing the auxiliary variables in the joint optimization model using the dual-branch processing mechanism based on the state-space model Mamba specifically includes:

[0014] Use the state-space model Mamba to replace the solution method of the auxiliary variable Z in ADMM;

[0015] Use an orthogonal projection matrix to decompose the auxiliary variable along the spectral mode, spatial height mode, and spatial width mode to generate orthogonal multi-modal feature components; each modal component is processed through a module Enhanced Mamba Block that integrates selective state parameterization and adaptive time-scale adjustment;

[0016] The main branch uses a sequence of linear transformations with layer normalization and the SiLU activation function, and maintains feature consistency by adjusting the dimensions of the linear layer to obtain the auxiliary variable Global features;

[0017] The context branch performs spatial processing using small convolutional kernels and adds attention fusion to obtain auxiliary variables Local features;

[0018] Use layer normalization and linear projection to fuse the features obtained from the two branches to complete the optimization of the auxiliary variables Optimization.

[0019] Furthermore, before obtaining the sparse coefficient features, it further includes:

[0020] Obtain the transpose of the dictionary matrix And multiply it with the dictionary matrix To obtain the multiplication result; According to the multiplication result, use a 3×3 convolution to replace the matrix inversion method to complete the optimization of the sparse coefficient matrix

[0021] Optimization.

[0022] Furthermore, the specific steps for obtaining multi-scale features include:

[0023] Convert the hyperspectral image dataset with dimensions (B, C, H, W), where B is the batch size, C is the number of spectral channels, and H and W are the spatial dimensions; convert it to (B, 1, C, H, W) through dimension expansion;

[0024] Input the dimension-converted hyperspectral image dataset into a 3D convolutional network for feature extraction to obtain spatial information and spectral information. The 3D convolutional network uses convolutional kernels of (7, 3, 3), (5, 3, 3), and (3, 3, 3), and uses a BatchNorm normalization layer and a ReLU non-linear activation function, and the number of channels changes sequentially to 8, 16, and 32;

[0025] Use channel attention mechanism to enhance the features of the spatial information and spectral information to obtain enhanced features;

[0026] Input the enhanced features into a 2D convolutional network for multi-scale spatial feature extraction to obtain spatial information. The 2D convolutional network has 32, 128, 256, and 512 channels, and uses max-pooling operations for hierarchical transformation of the feature map size.

[0027] Furthermore, the classifier uses a two-layer fully connected network.

[0028] Furthermore, the classifier is trained by a learning rate scheduling strategy that combines the training method warmup that can gradually increase the learning rate and cosine annealing.

[0029]

[0029] Furthermore, the loss function of the classifier L total , and the formula is:

[0030] ;

[0031] wherein, is the classification cross-entropy loss function, is the dictionary coherence loss function, w 1 and w 2 are both adjustable weight parameters.

[0032] An embodiment of the present invention provides a hyperspectral image classification system, including:

[0033] A data acquisition module, configured to acquire a hyperspectral image dataset containing multiple ground object categories, and sequentially use a 3D convolutional network and a 2D convolutional network for feature extraction to obtain multi-scale features;

[0034] A joint optimization model construction module, configured to use the alternating direction method of multipliers (ADMM) to construct a joint optimization objective function for hyperspectral image classification for variable iterative solution as a joint optimization model; use a dual-branch processing mechanism based on the state space model Mamba to optimize the auxiliary variables in the joint optimization model; wherein, the dual-branch processing mechanism based on the state space model Mamba includes: decomposing the auxiliary variables into multiple independent modes through modal decomposition, using the main branch and the context branch to extract the global features and local features of the multiple independent modes respectively, and using layer normalization and linear projection to combine the global features and local features obtained by the two branches to obtain optimized auxiliary variables; the main branch uses multiple linear transformations, each linear transformation with layer normalization and the SiLU activation function, and maintains feature consistency by adjusting the dimension of the linear layer; the context branch uses small convolutional kernels for spatial processing and adds attention fusion;

[0035] A classification result output module, configured to input the multi-scale features into the joint optimization model for variable iterative solution to obtain sparse coefficient features and dictionary features; splice the sparse coefficient features and the dictionary features to obtain fused features; input the fused features into a classifier to obtain the classification result of the hyperspectral image.

[0036] An embodiment of the present invention provides a hyperspectral image classification method and system. Compared with the prior art, its beneficial effects are as follows:

[0037] The 3D convolutional network and 2D convolutional network are successively used for feature extraction on the hyperspectral image dataset to obtain multi-scale features; the alternating direction method of multipliers (ADMM) is used to construct a joint optimization objective function for hyperspectral image classification for variable iterative solution, which is used as a joint optimization model; a dual-branch processing mechanism based on the state space model Mamba is used to optimize the auxiliary variables in the joint optimization model; among them, the dual-branch processing mechanism based on the state space model Mamba includes: decomposing the auxiliary variables into multiple independent modes through modal decomposition, using the main branch and the context branch to extract the global features and local features of the multiple independent modes respectively, and using layer normalization and linear projection to combine the global features and local features obtained by the two branches to obtain optimized auxiliary variables; the main branch uses multiple linear transformations, each linear transformation with layer normalization and SiLU activation function, and maintains feature consistency by adjusting the dimension of the linear layer; the context branch uses small convolutional kernels for spatial processing and adds attention fusion; the multi-scale features are input into the joint optimization model for variable iterative solution to obtain sparse coefficient features and dictionary features; the sparse coefficient features and dictionary features are concatenated to obtain fused features; the fused features are input into the classifier to obtain the classification result of the hyperspectral image.

[0038] Among them, the state space model Mamba has both a state transition mechanism and a state update mechanism. The learnable parameters in the state transition mechanism enable the model to adaptively adjust the structure to match the characteristics of hyperspectral image data, capture the linear relationships and complex structures in the hyperspectral image data, so as to adapt to the complex structural characteristics of hyperspectral image data; at the same time, the state update mechanism enables the update of the model state to follow the update rule of linear time, that is, the state estimation at the current time depends on the state estimation at the previous time and the observation data at the current time, so that the state space model Mamba realizes a state update mechanism with linear complexity and captures the long-range dependence characteristics of hyperspectral image data; finally, the classification result of hyperspectral image data can be accurately obtained through the state space model Mamba. Description of the Drawings

[0039] Figure 1 It is the overall flowchart of a hyperspectral image classification method provided by an embodiment of the present invention;

[0040] Figure 2 It is the feature extraction flowchart of a hyperspectral image classification method provided by an embodiment of the present invention;

[0041] Figure 3 It is the admm parameter iteration flowchart of a hyperspectral image classification method provided by an embodiment of the present invention;

[0042] Figure 4This is the classification result diagram of the Indianpines hyperspectral image for a hyperspectral image classification method provided by an embodiment of the present invention. Detailed implementation manners

[0043] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will describe the detailed implementation manners of the present invention with reference to the accompanying drawings. Many specific details are set forth in the following description to fully understand the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0044] See Figure 1 , an embodiment of the present invention provides a hyperspectral image classification method, including the following steps:

[0045] Step 1: Obtain a hyperspectral image dataset containing multiple ground object categories, and sequentially use a 3D convolutional network and a 2D convolutional network for feature extraction to obtain multi-scale features. Specifically, through the cascaded processing of a three-dimensional convolutional network (3D CNN) and a two-dimensional convolutional network (2D CNN), multi-scale features with spatial-spectral joint characteristics are extracted.

[0046] Step 2: Use the alternating direction method of multipliers (ADMM) to construct a joint optimization objective function for hyperspectral image classification for variable iterative solution, as a joint optimization model. Use the dual-branch processing mechanism based on the state space model Mamba to optimize the auxiliary variables in the joint optimization model; wherein, the dual-branch processing mechanism based on the state space model Mamba includes: decomposing the auxiliary variables into multiple independent modes through modal decomposition, using the main branch and the context branch to extract the global features and local features of the multiple independent modes respectively, and using layer normalization and linear projection to combine the global features and local features obtained by the two branches to obtain the optimized auxiliary variables; the main branch uses multiple linear transformations, each linear transformation with layer normalization and the SiLU activation function, and maintains feature consistency by adjusting the dimension of the linear layer; the context branch uses small convolutional kernels for spatial processing and adds attention fusion.

[0047] Specifically, by introducing the auxiliary variable Z to achieve the separability of the objective function, its core convex optimization problem can be expressed as:

[0048] .

[0049] By introducing the Lagrange multiplier U and the penalty parameter, the augmented Lagrangian function, that is, the joint optimization objective function L, is defined as:

[0050] .

[0051] Among them, is the multi-scale feature, is the learnable dictionary matrix, is the sparse coefficient matrix, is the auxiliary variable, is the implicit prior regularization term for the hyperspectral image, is the Lagrange multiplier, and are both trade-off parameters. The dictionary matrix is set as a trainable parameter to implicitly learn the dynamic dictionary in an end-to-end manner.

[0052] By setting the dictionary D as a learnable parameter and alternately iteratively minimizing the sub-problems of the objective function L with respect to the variables S, Z, U within the ADMM framework while keeping other variables fixed, the problem is effectively solved.

[0053] For the optimization of the auxiliary variable Z, a dual-branch network based on the Mamba state space model is designed to replace the traditional soft thresholding algorithm, and the prior characteristics of the hyperspectral image are used for data-driven optimization. The optimization process first decomposes the input features into multiple independent modes through modal decomposition, and then combines the dual-branch mechanism to refine and fuse the global and local information of the decomposed features, and finally generates the auxiliary variable Z that satisfies the sparse constraint. Among them, the dual-branch processing mechanism includes the main branch and the context branch, which are used to extract the global features and local spatial features of the auxiliary variable Z respectively, and effectively integrate the features of the two branches through layer normalization and linear projection operations; the main branch uses multiple linear transformations, each linear transformation with layer normalization and SiLU activation function, and maintains the feature consistency by adjusting the dimension of the linear layer; the context branch uses a small convolution kernel ( convolution kernel) for spatial processing and adds attention fusion.

[0054] The detailed content of the main branch includes: using a multi-layer linear transformation sequence to model the global features. Each layer of transformation is equipped with layer normalization, SiLU activation function, and Dropout regularization module. By dynamically adjusting the dimensions of the linear layer, the feature consistency is ensured. The input feature z_combined, which is the combined feature processed by Mamba, is reshaped into [batch, height*width, channels] to adapt to the requirements of linear transformation. The processing flow includes dimension elevation, feature normalization, non-linear activation, and dimension reduction, and finally outputs the feature branch1out with the same dimension as the input. Specifically, the main branch performs feature reconstruction and enhancement of the expression ability on z_combined, expands the feature dimension through the dimension elevation operation, uses LayerNorm to maintain the stability of the feature distribution, the SiLU activation function introduces non-linear expression, and the Dropout mechanism improves the robustness of the model, thereby outputting the feature branch1out with high consistency and optimized expression ability.

[0055] The detailed content of the context branch includes: extracting local spatial features from the input feature z_combined through a 3×3 convolutional kernel, combining batch normalization (BatchNorm) to stabilize the training process, and achieving multi-scale feature fusion through the GlobalLocalAttention mechanism. The input feature is reshaped into [batch, channels, height, width] to adapt to the convolutional operation. The processing flow includes local spatial feature extraction, global pooling, and attention fusion of local convolution, and finally outputs the feature branch2out. Specifically, the context branch captures local spatial correlations through convolutional operations, and the GlobalLocalAttention module integrates global context and local detail information, significantly enhancing the spatial representation ability and multi-scale context awareness ability of the features.

[0056] The final feature fusion includes: concatenating branch1out output by the main branch and branch2out output by the context branch along the channel dimension to form a comprehensive feature tensor. The fused feature is then optimized through a multi-layer processing sequence, including layer normalization, linear dimensionality reduction, SiLU non-linear activation, and final linear projection, to generate the optimized auxiliary variable Z.

[0057] Step 3: Input the multi-scale features into the joint optimization model for variable iterative solution to obtain sparse coefficient features and dictionary features; concatenate the sparse coefficient features and dictionary features to obtain the fused feature; input the fused feature into the classifier to obtain the classification result of the hyperspectral image. Among them, the sparse coefficient features and dictionary features obtained during the optimization process are concatenated along the channel dimension.

[0058] Step 1: Data Acquisition and Preprocessing.

[0059] 1.1 Obtain a hyperspectral image dataset containing K ground object categories , with class labels . Denote the -th sample, C as the spectral dimension (number of spectral channels), and H×W as the spatial dimension, representing height and width respectively.

[0060] 1.2 Spatial-Spectral Joint Feature Extraction:

[0061] The feature extraction module constructs a three-layer stacked 3D convolutional network for extracting spatial-spectral joint features. The input data dimension is (B, C, H, W), where B is the batch size, C is the number of spectral channels, and H and W are the spatial dimensions. By expanding the dimension to (B, 1, C, H, W), the network successively uses convolutional kernels of (7, 3, 3), (5, 3, 3), and (3, 3, 3), followed by a BatchNorm normalization layer and a ReLU non-linear activation function after each layer. The number of channels changes to 8, 16, 32 in sequence, and the features after three layers of 3D convolution are reshaped into the form of (B, 32, H, W). Subsequently, weighted enhancement is performed through a channel attention mechanism, and then a three-layer 2D convolutional network is used to achieve multi-scale spatial feature extraction. The number of channels is configured as 32→128→256→512, and the maximum pooling operation is used to achieve hierarchical transformation of the feature map size, effectively capturing multi-scale spatial structure information, as Figure 2 shown.

[0062] Step 2: Construct a Multi-Constraint Optimization Framework.

[0063] 2.1 The solver based on ADMM introduces an auxiliary variable Z to achieve the separability of the objective function. Its core convex optimization problem can be expressed as:

[0064] .

[0065] By introducing the Lagrange multiplier U and the penalty parameter, the augmented Lagrangian function, that is, the joint optimization objective function L, is defined as:

[0066] .

[0067] Where: is the reconstructed hyperspectral data, that is, the multi-scale features after feature extraction using the 3D convolutional network and the 2D convolutional network, is the learnable dictionary matrix, is the sparse coefficient matrix, is the auxiliary variable, is the implicit prior regularization term for hyperspectral images, is the Lagrange multiplier, is the learnable trade-off parameter.

[0068] 2.2 Coefficient Optimization Design:

[0069] Through matrix operations, the closed-form solution of the sub-problem regarding the coefficient matrix is obtained: . Calculate the transpose of the dictionary and multiply it with , then replace the matrix inversion with a 3×3 trainable convolution to obtain , and then calculate the updated coefficient S in sequence. First, calculate the product of the transpose of the dictionary and the dictionary D, and then use a 3×3 trainable convolution kernel to approximately simulate the inverse operation of to generate an intermediate result, and iteratively update it in sequence to obtain the optimized sparse coefficient S. At the same time, the dictionary is set as the trainable parameter of . The feature distribution of the dictionary is stabilized through Layer Normalization, and then it is input into a two-layer fully connected network to map it into a 128-dimensional feature vector for enhancing the classification performance.

[0070] 2.3 Mamba Auxiliary Variable Optimization:

[0071] Use the Mamba state space model to replace traditional algorithms (such as soft thresholding), construct a hyperspectral image prior network, and optimize the Z variable through a data-driven adaptive learning strategy. For the initial sequence , through the modal decomposition strategy, is converted into unfolded sequences in three main directions. The orthogonal projection matrix is used to decompose the input tensor along the three main modes: the spectral mode , the spatial height mode and the spatial width mode . Each modal component is processed by an enhanced Mamba module (Enhanced Mamba Block), which integrates selective state parameterization and adaptive time scale adjustment. The state transition mechanism calculates its state parameters through a learnable projection matrix: , where is the weight matrix, is the bias vector. The time scale is modulated by the softplus activation function : , and the state transition matrix is derived through an exponential transformation . The state update follows the recursive formula , where represents the Hadamard product, Adaptive information flow at different time scales is achieved.

[0072] Among them, represents the input feature tensor of the hyperspectral image data at time step t , which is a certain modal component obtained after modal decomposition of the initial sequence ( Z 1, Z 2, Z 3) at the current time step. represents the discrete time interval parameter at time step t , which is used to adjust the dynamic characteristics of the state transition matrix A t and is calculated through the learnable parameter log δ , modulated by the softplus activation function. represents the hidden state at time step , which is the output of the Mamba state space model at the previous moment.

[0073] The optimization process adopts a dual-branch processing mechanism: the main branch uses a sequence of linear transformations with layer normalization LayerNorm and SiLU activation, and maintains feature consistency through dimensional configuration; the context branch uses convolution kernels for spatial processing to implement the GlobalLocalAttention mechanism , combines the global feature and the local feature , and obtains through attention fusion. The final output is obtained through the fusion module : , which uses layer normalization and linear projection to combine the features of the two branches: .

[0074] 2.4 Lagrange multiplier update:

[0075] The Lagrange multiplier is iteratively updated , where ρ is the step size parameter, and S and Z are the solutions of the primal problem and the dual problem respectively. Each iteration adjusts U to coordinate the consistency of S and Z.

[0076] 2.5 Dictionary matrix update:

[0077] The construction method of the implicit dynamic dictionary D is as follows: dictionary atoms are expressed by the independently learnable parameters of the neural network, which are set as the learnable network parameters, and the implicit dictionary learning is realized through end-to-end training; during the backpropagation process, the dictionary coherence loss and the classification loss are jointly optimized to realize task-oriented atom generation.

[0078] Step 3: Classification decision implementation.

[0079] 3.1 Feature fusion and classifier:

[0080] First, the sparse coefficient feature F_sparse obtained by the ADMM network and the dictionary feature F_dict are concatenated in the channel dimension to obtain the fused feature F_combined. The classifier uses a two-layer fully connected network. The first layer uses the ReLU activation function, and the second layer outputs the probability distribution of each category, and introduces learnable category weights for balancing. To optimize the model training process, a learning rate scheduling strategy combining warmup and cosine annealing is adopted: in the first 10 warmup rounds, the learning rate linearly increases from 0 to the initial value of 0.001; in the subsequent 90 rounds of training, the learning rate gradually decays according to the cosine function to achieve better optimization effects, as Figure 3 shown.

[0081] 3.2 Loss function:

[0082] The loss function consists of two components : the classification cross-entropy loss , where K is the number of categories, is the true label, is the predicted probability. The dictionary coherence loss , where is the th normalized dictionary feature vector, is the Kronecker function, is the number of dictionary features. The two losses are combined through the learnable weights w1 and w2, which are initialized with w1 = 1.0 and w2 = 0.1 to achieve adaptive weight adjustment.

[0083] The present invention also has the following effects:

[0084] By expanding the traditional ADMM iteration process into a trainable network layer and parameterizing the dictionary matrix as the trainable weights of the network, end-to-end implicit dictionary learning is realized. This design breaks through the limitations of traditional explicit dictionary construction methods and forms a dual-effect collaborative mechanism of data-driven regularization terms and model-driven optimization.

[0085] The embodiment of the present invention provides a classification system for hyperspectral images, including:

[0086] A data acquisition module, which is used to acquire a hyperspectral image dataset containing multiple ground object categories, and sequentially use a 3D convolutional network and a 2D convolutional network for feature extraction to obtain multi-scale features.

[0087] A joint optimization model construction module, which is used to construct a joint optimization objective function for hyperspectral image classification using the Alternating Direction Method of Multipliers (ADMM) to perform variable iterative solution as the joint optimization model; use a dual-branch processing mechanism based on the State Space Model Mamba to optimize the auxiliary variables in the joint optimization model; among them, the dual-branch processing mechanism based on the State Space Model Mamba includes: decomposing the auxiliary variables into multiple independent modes through modal decomposition, using the main branch and the context branch to extract the global features and local features of multiple independent modes respectively, and using layer normalization and linear projection to combine the global features and local features obtained by the two branches to obtain optimized auxiliary variables; the main branch uses multiple linear transformations, each linear transformation with layer normalization and SiLU activation function, and maintains feature consistency by adjusting the dimension of the linear layer; the context branch uses small convolutional kernels for spatial processing and adds attention fusion.

[0088] A classification result output module, which is used to input the multi-scale features into the joint optimization model for variable iterative solution to obtain sparse coefficient features and dictionary features; splice the sparse coefficient features and dictionary features to obtain fused features; input the fused features into a classifier to obtain the classification result of the hyperspectral image.

[0089] A specific embodiment is as follows:

[0090] This embodiment discloses a method for classifying hyperspectral images, and the specific steps include the following:

[0091] S1. Use the hyperspectral image obtained by an airborne visible / infrared imaging spectrometer (AVIRIS) at a certain test site. The spatial resolution of this hyperspectral image is 20 meters, and the original data contains 224 bands. After removing the low-quality and low signal-to-noise bands caused by strip noise and water absorption phenomena, 200 bands are retained for classification. The image covers 16 different land cover categories such as corn, soybeans, wheat and other crops, forests, buildings, grasslands, etc. The data size of the experimental area is 145×145 pixels, and there are 10,366 pixel points in the groundTruth.

[0092] S2. Use the joint optimization objective function and classifier for the processed image to obtain the classification result. The classification result is as Figure 4 , and the experimental results show that the overall accuracy reaches 98.10%.

[0093] The above-described embodiments merely represent several implementation manners of the present invention. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all fall within the protection scope of the present invention. Therefore, the protection scope of the present invention patent shall be subject to the appended claims.

Claims

1. A method for classifying hyperspectral images, characterized in that: The following steps are involved: Obtain a hyperspectral image dataset containing multiple ground object categories, and use 3D convolutional networks and 2D convolutional networks to extract features in turn to obtain multi-scale features; The alternating direction multiplier method (ADMM) is used to construct a joint optimization objective function for hyperspectral image classification and iterate the variables as a joint optimization model. The joint optimization objective function L is formulated as follows: ； in, is a multi-scale feature, is the dictionary matrix, is a sparse coefficient matrix, is an auxiliary variable, is an implicit prior regularization term for hyperspectral images, is the Lagrange multiplier, and All are trade-off parameters; A dual-branch processing mechanism based on the state-space model Mamba is used to optimize the auxiliary variables in the joint optimization model; wherein the dual-branch processing mechanism based on the state-space model Mamba includes: decomposing the auxiliary variables into multiple independent modes through modal decomposition, using the main branch and the context branch to extract the global features and local features of the multiple independent modes respectively, and using layer normalization and linear projection to combine the global features and local features obtained by the two branches to obtain optimized auxiliary variables; the main branch uses multiple linear transformations, each linear transformation has layer normalization and SiLU activation function, and the feature consistency is maintained by adjusting the dimension of the linear layer; the context branch uses a small convolution kernel for spatial processing, and adds attention fusion; The method of optimizing auxiliary variables in the joint optimization model using a dual-branch processing mechanism based on the state space model Mamba includes: The state space model Mamba is used to replace the solution method of the auxiliary variable Z in ADMM; The orthogonal projection matrix is ​​used to map the auxiliary variables along the spectral mode, spatial height mode and spatial width mode. Decomposition is performed to generate orthogonal multimodal feature components; each modal component is processed by the Enhanced Mamba Block, which integrates selective state parameterization and adaptive time scale adjustment; The main branch uses a sequence of linear transformations with layer normalization and SiLU activation functions to obtain auxiliary variables by adjusting the dimensions of the linear layer to maintain feature consistency. The global characteristics of The context branch uses a small convolution kernel for spatial processing and adds attention fusion to obtain auxiliary variables Local features of Use layer normalization and linear projection to fuse the features obtained from the two branches to complete the auxiliary variables Optimization; The multi-scale features are input into the joint optimization model for variable iterative solution to obtain sparse coefficient features and dictionary features; the sparse coefficient features and dictionary features are spliced ​​to obtain fusion features; the fusion features are input into the classifier to obtain the classification results of the hyperspectral image.

2. A method for classifying hyperspectral images as claimed in claim 1, characterized in that: Before obtaining the sparse coefficient feature, the method further includes: Get dictionary matrix Transpose and with the dictionary matrix Multiply to obtain the multiplication result; According to the multiplication result, the sparse coefficient matrix in the joint optimization model is completed by replacing the matrix inversion method with a 3×3 convolution. Optimization.

3. A method for classifying hyperspectral images as claimed in claim 1, characterized in that: The step of obtaining multi-scale features specifically includes: The dimension of the hyperspectral image dataset is (B, C, H, W), where B is the batch size, C is the number of spectral channels, and H and W are spatial dimensions; it is converted to (B, 1, C, H, W) through dimension expansion; The dimension-converted hyperspectral image dataset is input into a 3D convolutional network for feature extraction to obtain spatial and spectral information. The 3D convolutional network uses convolution kernels of (7, 3, 3), (5, 3, 3), and (3, 3, 3), and uses a BatchNorm normalization layer and a ReLU nonlinear activation function. The number of channels changes to 8, 16, and 32 respectively. The spatial information and spectral information are enhanced using the channel attention mechanism to obtain enhanced features; The enhanced features are input into a 2D convolutional network for multi-scale spatial feature extraction to obtain spatial information. The number of channels of the 2D convolutional network is 32, 128, 256 and 512, and the maximum pooling operation is used to perform hierarchical transformation of the feature map size.

4. A method for classifying hyperspectral images as claimed in claim 1, characterized in that: The classifier adopts a two-layer fully connected network.

5. The method for classifying a hyperspectral image according to claim 1, wherein: The classifier is trained by a learning rate scheduling strategy combining warmup, a training method capable of gradually increasing the learning rate, and cosine annealing.

6. A method for classifying hyperspectral images as claimed in claim 1, characterized in that: The loss function of the classifier L total , the formula is: ； in, is the classification cross entropy loss function, is the dictionary coherence loss function, w 1 and w 2 are all adjustable weight parameters.

7. A hyperspectral image classification system, characterized in that: include: The data acquisition module is used to acquire a hyperspectral image dataset containing multiple ground object categories, and sequentially use a 3D convolutional network and a 2D convolutional network to extract features and obtain multi-scale features; The joint optimization model building module is used to use the alternating direction multiplier method ADMM to build a joint optimization objective function for hyperspectral image classification to perform variable iterative solution as a joint optimization model; wherein the joint optimization objective function L is formulated as: ;in, is a multi-scale feature, is the dictionary matrix, is a sparse coefficient matrix, is an auxiliary variable, is an implicit prior regularization term for hyperspectral images, is the Lagrange multiplier, and are all trade-off parameters; a dual-branch processing mechanism based on the state-space model Mamba is used to optimize the auxiliary variables in the joint optimization model; wherein, the dual-branch processing mechanism based on the state-space model Mamba includes: decomposing the auxiliary variables into multiple independent modes through modal decomposition, using the main branch and the context branch to extract the global features and local features of the multiple independent modes respectively, and using layer normalization and linear projection to combine the global features and local features obtained by the two branches to obtain optimized auxiliary variables; the main branch uses multiple linear transformations, each linear transformation has layer normalization and SiLU activation function, and the feature consistency is maintained by adjusting the dimension of the linear layer; the context branch uses a small convolution kernel for spatial processing, and adds attention fusion; wherein, the dual-branch processing mechanism based on the state-space model Mamba is used to optimize the auxiliary variables in the joint optimization model, including: using the state-space model Mamba to replace the solution method of the auxiliary variable Z in ADMM; using an orthogonal projection matrix to transform the auxiliary variables along the spectral mode, spatial height mode and spatial width mode. Decompose to generate orthogonal multimodal feature components; each modal component is processed by the Enhanced MambaBlock module that integrates selective state parameterization and adaptive time scale adjustment; the main branch uses a linear transformation sequence with layer normalization and SiLU activation function to maintain feature consistency by adjusting the dimension of the linear layer to obtain auxiliary variables The context branch uses a small convolution kernel for spatial processing and adds attention fusion to obtain auxiliary variables. The local features of the two branches are fused using layer normalization and linear projection to complete the auxiliary variables Optimization; The classification result output module is used to input the multi-scale features into the joint optimization model for variable iterative solution to obtain sparse coefficient features and dictionary features; splice the sparse coefficient features with the dictionary features to obtain fusion features; input the fusion features into the classifier to obtain the classification results of the hyperspectral image.

Images