Image classification method and system based on multi-branch grouping mamba network

Abstract

The application belongs to the technical field of remote sensing image processing and deep learning, and particularly relates to an image classification method and system based on a multi-branch grouping Mamba network, comprising a data preparation and preprocessing module, an initial feature embedding module, a multi-domain parallel feature extraction module, an adaptive multi-domain fusion module and a classification prediction module. The technical scheme realizes high-precision hyperspectral feature classification while maintaining a light parameter quantity by coordinating the discriminative information of three complementary domains of space, spectrum and frequency, and is particularly suitable for remote sensing scenes with limited labeled samples, high spectral similarity and complex spatial texture.

Description

Technical Field

[0001] This application belongs to the field of remote sensing image processing and deep learning technology, specifically relating to an image classification method and system based on a multi-branch grouped Mamba network. Background Technology

[0002] The inherent characteristics of hyperspectral data, such as high dimensionality, scarce labeled samples, strong noise, and strong correlations between bands, make classification tasks extremely challenging. Convolutional neural networks (CNNs) are limited by fixed local receptive fields, making it difficult to model long-range dependencies; Transformers achieve global modeling through self-attention, but their O(N²) computational complexity is unbearable for high-dimensional hyperspectral data; the Mamba architecture, based on a selectively structured state-space model, achieves global sequence modeling with linear complexity and has become a research hotspot in the field of hyperspectral classification. However, existing Mamba methods generally suffer from insufficient handling of spectral redundancy and weak fusion of heterogeneous features.

[0003] Furthermore, land cover materials are often only effectively distinguishable within specific representation domains: spectrally similar categories may be distinguished by spatial texture, spatially similar categories by spectral features, while categories easily confused in both spatial and spectral domains may require discrimination based on periodic patterns in the frequency domain. Existing Mamba methods are still limited to the spatial and spectral domains, neglecting the global discriminative power of periodicity in the frequency domain. Therefore, effective classification requires extracting features from three complementary domains: spatial, spectral, and frequency. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides an image classification method and system based on a multi-branch grouped Mamba network. The aim is to achieve high-precision hyperspectral land cover classification while maintaining a lightweight parameter set by coordinating the discriminative information of three positively complementary representation domains: spatial, spectral, and frequency.

[0005] An image classification method based on a multi-branch grouped Mamba network includes the following steps: S1: Data preparation and preprocessing: Read the hyperspectral image, extract a fixed-size neighborhood 3D patch centered on each labeled pixel as the input sample, perform standardization processing, and then divide it into training set and test set; S2: Initial Feature Embedding: The input map is mapped to the latent feature space through convolutional layers to achieve preliminary information aggregation across spectral channels; S3: Multi-domain parallel feature extraction: Embedded features are fed into three parallel branches simultaneously: Spatial grouping Mamba extractor PGM is responsible for capturing spatial context features, bidirectional spectral Mamba extractor BCM is responsible for modeling long-distance dependencies along the spectral dimension, and Fourier-based feature extractor FFE is responsible for extracting global texture and periodic information from the frequency domain. The three branches output spatial feature maps, spectral feature maps and frequency domain feature maps respectively. S4: Adaptive multi-domain fusion: The feature maps of the three domains are summed and fed into a lightweight weight generation network to generate adaptive fusion weights for each pixel. After thresholding to filter out low-contribution domains, the weighted sums are performed to obtain the fused features. S5: Classification prediction: The fused features are mapped to the class space through a fully connected layer, and the class with the highest probability is taken as the classification label of the pixel; S6: The cross-entropy loss function is used to calculate the loss between the predicted value and the true label. All network parameters are updated through backpropagation using the AdamW optimizer, and training is performed iteratively until convergence.

[0006] Preferably, in the spatial grouping Mamba extractor PGM, an adjacent group selective state space model AGSM is designed: the embedded features are continuously divided into multiple non-overlapping groups along the spectral dimension, each group contains several continuous spectral channels, and local spectral correlation is maintained; each group of features is independently fed into a unidirectional selective state space model (S module), and a different spatial scanning direction is assigned to each group. The state space evolution of each group is performed in parallel. Finally, the outputs of each group are reshaped and spliced ​​along the feature dimension to obtain a spatial feature map.

[0007] Preferably, the adjacent group selective state-space model AGSM implements a sequence-aware parameterization mechanism, enabling the projection matrix to be dynamically adjusted according to the input sequence content, thereby achieving interactive learning between sequence states. Its parameterization process is described as follows: ; ; Among them, parameters and All are input Direct mapping generation; b represents the batch size, L represents the sequence length, D represents the feature dimension of each sequence element; N represents the hidden state dimension of the state space; Let these represent linear mapping functions with output dimensions D and N, respectively. This represents the input-related discretization step size parameter, used to adjust the time scale during the discretization process of the unidirectional selective state-space model, thereby enabling the state update process to adapt to changes in the input content.

[0008] Preferably, the adjacent grouping selective state-space model AGSM includes: Adjacent groups: Given an embedding feature cube It is continuously divided into G non-overlapping groups along the spectral dimension D, and the division method is as follows: ; in For feature cube The One channel, each group Include A continuous spectral channel, preserving local spectral consistency between adjacent bands; Evolution of the unidirectional selective state-space model: Each group of sequences is fed into an independent unidirectional selective state-space model, and each group uses an independent set of trainable parameters; taking the first group as an example... Taking a group as an example, its state-space evolution process is as follows: ; ; in , , and For the first Independent trainable parameters specific to a group. Let t be the implicit latent state at time t; the above state space evolution process, combined with the independent parameters of the k-th group and the input-related selection mechanism, enables each feature group to adaptively update its state according to its own sequence content.

[0009] Feature Reconstruction: The remaining G-1 feature groups perform the same flattening-scanning-evolution process in parallel, with each group using independent unidirectional selective state-space model parameters and assigned different scan directions; upon completion, each group outputs a sequence. Reconstruct them back to their original two-dimensional spatial structures. Then, the features are stitched together along the feature dimensions to reconstruct the final spatial feature map. .

[0010] Preferably, the bidirectional spectral Mamba extractor (BCM) focuses on modeling global semantic dependencies along the spectral dimension, including: Sequence construction: embedding features of Flattening the spatial dimensions yields the feature matrix. Transpose the matrix as The sequence length is D, and the feature dimension of each sequence element is . ; Bidirectional scanning: forward scanning That is, scanning sequentially from band 1 to band D; reverse scanning. The sequence is reversed and scanned from band D to band 1; the two sequences are respectively fed into two independent unidirectional selective state-space models. The forward scan state space evolution process is as follows: ; ; Reverse scanning uses independent parameter sets ( , , , Processing results ; Two-way information fusion: The output sequence of the reverse scan is flipped to the original band order and added element-wise to the output of the forward scan. ; fused sequence The original spatial cube structure is restored through transpose and reshaping operations, generating the final spectral feature map. .

[0011] Preferably, the Fourier-based feature extractor (FFE) comprises two parallel sub-branches: the spatial frequency learning branch performs a Fourier transform on the input features along the spatial dimension, uses low-pass filtering to retain the main frequency components, performs frequency modulation in the frequency domain using learnable complex weights, and then converts them back to the spatial domain via an inverse Fourier transform; the spectral frequency learning branch performs a Fourier transform along the spectral dimension, retains all frequency components, and restores them after complex domain modulation; the outputs of the two branches are adaptively fused using learnable weights to obtain a frequency domain feature map; the Fourier-based feature extractor (FFE) extracts spatial texture periodicity and spectral harmonic patterns from the frequency domain, providing a global discriminative perspective that is positively complementary to the spatial-spectral domain.

[0012] Preferably, the Fourier-based feature extractor (FFE) comprises two parallel sub-branches, learning frequency representations along the spatial and spectral dimensions, respectively; for features from the initial embedding layer... First, flatten it into , Then, layer normalization is applied, and the normalized features are fed into two branches; Spatial frequency learning branch given input features: First, a Fast Fourier Transform is applied along the spatial dimension N to map the features to the spatial frequency domain. The spatial branch employs a low-pass filtering strategy, selecting only m low-frequency modes for subsequent processing, which can be expressed as: ; Subsequently, frequency modulation is achieved by performing complex multiplication on these selected frequency modes with learnable complex weights R: ; in The complex weight matrix is ​​a learnable matrix. This represents the Hadamard product; the modulated features are padded with zeros to restore the original frequency dimension. Then, the inverse FFT is used to transform it back to the spatial domain to obtain the spatially filtered features; Spectral frequency learning branch: Perform FFT on the input features along the spectral dimension D: ; Unlike the spatial branch, this branch preserves all frequency components to capture complete spectral harmonic information; frequency modulation is achieved through complex multiplication and learnable weights. ; in The complex weight matrix is ​​a learnable matrix, which is modulated and then transformed back to the spatial domain by inverse FFT; Two-branch adaptive fusion: The outputs of the two branches are passed through two learnable scalar weights. and Adaptive fusion is performed to generate the final frequency domain feature output. .

[0013] Preferably, the feature maps of the spatial, spectral, and frequency domains are first summed to obtain aggregated features, which are then fed into a lightweight weight generation network. This network outputs the fusion weights of the three domains positionally for each pixel. A renormalization mechanism is introduced to suppress low-contribution domains and enhance high-contribution domains. Finally, the fusion features are obtained through weighted summation. The specific method is as follows: First, global context information is aggregated by summing the output feature maps of the three branches: ; Then, the aggregation feature The data is fed into a lightweight weight generation network and mapped to a three-channel weight vector. : ; The three channels correspond to the spatial domain weights respectively. Spectral domain weights and frequency domain weights The weight normalization mechanism utilizes the normalized weights. , , The corresponding feature maps are then weighted and summed to obtain the final fused features. : ; in, This represents the Hadamard product.

[0014] Preferably, it includes a data preparation and preprocessing module, an initial feature embedding module, a multi-domain parallel feature extraction module, an adaptive multi-domain fusion module, and a classification prediction module; Data preparation and preprocessing module: Read the hyperspectral image, extract a fixed-size neighborhood 3D patch centered on each labeled pixel as the input sample, perform standardization processing, and divide it into training set and test set; Initial feature embedding module: Maps the input map patches to the latent feature space through convolutional layers, achieving preliminary information aggregation across spectral channels; The multi-domain parallel feature extraction module includes a spatial grouping Mamba extractor (PGM) module, a bidirectional spectral Mamba extractor (BCM) module, and a Fourier-based feature extractor (FFE) module. The spatial grouping Mamba extractor PGM module extracts spatial features through adjacent grouping and unidirectional selective state-space model scanning; the bidirectional spectral Mamba extractor BCM module extracts spectral features through bidirectional spectral unidirectional selective state-space model scanning; the Fourier-based feature extractor FFE module extracts frequency domain features through Fourier transforms in both spatial and spectral dimensions; the three branches are executed in complete parallel, each maintaining the independence of its representation domain. The adaptive multi-domain fusion module sums the feature maps of the spatial, spectral, and frequency domains to obtain aggregated features, which are then fed into a lightweight weight generation network to output the fusion weights of the three domains positionally for each pixel. A normalization mechanism is introduced to suppress low-contribution domains and enhance high-contribution domains, and finally, the fusion features are obtained by weighted summation. Classification prediction module: The fused features are mapped to the category space through a fully connected layer, and the category with the highest probability is taken as the classification label of the pixel.

[0015] Preferably, the spatial grouping Mamba extractor PGM designs an adjacent grouping selective state space model: along the spectral dimension, the embedded features are continuously divided into multiple non-overlapping groups, each group contains several continuous spectral channels, maintaining local spectral correlation, each group of features is independently fed into a unidirectional selective state space model, and a different spatial scanning direction is assigned to each group. The state space evolution of each group is performed in parallel, and the outputs of each group are reshaped and spliced ​​along the feature dimension to obtain a spatial feature map.

[0016] Compared with the prior art, the beneficial effects of this application are as follows:

[0017] This invention achieves significant results in dynamic task allocation scenarios by introducing a collaborative optimization mechanism of reinforcement learning and adaptive genetic algorithms. Firstly, the global scheduling strategy based on reinforcement learning can achieve more accurate task allocation according to workstation load, task characteristics, and subtask dependencies, significantly improving decision-making stability in dynamic environments.

[0018] Secondly, by using an adaptive genetic algorithm to locally optimize the task sequence at the target workstation, the blocking of high-priority tasks is effectively reduced, resulting in a more rational task processing order. Furthermore, this invention introduces discretization and subtask dependency constraints into the state construction, enabling the scheduling model to accurately identify the decomposition order requirements, avoid scheduling conflicts, and reduce the average task waiting time by approximately 15%. In summary, this invention significantly improves the operating efficiency and resource utilization of the decomposition system, demonstrating good engineering application value. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the system architecture; Figure 2 A schematic diagram of the AGSM (Adjacent Group Selective State Space Model); Figure 3 This is a graph showing the calculation efficiency. Detailed Implementation

[0020] An image classification method based on a multi-branch grouped Mamba network includes the following steps: S1: Data preparation and preprocessing: Read the hyperspectral image, extract a fixed-size neighborhood 3D patch centered on each labeled pixel as the input sample, perform standardization processing, and then divide it into training set and test set; S2: Initial Feature Embedding: The input map is mapped to the latent feature space through convolutional layers to achieve preliminary information aggregation across spectral channels;

[0021] S3: Multi-domain parallel feature extraction: Embedded features are fed into three parallel branches simultaneously: the spatial grouping Mamba extractor PGM is responsible for capturing spatial context features, the bidirectional spectral Mamba extractor BCM is responsible for modeling long-distance dependencies along the spectral dimension, and the Fourier-based feature extractor FFE is responsible for extracting global texture and periodic information from the frequency domain. The three branches output spatial feature maps, spectral feature maps, and frequency domain feature maps, respectively.

[0022] In the spatial grouping Mamba extractor PGM, a selective state-space mechanism AGSM for adjacent groups is designed: the embedded features are continuously divided into multiple non-overlapping groups along the spectral dimension, each group contains several continuous spectral channels, and local spectral correlation is maintained; each group of features is independently fed into a unidirectional selective state-space model (S module), and a different spatial scanning direction is assigned to each group. The state-space evolution of each group is performed in parallel. Finally, the outputs of each group are reshaped and spliced ​​along the feature dimension to obtain a spatial feature map.

[0023] The Adjacent Group Selective State Space Model (AGSM) implements a sequence-aware parameterization mechanism, enabling the projection matrix to dynamically adjust according to the input sequence content, thereby achieving interactive learning between sequence states. Its parameterization process is described as follows: ; ; Among them, parameters and All are input Direct mapping generation; b represents the batch size, L represents the sequence length, D represents the feature dimension of each sequence element; N represents the hidden state dimension of the state space. Let these represent linear mapping functions with output dimensions D and N, respectively. This represents the input-related discretization step size parameter, used to adjust the time scale during the discretization process of the unidirectional selective state-space model, thereby enabling the state update process to adapt to changes in the input content.

[0024] Adjacent groups:

[0025] Given an embedding feature cube Along the spectral dimension It is continuously divided into G non-overlapping groups, and the division method is as follows: ; in For feature cube The One channel, each group Include A continuous spectral channel preserves local spectral consistency between adjacent bands.

[0026] Spatial flattening and serialization: To enable the unidirectional selective state-space model to process two-dimensional spatial data, each three-dimensional feature group needs to be flattened into a one-dimensional sequence. Taking the first group... For example, a spatial flattening operation is performed according to the specific spatial scanning direction assigned to the group to generate a one-dimensional sequence. Each element is The four groups are assigned four spatial scanning directions: scanning line by line from top to bottom, scanning line by line from bottom to top, scanning column by column from left to right, and scanning column by column from right to left.

[0027] Evolution of the unidirectional selective state-space model: Each group of sequences is fed into an independent unidirectional selective state-space model, and each group uses an independent set of trainable parameters; taking the first group as an example... Taking a group as an example, its state-space evolution process is as follows: ; ; in and For the first Independent trainable parameters specific to a group. Let t be the implicit latent state at time t; the above state space evolution process, combined with the independent parameters of the k-th group and the input-related selection mechanism, enables each feature group to adaptively update its state according to its own sequence content.

[0028] Feature Reconstruction: The remaining G-1 feature groups perform the same flattening-scanning-evolution process in parallel, with each group using independent unidirectional selective state-space model parameters and assigned different scan directions; upon completion, each group outputs a sequence. Reconstruct them back to their original two-dimensional spatial structures. Then, the features are stitched together along the feature dimensions to reconstruct the final spatial feature map. .

[0029] The Bidirectional Spectral Mamba Extractor (BCM) focuses on modeling global semantic dependencies along the spectral dimension, including: Sequence construction: embedding features of Flattening the spatial dimensions yields the feature matrix. Transpose the matrix as The sequence length is D, and the feature dimension of each sequence element is . ; Bidirectional scanning: forward scanning That is, scanning sequentially from band 1 to band D; reverse scanning. The sequence is reversed and scanned from band D to band 1; the two sequences are respectively fed into two independent unidirectional selective state-space models. The forward scan state space evolution process is as follows: ; ; Reverse scanning uses independent parameter sets Processed .

[0030] Two-way information fusion: The output sequence of the reverse scan is flipped to the original band order and added element-wise to the output of the forward scan. ; fused sequence The original spatial cube structure is restored through transpose and reshaping operations, generating the final spectral feature map. .

[0031] The Fourier-based Feature Extractor (FFE) comprises two parallel sub-branches: the spatial frequency learning branch performs a Fourier transform on the input features along the spatial dimension, uses low-pass filtering to retain the main frequency components, modulates the frequency in the frequency domain using learnable complex weights, and then converts it back to the spatial domain via an inverse Fourier transform; the spectral frequency learning branch performs a Fourier transform along the spectral dimension, retains all frequency components, modulates them in the complex domain, and then restores them; the outputs of the two branches are adaptively fused using learnable weights to obtain a frequency domain feature map; the Fourier-based Feature Extractor (FFE) extracts spatial texture periodicity and spectral harmonic patterns from the frequency domain, providing a global discriminative perspective that is positively complementary to the spatial-spectral domain.

[0032] The Fourier-based feature extractor (FFE) comprises two parallel sub-branches that learn frequency representations along the spatial and spectral dimensions, respectively; for features from the initial embedding layer... First, flatten it into , Then, layer normalization is applied, and the normalized features are fed into two branches;

[0033] Spatial frequency learning branch given input features: First, a Fast Fourier Transform is applied along the spatial dimension N to map the features to the spatial frequency domain. The spatial branch employs a low-pass filtering strategy, selecting only m low-frequency modes for subsequent processing, which can be expressed as: ; Subsequently, frequency modulation is achieved by performing complex multiplication on these selected frequency modes with learnable complex weights R: ; in The complex weight matrix is ​​a learnable matrix. This represents the Hadamard product; the modulated features are padded with zeros to restore the original frequency dimension. Then, the inverse FFT is used to transform it back to the spatial domain to obtain the spatially filtered features; Spectral frequency learning branch: Perform FFT on the input features along the spectral dimension D: .

[0034] Unlike the spatial branch, this branch preserves all frequency components to capture complete spectral harmonic information; frequency modulation is achieved through complex multiplication and learnable weights. ; in The complex weight matrix is ​​a learnable matrix, which is modulated and then transformed back to the spatial domain by inverse FFT.

[0035] Two-branch adaptive fusion: The outputs of the two branches are passed through two learnable scalar weights. and Adaptive fusion is performed to generate the final frequency domain feature output. .

[0036] S4: Adaptive Multi-Domain Fusion: The feature maps of the three domains are summed and fed into a lightweight weight generation network to generate adaptive fusion weights for each pixel. After thresholding to filter out low-contribution domains, the weighted sums are performed to obtain the fused features.

[0037] First, global context information is aggregated by summing the output feature maps of the three branches: ; Then, the aggregation feature The data is fed into a lightweight weight generation network and mapped to a three-channel weight vector. : ; The three channels correspond to the spatial domain weights respectively. Spectral domain weights and frequency domain weights The weight normalization mechanism utilizes the normalized weights. , , The corresponding feature maps are then weighted and summed to obtain the final fused features. : ; in, This represents the Hadamard product.

[0038] S5: Classification prediction: The fused features are mapped to the class space through a fully connected layer, and the class with the highest probability is taken as the classification label of the pixel.

[0039] S6: The cross-entropy loss function is used to calculate the loss between the predicted value and the true label. All network parameters are updated through backpropagation using the AdamW optimizer, and training is performed iteratively until convergence.

[0040] The resulting fused feature vector F_fused is finally used for class prediction through a fully connected (FC) layer.

[0041] To quantitatively evaluate the PTMamba model of this invention, three standard indicators are used: (1) Overall Accuracy (OA), which reflects the overall classification performance of the model, i.e., the proportion of correctly classified pixels to the total number of test pixels; (2) Average Accuracy (AA), which evaluates the average classification performance under class imbalance conditions, i.e., the arithmetic mean of the classification accuracy of each class, to avoid the model being biased towards the class with a larger proportion of pixels; (3) Kappa coefficient ( This measures the degree of agreement between the classification results and the ground truth after removing random consistency. A value closer to 1 indicates a higher consistency between the classification result and the true label. The three metrics comprehensively evaluate the classification model from three dimensions: overall performance, class fairness, and statistical consistency.

[0042] To fully validate the effectiveness of PTMamba, a systematic evaluation was conducted on four publicly available hyperspectral datasets: Indian Pines (IP), Pavia University (PU), Houston2013 (HUS), and WHU-Hi-LongKou (LK). A comprehensive comparison was made with CNN-based methods (1-D CNN, 2-D CNN, CLOLN), Transformer-based methods (SpectralFormer, morphFormer, SSFTT, GraphGST), and Mamba-based methods (SpiralMamba, DBMGNet, S2Mamba). The comparison results for overall accuracy (OA), average accuracy (AA), and Kappa coefficient (kappa) are shown in Table 1.

[0043] Table 1. Overall performance comparison of PTMamba and existing methods on four datasets.

[0044] As shown in Table 1, PTMamba achieved the best overall accuracy (OA) of 98.38%, 97.80%, 94.19%, and 99.13% on the four datasets, surpassing all other methods compared. The average accuracy (AA) and Kappa coefficient (κ) also reached the best levels on all datasets. Taking the LK dataset as an example, PTMamba achieved the best performance with OA of 99.13%, AA of 98.95%, and κ of 0.9885. This fully demonstrates the essential advantages of this invention in continuous spectral sequence modeling for precision agriculture.

[0045] Computational efficiency assessment

[0046] In addition to demonstrating excellent generalization capabilities across various complex remote sensing scenarios, this system also exhibits significant advantages in computational resource utilization. Taking the Indian Pines dataset as an example, for instance... Figure 3 As shown, CNN-type models, due to insufficient complexity to fully capture the high-dimensional features of hyperspectral data, cluster in the lower left quadrant, exhibiting a "low cost, low return" characteristic and limited classification accuracy. Transformer-type models show significant heterogeneity in efficiency and performance, revealing their challenge in balancing computational cost and classification accuracy: some methods achieve acceptable accuracy but have a huge number of parameters, while others have fewer parameters but experience a significant drop in accuracy. Mamba-type models diverge into two distinct branches: on the one hand, methods such as SpiralMamba and DBMGNet sacrifice lightweightness for high accuracy; on the other hand, PTMamba achieves the highest overall accuracy with the fewest parameters, reaching Pareto optimality among all compared models. Compared to other methods, PTMamba achieves higher accuracy with fewer parameters. This result strongly demonstrates that PTMamba does not simply trade accuracy for efficiency or efficiency for accuracy, but rather achieves the optimal balance between lightweight design and high-precision performance through an advanced multi-domain parallel independent extraction architecture and adjacent group scanning strategy.

Claims

1. An image classification method based on a multi-branch grouped Mamba network, characterized in that, Includes the following steps: S1: Data preparation and preprocessing: Read the hyperspectral image, extract a fixed-size neighborhood 3D patch centered on each labeled pixel as the input sample, perform standardization processing, and then divide it into training set and test set; S2: Initial Feature Embedding: The input map is mapped to the latent feature space through convolutional layers to achieve preliminary information aggregation across spectral channels; S3: Multi-domain parallel feature extraction: Embedded features are fed into three parallel branches simultaneously: Spatial grouping Mamba extractor PGM is responsible for capturing spatial context features, bidirectional spectral Mamba extractor BCM is responsible for modeling long-distance dependencies along the spectral dimension, and Fourier-based feature extractor FFE is responsible for extracting global texture and periodic information from the frequency domain. The three branches output spatial feature maps, spectral feature maps and frequency domain feature maps respectively. S4: Adaptive multi-domain fusion: The feature maps of the three domains are summed and fed into a lightweight weight generation network to generate adaptive fusion weights for each pixel. After thresholding to filter out low-contribution domains, the weighted sums are performed to obtain the fused features. S5: Classification prediction: The fused features are mapped to the class space through a fully connected layer, and the class with the highest probability is taken as the classification label of the pixel; S6: The cross-entropy loss function is used to calculate the loss between the predicted value and the true label. All network parameters are updated through backpropagation using the AdamW optimizer, and training is performed iteratively until convergence.

2. The image classification method based on a multi-branch grouped Mamba network according to claim 1, characterized in that, In the spatial grouping Mamba extractor PGM, an adjacent group selective state space model AGSM is designed: the embedded features are continuously divided into multiple non-overlapping groups along the spectral dimension, each group contains several continuous spectral channels, and local spectral correlation is maintained; each group of features is independently fed into a unidirectional selective state space model, and a different spatial scanning direction is assigned to each group. The state space evolution of each group is performed in parallel. Finally, the outputs of each group are reshaped and spliced ​​along the feature dimension to obtain a spatial feature map.

3. The image classification method based on a multi-branch grouped Mamba network according to claim 2, characterized in that, The Adjacent Group Selective State Space Model (AGSM) implements a sequence-aware parameterization mechanism, enabling the projection matrix to dynamically adjust according to the input sequence content, thereby achieving interactive learning between sequence states. Its parameterization process is described as follows: ； ； Among them, parameters and All are input Direct mapping generation; b represents the batch size, L represents the sequence length, D represents the feature dimension of each sequence element; N represents the hidden state dimension of the state space; Let these represent linear mapping functions with output dimensions D and N, respectively. This indicates the input-related discretization step size parameter.

4. The image classification method based on a multi-branch grouped Mamba network according to claim 3, characterized in that, Adjacent Group Selective State-Space Model (AGSM) includes: Adjacent groups: Given an embedding feature cube It is continuously divided into G non-overlapping groups along the spectral dimension D, and the division method is as follows: ； in For feature cube The One channel, each group Include A continuous spectral channel, preserving local spectral consistency between adjacent bands; Evolution of the unidirectional selective state-space model: Each group of sequences is fed into an independent unidirectional selective state-space model, and each group uses an independent set of trainable parameters; taking the first group as an example... Taking a group as an example, its state-space evolution process is as follows: ； ； in and For the first Independent trainable parameters specific to a group. Let be the implicit latent state at time t; the above state space evolution process, combined with the independent parameters of the k-th group and the input-related selection mechanism, enables each feature group to adaptively update its state according to its own sequence content. Feature Reconstruction: The remaining G-1 feature groups perform the same flattening-scanning-evolution process in parallel, with each group using independent unidirectional selective state-space model parameters and assigned different scan directions; upon completion, each group outputs a sequence. Reconstruct them back to their original two-dimensional spatial structures. Then, the features are stitched together along the feature dimensions to reconstruct the final spatial feature map. .

5. The image classification method based on a multi-branch grouped Mamba network according to claim 3, characterized in that, The Bidirectional Spectral Mamba Extractor (BCM) focuses on modeling global semantic dependencies along the spectral dimension, including: Sequence construction: embedding features of Flattening the spatial dimensions yields the feature matrix. Transpose the matrix as The sequence length is D, and the feature dimension of each sequence element is . ; Bidirectional scanning: forward scanning That is, scanning sequentially from band 1 to band D; reverse scanning. The sequence is reversed and scanned from band D to band 1; the two sequences are respectively fed into two independent unidirectional selective state-space models. The forward scan state space evolution process is as follows: ； ； Reverse scanning uses independent parameter sets ( , , , Processing results ; Two-way information fusion: The output sequence of the reverse scan is flipped to the original band order and added element-wise to the output of the forward scan. ； fused sequence The original spatial cube structure is restored through transpose and reshaping operations, generating the final spectral feature map. .

6. The image classification method based on a multi-branch grouped Mamba network according to claim 1, characterized in that, The Fourier-based Feature Extractor (FFE) comprises two parallel sub-branches: the spatial frequency learning branch performs a Fourier transform on the input features along the spatial dimension, uses low-pass filtering to retain the main frequency components, modulates the frequency in the frequency domain using learnable complex weights, and then converts it back to the spatial domain via an inverse Fourier transform; the spectral frequency learning branch performs a Fourier transform along the spectral dimension, retains all frequency components, modulates them in the complex domain, and then restores them; the outputs of the two branches are adaptively fused using learnable weights to obtain a frequency domain feature map; the Fourier-based Feature Extractor (FFE) extracts spatial texture periodicity and spectral harmonic patterns from the frequency domain, providing a global discriminative perspective that is positively complementary to the spatial-spectral domain.

7. The image classification method based on a multi-branch grouped Mamba network according to claim 6, characterized in that, The Fourier-based feature extractor (FFE) comprises two parallel sub-branches that learn frequency representations along the spatial and spectral dimensions, respectively; for features from the initial embedding layer... First, flatten it into Then, layer normalization is applied, and the normalized features are fed into two branches; Spatial frequency learning branch given input features: First, a Fast Fourier Transform is applied along the spatial dimension N to map the features to the spatial frequency domain. The spatial branch employs a low-pass filtering strategy, selecting only m low-frequency modes for subsequent processing, which can be expressed as: ； Subsequently, frequency modulation is achieved by performing complex multiplication on these selected frequency modes with learnable complex weights R: ； in The complex weight matrix is ​​a learnable matrix. This represents the Hadamard product; the modulated features are padded with zeros to restore the original frequency dimension. Then, the inverse FFT is used to transform it back to the spatial domain to obtain the spatially filtered features; Spectral frequency learning branch: Perform FFT on the input features along the spectral dimension D: ； Unlike the spatial branch, this branch preserves all frequency components to capture complete spectral harmonic information; frequency modulation is achieved through complex multiplication and learnable weights. ； in The complex weight matrix is ​​a learnable matrix, which is modulated and then transformed back to the spatial domain by inverse FFT; Two-branch adaptive fusion: The outputs of the two branches are passed through two learnable scalar weights. and Adaptive fusion is performed to generate the final frequency domain feature output. .

8. The image classification method based on a multi-branch grouped Mamba network according to claim 7, characterized in that, First, the feature maps of the spatial, spectral, and frequency domains are summed to obtain aggregated features, which are then fed into a lightweight weighted generation network. This network outputs the fusion weights of the three domains positionally for each pixel. A renormalization mechanism is introduced to suppress low-contribution domains and enhance high-contribution domains. Finally, the fusion features are obtained through weighted summation. The specific method is as follows: First, global context information is aggregated by summing the output feature maps of the three branches: ； Then, the aggregation feature The data is fed into a lightweight weight generation network and mapped to a three-channel weight vector. : ； The three channels correspond to the spatial domain weights respectively. Spectral domain weights and frequency domain weights The weight normalization mechanism utilizes the normalized weights. The corresponding feature maps are then weighted and summed to obtain the final fused features. : ； in, This represents the Hadamard product.

9. The image classification system based on a multi-branch grouped Mamba network according to claim 7, characterized in that, It includes a data preparation and preprocessing module, an initial feature embedding module, a multi-domain parallel feature extraction module, an adaptive multi-domain fusion module, and a classification prediction module; Data preparation and preprocessing module: Read the hyperspectral image, extract a fixed-size neighborhood 3D patch centered on each labeled pixel as the input sample, perform standardization processing, and divide it into training set and test set; Initial feature embedding module: Maps the input map patches to the latent feature space through convolutional layers, achieving preliminary information aggregation across spectral channels; The multi-domain parallel feature extraction module includes a spatial grouping Mamba extractor (PGM) module, a bidirectional spectral Mamba extractor (BCM) module, and a Fourier-based feature extractor (FFE) module. The spatial grouping Mamba extractor PGM module extracts spatial features through adjacent grouping and unidirectional selective state-space model scanning; the bidirectional spectral Mamba extractor BCM module extracts spectral features through bidirectional spectral unidirectional selective state-space model scanning; the Fourier-based feature extractor FFE module extracts frequency domain features through Fourier transforms in both spatial and spectral dimensions; the three branches are executed in complete parallel, each maintaining the independence of its representation domain. The adaptive multi-domain fusion module sums the feature maps of the spatial, spectral, and frequency domains to obtain aggregated features, which are then fed into a lightweight weight generation network to output the fusion weights of the three domains positionally for each pixel. A normalization mechanism is introduced to suppress low-contribution domains and enhance high-contribution domains, and finally, the fusion features are obtained by weighted summation. Classification prediction module: The fused features are mapped to the category space through a fully connected layer, and the category with the highest probability is taken as the classification label of the pixel.

10. The image classification system based on a multi-branch grouped Mamba network according to claim 9, characterized in that, Spatial Grouping Mamba Extractor PGM: Design of Adjacent Grouping Selective State Space Model: The embedded features are continuously divided into multiple non-overlapping groups along the spectral dimension. Each group contains several continuous spectral channels, maintaining local spectral correlation. Each group of features is independently fed into a unidirectional selective state space model, and a different spatial scanning direction is assigned to each group. The state space evolution of each group is performed in parallel. The outputs of each group are reshaped and then stitched along the feature dimension to obtain a spatial feature map.

Images