Hyperspectral image classification method based on spectral space decoupling of state space model

Abstract

The invention provides a hyperspectral image classification method based on spectral space decoupling of a state space model, and the method comprises the steps: carrying out the feature extraction of a to-be-classified hyperspectral image, and obtaining a shallow feature; inputting the shallow features into a spatial semantic modeling module based on a state space model to obtain global spatial semantic information output by the spatial semantic modeling module; inputting the shallow features into a spectrum sensing module based on a state space model to obtain global spectrum sensing information output by the spectrum sensing module; performing channel-level point product based on the global spatial semantic information and the global spectrum sensing information to obtain spatial spectrum coupling information; carrying out average pooling and maximum value indexing based on the spatial spectrum coupling information to obtain a classification result of the hyperspectral image to be classified; according to the invention, combined utilization of the spatial information and the spectral information can be realized to improve the classification accuracy of the hyperspectral image.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a hyperspectral image classification method based on spectral-spatial decoupling of a state-space model. Background Art

[0002] Hyperspectral images are three-dimensional data bodies composed of spatial information in the space dimension and spectral information in the channel dimension. Spectral information and spatial information are crucial for hyperspectral classification.

[0003] Traditional spectral-spatial classification methods rely too much on manual features and have weak robustness and generalization in complex scenes. Benefiting from the powerful deep, nonlinear information mining capabilities, deep learning-based algorithms can effectively solve this problem; however, due to the complex nonlinear coupling between spectral and spatial information, the results generated by hyperspectral classification methods based on three-dimensional information extraction in complex scenes are often unreliable.

[0004] This shows that the hyperspectral classification method in related technologies has the technical problem of low classification accuracy in complex scenarios. Summary of the invention

[0005] The present invention provides a hyperspectral image classification method based on spectral-spatial decoupling of a state-space model, which is used to solve the defect of low classification accuracy in complex scenes in the hyperspectral classification method in the prior art, and realizes the joint use of spatial information and spectral information to improve the classification accuracy of hyperspectral images.

[0006] The present invention provides a hyperspectral image classification method based on spectral-spatial decoupling of a state-space model, comprising the following steps.

[0007] The hyperspectral image to be classified is subjected to feature extraction to obtain shallow features; the shallow features are input into a spatial semantic modeling module based on a state-space model to obtain global spatial semantic information output by the spatial semantic modeling module; the shallow features are input into a spectral perception module based on a state-space model to obtain global spectral perception information output by the spectral perception module; channel-level dot multiplication is performed based on the global spatial semantic information and the global spectral perception information to obtain spatial-spectral coupling information; average pooling and maximum value indexing are performed based on the spatial-spectral coupling information to obtain a classification result of the hyperspectral image to be classified.

[0008] According to a hyperspectral image classification method based on spectral-spatial decoupling of a state-space model provided by the present invention, the shallow features are input into a spatial semantic modeling module based on a state-space model to obtain global spatial semantic information output by the spatial semantic modeling module, including: performing spatial information aggregation on the shallow features to obtain multi-scale spatial information; performing local association on the multi-scale spatial information to obtain local association information; performing semantic extraction on the local association information to obtain global spatial semantic information; the shallow features are input into a spectral perception module based on a state-space model to obtain global spectral perception information output by the spectral perception module, including: performing spectral information extraction on the shallow features to obtain multi-scale spectral information; performing spectral grouping on the multi-scale spectral information to obtain a band group sequence; performing spectral information analysis on the band group sequence to obtain global spectral perception information.

[0009] According to a hyperspectral image classification method based on spectral-spatial decoupling of a state-space model provided by the present invention, the spatial information of the shallow features is aggregated to obtain multi-scale spatial information, including: extracting features of the shallow features through 3×3 convolution to obtain adjacent pixel information; extracting features of the shallow features through 5×5 average pooling to obtain mid-scale rotation-invariant information; extracting features of the shallow features through 1×1 convolution to obtain pixel-level information; cascading and aggregating the adjacent pixel information, the mid-scale rotation-invariant information, the pixel-level information and the shallow features to obtain multi-scale spatial information, wherein the spatial scales of the adjacent pixel information, the mid-scale rotation-invariant information, the pixel-level information and the shallow features are the same.

[0010] According to a hyperspectral image classification method based on spectral-spatial decoupling of a state-space model provided by the present invention, the spectral information of the shallow features is extracted to obtain multi-scale spectral information, including: convolving, normalizing and activating the shallow features through a 1×1 convolution with a channel number of 7 to obtain first-scale spectral information; convolving, normalizing and activating the shallow features through a 1×1 convolution with a channel number of 3 to obtain second-scale spectral information; adding and globally pooling the first-scale spectral information, the second-scale spectral information and the shallow features to obtain multi-scale spectral information.

[0011] According to a hyperspectral image classification method based on spectral-spatial decoupling of a state-space model provided by the present invention, the multi-scale spectral information is spectrally grouped to obtain a band group sequence, including: converting the number of channels of the multi-scale spectral information into a target number of channels according to the following formula to obtain a band group sequence: ; in, represents the dimension of the band group sequence, represents the number of bands in the band group sequence, represents the interval number of the starting bands of adjacent band group sequences of the band group sequence, represents the number of channels of the multi-scale spectral information, represents the remainder operation, and Indicates rounding down and rounding up respectively.

[0012] According to a hyperspectral image classification method based on spectral-spatial decoupling of a state-space model provided by the present invention, average pooling and maximum indexing are performed based on the spatial-spectral coupling information to obtain a classification result of the hyperspectral image to be classified, including: performing 3×3 average pooling and 1×1 convolution on the spatial-spectral coupling information to obtain an output result; performing maximum indexing on the output result to determine the probability that each pixel in the hyperspectral image to be classified belongs to a target category among multiple preset categories.

[0013] The present invention also provides a hyperspectral image classification device with spectral-spatial decoupling based on a state-space model, comprising the following modules: an extraction module, used to perform feature extraction on the hyperspectral image to be classified to obtain shallow features; a spatial semantic module, used to input the shallow features into a spatial semantic modeling module based on a state-space model to obtain global spatial semantic information output by the spatial semantic modeling module; a spectral perception module, used to input the shallow features into a spectral perception module based on a state-space model to obtain global spectral perception information output by the spectral perception module; a channel dot multiplication module, used to perform channel-level dot multiplication based on the global spatial semantic information and the global spectral perception information to obtain spatial-spectral coupling information; a classification module, used to perform average pooling and maximum value indexing based on the spatial-spectral coupling information to obtain the classification result of the hyperspectral image to be classified.

[0014] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, a hyperspectral image classification method based on spectral-spatial decoupling of a state-space model as described above is implemented.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-described hyperspectral image classification methods based on spectral-spatial decoupling of a state-space model.

[0016] The present invention also provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the method for classifying hyperspectral images based on spectral-spatial decoupling of a state-space model as described above is implemented.

[0017] The hyperspectral image classification method based on spectral-spatial decoupling of a state-space model provided by the present invention obtains shallow features by extracting features from the hyperspectral image to be classified, thereby providing effective data input for subsequent spatial semantic modeling and spectral perception by extracting key information in the image; the shallow features are input into a spatial semantic modeling module based on the state-space model, and the state-space model can describe the spatial relationship between each pixel point in the image, thereby extracting the global spatial semantic information of the image; the shallow features are input into a spectral perception module based on the state-space model, and the spectral features of different substances in the image are analyzed, thereby obtaining global spectral perception information; The global spatial semantic information is multiplied by the global spectral perception information at the channel level to obtain the spatial-spectral coupling information, thereby effectively combining the spatial features and the spectral features to form a more comprehensive and accurate image feature description, thereby improving the accuracy of classification; average pooling and maximum indexing are performed based on the spatial-spectral coupling information. Average pooling can smooth image features and reduce noise interference; while maximum indexing can highlight the key features in the image and improve the robustness of classification, so as to obtain accurate and stable classification results for the hyperspectral images to be classified, thereby solving the technical problem of low classification accuracy in complex scenes in the hyperspectral classification method in related technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced one by one below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0019] Figure 1 It is a schematic flow chart of a hyperspectral image classification method based on spectral-spatial decoupling of a state-space model provided by the present invention.

[0020] Figure 2 It is a schematic diagram of the overall process of the hyperspectral image classification method based on spectral-spatial decoupling of the state-space model provided by the present invention.

[0021] Figure 3 It is a structural schematic diagram of a hyperspectral image classification device based on spectral-spatial decoupling of a state-space model provided by the present invention.

[0022] Figure 4It is a schematic diagram of the physical structure of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0023] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0024] Hyperspectral images are well-known for their ability to capture the fine spectral features of surface objects, and have unique advantages in accurately identifying various ground features. This ability makes hyperspectral images widely used in many fields such as agricultural protection, urban development and environmental protection. Hyperspectral classification technology is the key step in these applications.

[0025] Hyperspectral images are three-dimensional data bodies composed of spatial information in the space dimension and spectral information in the channel dimension. Spectral and spatial information are crucial for hyperspectral classification. Traditional spectral-space-based classification methods rely too much on manual features and have weak robustness and generalization in complex scenes. Benefiting from the powerful deep and nonlinear information mining capabilities, deep learning-based algorithms can effectively solve this problem. Spectral-space hyperspectral classification methods based on deep learning can be roughly divided into two categories, namely, hyperspectral classification methods based on three-dimensional information extraction and spectral-space separation. The former conforms to the unified attribute of hyperspectral images and can mine spectral-spatial correlation information while effectively protecting the physical structure of hyperspectral images. However, due to the complex nonlinear coupling between spectral and spatial information, the results generated by three-dimensional methods are often unreliable, especially in complex scenes. The latter mines spectral and spatial information in a separate manner by designing different modules, which can mine more representative spectral and spatial information in a more efficient way.

[0026] Methods for spectral-spatial decoupling can be divided into dual-branch-based methods and cascade-based methods. The cascade-based method facilitates the extraction and utilization of spectral and spatial information by connecting spatial and spectral feature extraction modules in series. For example, in the related art, a spectral processing module based on a convolutional neural network and an adjacent pixel information aggregation module are designed to process spectral information and spatial information in turn. In the related art, sub-networks based on spectral and spatial transformers alternately extract spectral and spatial features. Although these cascade structures show good performance, the spectral-spatial coupling characteristics of hyperspectral images (HSI) allow the present invention to infer that the processing of one module may affect the feature processing of subsequent modules. Specifically, the processing in the previous module may reduce the features of interest to the subsequent modules.

[0027] In contrast, the dual-branch based method processes spectral and spatial features simultaneously, effectively avoiding the above interference. The related art uses a dual-branch self-attention sub-network to extract spatial and spectral information. In the related art, the structure of extracting spectral and spatial information is integrated into the module, achieving excellent performance. However, these methods mainly focus on the extraction of spectral and spatial information, and often ignore their fusion. In order to solve this problem, the related art enhances the coupling of spectral information and spatial information by strengthening the information flow between the respective extraction structures.

[0028] Although the dual-branch decoupled classification method improves the spatial and spectral information modeling capabilities of the model, the existing methods still find it difficult to achieve an effective balance between spatial local detail processing and image processing speed. According to the differences in the way spatial information is used, the hyperspectral classification methods based on deep learning can be divided into patch-based methods and fully connected methods. The patch-based method can effectively model local spatial details by mining the spatial association between the central pixel and the adjacent pixels in the patch. However, this type of method realizes ground object recognition pixel by pixel, and there is repeated calculation in the recognition of adjacent pixels, which wastes a lot of computing resources when using this method for large-scale mapping based on hyperspectral images. However, the fully connected method can effectively solve this problem. The fully connected algorithm inputs a hyperspectral image into the network and outputs a classification map of the corresponding scale. This process effectively avoids repeated calculations and can achieve fast and efficient hyperspectral image mapping.

[0029] However, most of the existing fully connected algorithms use convolutional neural networks and transformers to model spatial information, which makes the model limited by the receptive field and quadratic complexity, making it difficult to achieve lightweight global spatial and spectral information modeling. Recently, the selective state space model (mamba) has been introduced into the field of remote sensing, which can achieve long-range information modeling with linear complexity, making it possible to achieve lightweight and efficient object recognition based on fully connected hyperspectral classification models.

[0030] The present invention provides a hyperspectral image classification method based on spectral-spatial decoupling of a state-space model, which is used to at least solve the following problems.

[0031] Preservation of local spatial structure and utilization of global spatial semantics. The association between objects exists not only between adjacent pixels, but also at the scene scale. The joint use of local and global information helps alleviate the confusion problem when the model recognizes complex objects. Mamba's one-dimensional linear scanning mechanism enables it to model global spatial semantics in a lightweight way, but it needs to expand the two-dimensional information into a one-dimensional sequence when mining two-dimensional spatial information. This makes the pixels that are close to each other in space far apart in the sequence, which seriously destroys the original local spatial structure of the image.

[0032] Three-dimensional spatial-spectral information fusion. Hyperspectral images have the property of spatial-spectral integration, and the joint representation of spatial and spectral information can achieve high-precision recognition of objects. Therefore, the key points of the separation and extraction method should include decoupling and fusion, and the fusion of spatial and spectral information is the ultimate goal. However, previous methods mainly focused on the design of spectral and spatial information extraction modules. There is almost no correlation between the spatial module and the spectral module, and they are finally fused in an extensive form. Such a spatial-spectral information mining framework is not conducive to the joint use of spatial and spectral information.

[0033] The present invention designs a new mamba-based global spectral space hyperspectral classification method, which uses different branches to model one-dimensional spectral information and two-dimensional spatial information respectively, and introduces a new fusion paradigm to achieve the coupling of three-dimensional spatial spectrum information. Specifically, the method uses the spatial module to model local and global spatial information with a state space model, uses the spectral module to perceive the maximum correlation information between bands to generate spectral weights, and finally realizes the fusion of spatial and spectral information in a weighted manner. In order to effectively improve the robustness and scientificity of the model, the present invention embeds the L1 loss into the network to effectively constrain the output of the spectral module.

[0034] Optionally, the hyperspectral image classification method based on spectral-spatial decoupling of a state-space model in the embodiment of the present application can be executed by a server, or by a terminal device, or jointly by a server and a terminal device, taking the example of the hyperspectral image classification method based on spectral-spatial decoupling of a state-space model in the present embodiment being executed by a server.

[0035] Figure 1 Schematic diagram of the process of the hyperspectral image classification method based on spectral spatial decoupling of the state space model provided by the present invention. Figure 1 As shown, the method includes the following steps.

[0036] Step 101 , extracting features from the hyperspectral image to be classified to obtain shallow features.

[0037] In the embodiment of the present invention, it is assumed that Represents the input hyperspectral image to be classified, where Represent the number of rows and columns, and the number of spectral bands, respectively.

[0038] Using The shallow feature extraction module composed of convolution, batch normalization, and GELU activation function performs shallow feature extraction on the hyperspectral image to be classified. This process can be expressed as: (1) in, represents the shallow feature extraction module, Represents shallow features, represents the set of real numbers, Represents the number of rows of the hyperspectral image to be classified, represents the number of columns of the hyperspectral image to be classified, Represents shallow feature dimensions.

[0039] Afterwards, the embodiment of the present invention utilizes the spatial semantic modeling module and the spectral perception module to extract spectral information and spatial information in a decoupled manner.

[0040] In the embodiment of the present invention, the hyperspectral image classification method based on spectral-spatial decoupling of the state-space model is a dual-branch structure, including: a spatial semantic modeling module (SPAM) and a spectral perception module (SPEM).

[0041] Step 102, inputting the shallow features into a spatial semantic modeling module based on a state space model to obtain global spatial semantic information output by the spatial semantic modeling module.

[0042] In an embodiment of the present invention, the spatial semantic modeling module (SPAM) is composed of a multi-scale spatial module, a local spatial association modeling module, and a global spatial semantic modeling module.

[0043] According to a hyperspectral image classification method based on spectral-spatial decoupling of a state-space model provided by the present invention, shallow features are input into a spatial semantic modeling module based on the state-space model to obtain global spatial semantic information output by the spatial semantic modeling module, including: Aggregate spatial information of shallow features to obtain multi-scale spatial information; Perform local correlation on multi-scale spatial information to obtain local correlation information; Semantic extraction is performed on local correlation information to obtain global spatial semantic information.

[0044] In the Spatial Semantic Modeling Module (SPAM), the present invention first uses the multi-scale spatial module to realize the aggregation of multi-scale spatial information, and uses the local spatial association modeling module to process the local spatial association features. Afterwards, the present invention aggregates the multi-scale spatial information and the local spatial association through jump connections, which can help the network obtain multi-scale spatial (local) information and local association information at the same time. This process can be expressed as: (2) (3) in, Represents multi-scale spatial information, represents a multi-scale spatial module, Represents shallow features, Represents local association information, Represents the local spatial association modeling module. , Respectively represent the aggregated multi-scale spatial information and local spatial (correlation) information, represents the set of real numbers, Represents the number of rows of the hyperspectral image to be classified, represents the number of columns of the hyperspectral image to be classified, Represents shallow feature dimensions.

[0045] Afterwards, the global spatial semantic modeling module is set up to extract the global spatial semantics. Therefore, the output of the spatial semantic modeling module is It can be obtained by the following formula.

[0046] (4) in represents the global space semantic modeling module, Represents global spatial semantic information, Represents local association information.

[0047] Through the embodiments of the present invention, by aggregating spatial information of shallow features, information from different locations can be integrated to enhance the model's understanding of spatial structure; after local association processing of multi-scale spatial information, association relationships between different regions or features can be established; semantic extraction of local association information can further refine global spatial semantic information.

[0048] Step 103: input the shallow features into a spectrum perception module based on a state space model to obtain global spectrum perception information output by the spectrum perception module.

[0049] In the embodiment of the present invention, the spectral perception module (SPEM) is composed of a multi-scale spectral information extraction module, a spectral coarse grouping module, and a spectral dynamic association clustering module.

[0050] According to a hyperspectral image classification method based on spectral-spatial decoupling of a state-space model provided by the present invention, shallow features are input into a spectral perception module based on a state-space model to obtain global spectral perception information output by the spectral perception module, including: Extract spectral information of shallow features to obtain multi-scale spectral information; Spectrally group the multi-scale spectral information to obtain a band group sequence; The spectral information of the band group sequence is analyzed to obtain the global spectral perception information.

[0051] In the embodiment of the present invention, in the spectrum perception module (SPEM), a multi-scale spectrum information extraction module is first used to realize the primary modeling and fusion of multi-scale spectrum information, and then based on the correlation between adjacent spectral bands, a spectrum coarse grouping module is used to realize the coarse clustering of spectrum information, which can be specifically expressed by the following formula: (5) (6) in, represents multi-scale spectral information, represents the multi-scale spectral information extraction module, Represents shallow features, represents a sequence of band groups, Represents the spectrum coarse grouping module.

[0052] Here, express The output, Represents shallow feature dimensions; represents the spectrum coarse grouping module, where Indicates the number of bands contained in each band group. Indicates the number of bands between the starting bands of two band groups. express Output band group sequence, where Indicates the number of band groups.

[0053] Afterwards, the present invention uses the state space model to perform a detailed spectral information analysis on the band group sequence to achieve modeling of dynamic changes in spectral information along the channel direction. The processing of the band group sequence by the spectral dynamic association clustering module can be expressed as follows: (7) in, represents the spectral dynamic association clustering module, represents the output of the (maximum correlation) spectral perception module, i.e., the global spectral perception information, Represents shallow feature dimensions.

[0054] Through the embodiments of the present invention, extracting multi-scale spectral information helps to capture the subtle changes of the target object in different bands, thereby improving the accuracy of spectral information extraction; the formation of band group sequences can provide a more targeted data basis for subsequent spectral information analysis, making the analysis more focused and in-depth; performing spectral information analysis on the band group sequence can obtain global spectral perception information. Global spectral perception information is a comprehensive reflection of the characteristics of the target object in the entire spectral range, which helps to achieve overall cognition and understanding of the target object.

[0055] Step 104 , performing channel-level dot multiplication based on the global spatial semantic information and the global spectral perception information to obtain spatial-spectral coupling information.

[0056] In the embodiment of the present invention, the coupling of channel-level spectrum and spatial information is achieved through a dot product operation, which can be specifically expressed by the following formula: (8) in, represents the spatial-spectral coupling information, , represents the multiplication of corresponding elements, represents the set of real numbers, Represents the number of rows of the hyperspectral image to be classified, represents the number of columns of the hyperspectral image to be classified, represents the shallow feature dimension, Represents global spatial semantic information, Represents global spectral perception information.

[0057] According to the above formula, it can be clearly concluded that the output of the spectral information extraction module is the weight of the spatial information, so that the spectral branch can serve the spatial branch, ensuring the correlation between the spectral and spatial information when they are mined separately in a decoupled manner, which helps to achieve more effective and more explanatory coupling of spatial information and spectral information.

[0058] Step 105 , performing average pooling and maximum indexing based on the spatial-spectral coupling information to obtain a classification result of the hyperspectral image to be classified.

[0059] Average Pooling is a commonly used information aggregation method that can be used to reduce data noise while retaining important features.

[0060] In classification tasks, the maximum index is often used to determine the most likely category label for each pixel. Specifically, it is to find the index corresponding to the element with the largest value in the feature vector of each pixel, which usually corresponds to a preset category label. In this way, it can be determined which preset category each pixel is most likely to belong to.

[0061] According to a hyperspectral image classification method based on spectral-spatial decoupling of a state-space model provided by the present invention, average pooling and maximum indexing are performed based on spatial-spectral coupling information to obtain a classification result of a hyperspectral image to be classified, including: Perform 3×3 average pooling and 1×1 convolution on the spatial-spectral coupling information to obtain the output result; The output results are indexed by the maximum value to determine the probability that each pixel in the hyperspectral image to be classified belongs to a target category in multiple preset categories.

[0062] In the embodiment of the present invention, the output module composed of average pooling and channel convolution with a kernel size of 3 and the argmax function are used to obtain the category of the corresponding position. This process can be expressed as: (9) (10) in, Represents the classification result of the hyperspectral image to be classified, is the model output, where Indicates that the corresponding pixel in the image belongs to The probability of the class, represents the total number of categories, represents the set of real numbers, Represents the number of rows of the hyperspectral image to be classified, The number of columns representing the hyperspectral image to be classified.

[0063] Through the embodiments of the present invention, a new feature map (i.e., output result) can be obtained by performing 3×3 average pooling and 1×1 convolution processing on the spatial-spectral coupling information; then, the probability or label of each pixel belonging to a target category in multiple preset categories can be determined by the maximum value index (argmax function).

[0064] Through the above steps of the embodiment of the present invention, the spatial features and the spectral features are effectively combined together to form a more comprehensive and accurate image feature description, thereby improving the accuracy of classification; average pooling and maximum indexing are performed based on spatial-spectral coupling information, and average pooling can smooth image features and reduce noise interference; while maximum indexing can highlight key features in the image and improve the robustness of classification, so as to obtain accurate and stable classification results of the hyperspectral image to be classified, thereby solving the technical problem of low classification accuracy in complex scenes in the hyperspectral classification method in related technologies.

[0065] According to a hyperspectral image classification method based on spectral-spatial decoupling of a state-space model provided by the present invention, spatial information aggregation is performed on shallow features to obtain multi-scale spatial information, including: The shallow features are extracted through 3×3 convolution to obtain the adjacent pixel information; The shallow features are extracted through 5×5 average pooling to obtain the mid-scale rotation invariant information; The shallow features are extracted through 1×1 convolution to obtain pixel-level information; Adjacent pixel information, mid-scale rotation invariant information, pixel-level information and shallow features are cascaded and aggregated to obtain multi-scale spatial information, wherein the spatial scales of adjacent pixel information, mid-scale rotation invariant information, pixel-level information and shallow features are the same.

[0066] The 3×3 convolution kernel is one of the most commonly used convolution kernel sizes in image processing. It can capture the relationship between adjacent pixels in the image and extract local features such as edges and textures. 5×5 average pooling means calculating the average value of pixel values ​​in each 5×5 neighborhood to obtain a new pixel value. Compared with 3×3 convolution, 5×5 average pooling can capture a larger image area and extract mid-scale features. Since average pooling is robust to rotation and position changes in the image, it helps to extract rotation-invariant information. 1×1 channel convolution can be regarded as a linear transformation of the channel features of each pixel to extract pixel-level information.

[0067] In the embodiment of the present invention, the fusion of local and spatial information helps to improve the model's ability to identify objects with complex spatial structures. In addition, the accuracy of local information helps to improve the modeling of global information. Therefore, the spatial semantic modeling module performs spatial information modeling by adopting a cascade method of first local and then global. The following will specifically introduce the process of the spatial semantic modeling module (SPAM) realizing the joint utilization of local and global spatial information based on the state space model.

[0068] The state space model processes one-dimensional sequence information, which will cause the destruction of the spatial structure. In order to solve this problem, the present invention improves the protection of local two-dimensional information when the state space model performs spatial information modeling by adding a multi-scale space module to perform preliminary modeling of two-dimensional multi-scale spatial information. The multi-scale space module uses 3×3 convolution to extract adjacent pixel information, 5×5 average pooling to extract mid-scale rotation invariant information, and 1×1 channel convolution to extract pixel-level information. The multi-scale feature extraction process can be expressed as the following formula: (11) (12) (13) in, Represents adjacent pixel information, represents the Gaussian Error Linear Unit (GELU) activation function, represents batch normalization, represents 3×3 convolution, Indicates shallow features; represents the mid-scale rotation invariant information, represents 5×5 average pooling, Represents pixel-level information, Represents a 1×1 channel convolution. , , and Have the same spatial dimensions.

[0069] Afterwards, by concatenating the multi-scale features with the original information and dynamically integrating the spatial information of each pixel through 1×1 convolution, this process can be expressed as: (14) in, Represents multi-scale spatial information, represents the Gaussian error linear unit activation function, represents 1×1 convolution, Indicates cascade operation, They represent adjacent pixel information, mid-scale rotation invariant information, pixel-level information, and shallow features respectively.

[0070] In order to achieve further aggregation of multi-scale spatial information and enhance local key information, the present invention designs a local spatial association modeling module. The local spatial association modeling module requires the model to be processed in blocks, and these blocks (i.e., patches) are non-overlapping. Therefore, in order to meet the non-overlapping block requirements of images of different spatial sizes, the present invention first expands the image. Afterwards, the present invention blocks the hyperspectral image to be classified and uses the state space model to perform association modeling on the local spatial patches.

[0071] Assume that the size of the patch is p and the original size of the image is (rows * columns), the expanded size is (row*column), then , , Remaining operation, the number of patch blocks is Therefore, the image after patch segmentation can be expressed as ,in The state-space model's processing of patches is formulated as: (15) (16) (17) in, , represents the intermediate output of local spatial correlation modeling, ,in Represents the output of the local space modeling module; Indicates block operation. represents a linear transformation, Representation layer normalization, represents the flattening operation, Represents the hyperspectral image to be classified. Patch blocks, represents the selective state space model, represents the activation function, represents 1×1 convolution, represents a reshape operation, Represents element-wise multiplication.

[0072] Note that the mamba processing modules of each patch share weights to effectively suppress noise and obtain local correlation properties that conform to global rules. Afterwards, the present invention uses the synthesis operation patches to stitch and restore the image in sequence. Since the padding operation changes the original size of the image, the present invention finally cuts off the padding part. Therefore, this process can be described as: (18) in Represents the local spatial association modeling module ( ) output, Represents a splicing operation, Represents the output of the local space modeling module.

[0073] The state space model can model long-range dependencies with linear complexity, which effectively alleviates the limitations of convolutional neural networks and transformers in receptive field and model complexity. Therefore, the present invention uses it to model the global spatial semantics of hyperspectral.

[0074] First, the hyperspectral image to be classified is segmented pixel by pixel to generate a one-dimensional sequence , after which the hyperspectral image to be classified is input into the state space model for global spatial semantic mining. This process can be described as: (19) (20) (twenty one) in is the intermediate variable of the global state space block, is the output of the state space block, Indicates block operation. represents a linear transformation, Representation layer normalization, Represents a one-dimensional sequence generated by pixel-by-pixel segmentation of the hyperspectral image to be classified. represents the selective state space model, represents the activation function, represents 1×1 convolution, represents a reshape operation, Represents element-wise multiplication.

[0075] Finally, the present invention further processes the spatial information using an information integration module consisting of a linear layer, a normalization layer, and an activation layer in preparation for fusion with the spectral information.

[0076] Therefore, the spatial module finally outputs global spatial semantic information It can be obtained by the following formula.

[0077] (twenty two) in, Represents global spatial semantic information, represents the Gaussian error linear unit activation function, Representation layer normalization, represents a linear transformation, Represents the output of the state space block.

[0078] Through the embodiment of the present invention, the spatial semantic modeling process is explained. In the spatial branch, the present invention introduces a multi-scale spatial module, a local spatial association modeling module, and a global spatial semantic modeling module. The first two modules can effectively model local spatial information, and the last module can lightweight global information. Therefore, compared with the existing global mamba-based spectral space hyperspectral classification method, the method of the present invention can effectively protect the local spatial structure when mining global spatial semantics.

[0079] According to a hyperspectral image classification method based on spectral-spatial decoupling of a state-space model provided by the present invention, spectral information of shallow features is extracted to obtain multi-scale spectral information, including: The shallow features are convolved, normalized, and activated through a 1×1 convolution with 7 channels to obtain the first-scale spectral information; The shallow features are convolved, normalized, and activated through a 1×1 convolution with 3 channels to obtain the second-scale spectral information. The first-scale spectral information, the second-scale spectral information and the shallow features are added and globally pooled to obtain multi-scale spectral information.

[0080] Based on the directionality and correlation principle of spectral space, the present invention uses spectral grouping to generate a sequence along the channel direction so that multiple spectra jointly represent a feature block, and input the sequence into the state space model for spectral dynamic association clustering. This section will specifically introduce the process of achieving maximum correlation spectral perception based on the state space model.

[0081] Since the spectral grouping operation may cause similar spectra to be in different groups, in order to enhance the interaction between band groups and facilitate the state space model to perform spectral dynamic analysis. The present invention first uses a multi-scale spectral information extraction module to process spectral information. The multi-scale spectral information extraction process can be expressed as: (twenty three) (twenty four) (25) in, , Respectively represent multi-scale spectral information. In order to protect the spectral sequence, the present invention fuses the multi-scale spectral information by adding corresponding elements. represents the first-scale spectral information, represents the Gaussian error linear unit activation function, represents a 1×1 convolution with 7 channels. Indicates shallow features; represents the second scale spectral information, represents a 1×1 convolution with 3 channels. represents multi-scale spectral information, represents global pooling, Indicates the number of rows and columns of the hyperspectral image to be classified. Next, A coarse grouping will be performed to achieve primary clustering of spectral information.

[0082] Through the embodiments of the present invention, spectral information of different scales can be extracted from shallow features through 1×1 convolution operations with different numbers of channels; 1×1 convolution can not only maintain the consistency of features in the spatial dimension, but also combine and interact features in the channel dimension. By adjusting the number of output channels, nonlinear relationships between features and the extraction of complex features can be achieved, and the expression ability of the model can be enhanced; pooling methods such as global average pooling can capture the global characteristics of feature maps and reduce the model's excessive reliance on local information.

[0083] According to a hyperspectral image classification method based on spectral-spatial decoupling of a state-space model provided by the present invention, the multi-scale spectral information is spectrally grouped to obtain a band group sequence, including: The number of channels of the multi-scale spectral information is converted into the target number of channels according to the following formula to obtain a band group sequence: ; in, represents the dimension of the band group sequence, represents the number of bands in the band group sequence, represents the interval number of the starting bands of adjacent band group sequences of the band group sequence, represents the number of channels of the multi-scale spectral information, represents the remainder operation, and Indicates rounding down and rounding up respectively.

[0084] Adjacent spectra have high correlation, and redundant spectra will cause Hughes phenomenon. The grouping method can effectively reduce redundant spectra. The present invention first sets two parameters, namely, the number of bands m in the band group and the number of intervals n between the starting bands of adjacent band groups. The number of channels is c, and the grouping module outputs the result .therefore, and The relationship can be expressed as the following formula.

[0085] (26) in and Represents rounding down and rounding up respectively. Through the grouping operation, the original spectral information is transformed from the c-dimensional sequence into In addition, the sequential generation of spectral sequences also protects the directionality of spectral information and can be naturally combined with the state space model to model long-range spectral associations.

[0086] The information processed by the spectral dynamic association clustering module is , the processing process can be expressed as: (27) (28) (29) in, is the intermediate variable, , which is the output of the mamba processing module. It can be expressed as , Indicates block operation. represents a linear transformation, Representation layer normalization, represents a sequence of band groups, represents the selective state space model, represents the activation function, represents 1×1 convolution, Represents element-wise multiplication.

[0087] In order to better illustrate the process of the state space model processing the spectral sequence process, assume The output is , then mamba The processing process can be expressed as: (30) (31) in, is the parameter of the control state mapping after discretization, is the state corresponding to different positions in the sequence, Depends on input Specifically, Mamba uses , , , to generate parameters , , ,in , , Represents the mapping matrix.

[0088] From the formula, we can see that in the process of sequence traversal, through The update and transmission of adjacent spectra, the correlation information of adjacent spectra is transmitted, and the long-range dependency is discovered by the model in a way that conforms to the directional attributes of the spectral information. Restore to one-dimensional sequence , and use the linear layer to obtain the spectral weight of the corresponding position in the form of global correlation. This process can be expressed as.

[0089] (32) in, represents the global spectral perception information, represents the Gaussian error linear unit activation function, represents a linear transformation, Represents a one-dimensional sequence.

[0090] Through the embodiments of the present invention, in order to realize the joint utilization of spectral and spatial information, spectral branches are used to generate channel weights. In this way, spatial and spectral branches can achieve functional-level interaction, which not only ensures the decoupling of spatial-spectral information, but also promotes the fusion of spatial and spectral information. Compared with the existing weighted addition method, the channel weight is only one-dimensional data, so the method used in the present invention is lighter. In addition, the spectral weighting method used in the present invention ensures the functional interaction between spatial and spectral branches. Compared with the traditional dual-branch structure that independently extracts spatial and spectral information, the method of the present invention is more conducive to realizing the joint utilization of spatial and spectral information.

[0091] In order to give full play to the performance of the model, the present invention uses a joint constraint method that balances sample distribution and reduces the correlation of spectral weights to optimize network parameters. First, the present invention uses focal_loss to generate category loss from the sample globally to improve the model's ability to handle category imbalance problems. For details, please refer to the following formula: (33) (34) in, Represents the indicator function, used to represent the position ( ) on the category Is it a true label (or some form of label). If Equal to a specific tag value ,but =1; otherwise =0, represents the true value, represents the focal loss. represents the weight factor, represents the parameter of interest, where Indicates that the pixel at the corresponding position in the hyperspectral image to be classified belongs to The probability of the class, represents the total number of categories, Respectively represent the number of rows and columns in the hyperspectral image to be classified.

[0092] In order to ensure that the spectral module can achieve maximum correlation spectral perception, the L1 norm is used to perform deep supervision on the output of the spectral branch. For details, please refer to the following formula: (35) (36) in, express The elements, Represents a conditional expression, that is, when When the expression is true, the value is 1; otherwise, it is 0; To calculate from 1 to All How many are there in is greater than 1.5, Represents shallow feature dimensions; represents the deep supervision loss function, Respectively represent the number of rows and columns in the hyperspectral image to be classified, represents the global spectral perception information, represents the balance parameter. The setting is not all zero, so At least T data are greater than 1.5.

[0093] The overall loss function can be expressed as: (37) in, represents the overall loss function, represents the focal loss function, represents the deep supervision loss function.

[0094] refer to Figure 2 , Figure 2 It is a schematic diagram of the overall process of the hyperspectral image classification method with spectral-spatial decoupling based on the state-space model provided by the present invention; it mainly consists of five parts: a feature input head, a global spatial semantic modeling module, a global spectral perception module, a dot product fusion module (i.e., channel-level dot product) and a classification head.

[0095] Among them, the global spectral perception module includes: multi-scale spectral information extraction, spectral grouping and spectral dynamic association clustering; the global spatial semantic modeling module includes: multi-scale spatial information extraction, local spatial association information extraction and global spatial semantic information extraction.

[0096] Through the above-mentioned embodiments provided by the present invention, based on the one-dimensional scanning mechanism of the selective state space model (mamba), the decoupling and fusion of one-dimensional spectral and two-dimensional spatial information in three-dimensional HSI are explored, and a spectral-spatial decoupled global mamba for lightweight and high-precision HSI classification is proposed.

[0097] In order to effectively mine global and local 2D spatial information, this paper develops a spatial semantic modeling module (SPAM), which consists of a multi-scale spatial extraction module, a local spatial association module, and a global spatial context modeling module. By integrating these three modules, the spatial semantic modeling module takes advantage of Mamba's advantages in modeling global dependencies, while also addressing the challenges associated with Mamba's one-dimensional state transition to exploit the two-dimensional spatial structure.

[0098] A new method for the scientific integration of spatial and spectral information is proposed. First, a method for generating spectral weights based on global spectral perception is proposed. On this basis, point multiplication is introduced into the global classification framework for the first time, realizing the channel-level fusion of global spectral information and spatial information.

[0099] In the spatial branch, a multi-scale spatial module, a local spatial association modeling module, and a global spatial semantic modeling module are introduced. The first two modules can effectively model local spatial information, and the last module can lightweight global information. Therefore, compared with the existing global mamba-based spectral space hyperspectral classification method, the embodiment of the present invention can effectively protect the local spatial structure while mining the global spatial semantics.

[0100] In order to realize the joint utilization of spectral and spatial information, spectral branches are used to generate channel weights. In this way, spatial and spectral branches can achieve functional-level interaction, which not only ensures the decoupling of spatial-spectral information, but also promotes the fusion of spatial and spectral information. Compared with the existing weighted addition method, the channel weight is only one-dimensional data, so the method used in the present invention is lighter. In addition, the spectral weighting method used in the present invention ensures the functional interaction between spatial and spectral branches. Compared with the traditional dual-branch structure that independently extracts spatial and spectral information, the method of the present invention is more conducive to realizing the joint utilization of spatial and spectral information.

[0101] The hyperspectral image classification device based on spectral spatial decoupling of the state-space model provided by the present invention is described below. The hyperspectral image classification device based on spectral spatial decoupling of the state-space model described below and the hyperspectral image classification method based on spectral spatial decoupling of the state-space model described above can be referred to each other.

[0102] refer to Figure 3 , Figure 3 It is a structural schematic diagram of a hyperspectral image classification device based on spectral-spatial decoupling of a state-space model provided by the present invention.

[0103] An extraction module 301 is used to extract features from the hyperspectral image to be classified to obtain shallow features; A spatial semantic module 302 is used to input shallow features into a spatial semantic modeling module based on a state space model to obtain global spatial semantic information output by the spatial semantic modeling module; The spectrum perception module 303 is used to input the shallow features into the spectrum perception module based on the state space model to obtain the global spectrum perception information output by the spectrum perception module; A channel dot product module 304 is used to perform channel-level dot product based on global spatial semantic information and global spectral perception information to obtain spatial-spectral coupling information; The classification module 305 is used to perform average pooling and maximum indexing based on the spatial-spectral coupling information to obtain a classification result of the hyperspectral image to be classified.

[0104] Specifically, the hyperspectral image classification device with spectral-spatial decoupling based on the state-space model provided by the present invention can implement all the method steps implemented by the above-mentioned hyperspectral image classification method embodiment with spectral-spatial decoupling based on the state-space model, and can achieve the same technical effect. The parts and beneficial effects that are the same as those in the method embodiment will not be described in detail herein.

[0105] Figure 4 is a schematic diagram of the physical structure of the electronic device provided by the present invention, such as Figure 4 As shown, the electronic device may include: a processor 410, a communication interface 420, a memory 430 and a communication bus 440, wherein the processor 410, the communication interface 420 and the memory 430 communicate with each other through the communication bus 440. The processor 410 may call the logic instructions in the memory 430 to execute the hyperspectral image classification method based on the spectral space decoupling of the state space model, the method comprising: extracting features of the hyperspectral image to be classified to obtain shallow features; inputting the shallow features into the spatial semantic modeling module based on the state space model to obtain the global spatial semantic information output by the spatial semantic modeling module; inputting the shallow features into the spectral perception module based on the state space model to obtain the global spectral perception information output by the spectral perception module; performing channel-level dot multiplication based on the global spatial semantic information and the global spectral perception information to obtain the spatial spectral coupling information; performing average pooling and maximum value indexing based on the spatial spectral coupling information to obtain the classification result of the hyperspectral image to be classified.

[0106] In addition, the logic instructions in the above-mentioned memory 430 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0107] On the other hand, the present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the hyperspectral image classification method based on spectral spatial decoupling of the state-space model provided by the above methods. The method includes: extracting features of the hyperspectral image to be classified to obtain shallow features; inputting the shallow features into a spatial semantic modeling module based on the state-space model to obtain global spatial semantic information output by the spatial semantic modeling module; inputting the shallow features into a spectral perception module based on the state-space model to obtain global spectral perception information output by the spectral perception module; performing channel-level dot multiplication based on the global spatial semantic information and the global spectral perception information to obtain spatial-spectral coupling information; performing average pooling and maximum value indexing based on the spatial-spectral coupling information to obtain the classification result of the hyperspectral image to be classified.

[0108] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the hyperspectral image classification method based on spectral-spatial decoupling of the state-space model provided by the above-mentioned methods, the method comprising: extracting features of the hyperspectral image to be classified to obtain shallow features; inputting the shallow features into a spatial semantic modeling module based on the state-space model to obtain global spatial semantic information output by the spatial semantic modeling module; inputting the shallow features into a spectral perception module based on the state-space model to obtain global spectral perception information output by the spectral perception module; performing channel-level dot multiplication based on the global spatial semantic information and the global spectral perception information to obtain spatial-spectral coupling information; performing average pooling and maximum value indexing based on the spatial-spectral coupling information to obtain the classification result of the hyperspectral image to be classified.

[0109] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0110] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention 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 invention.

Claims

1. A hyperspectral image classification method based on spectral-spatial decoupling of a state-space model, characterized in that: include: Perform feature extraction on the hyperspectral image to be classified to obtain shallow features; Inputting the shallow features into a spatial semantic modeling module based on a state space model to obtain global spatial semantic information output by the spatial semantic modeling module; Inputting the shallow features into a spectral perception module based on a state-space model to obtain global spectral perception information output by the spectral perception module; Performing channel-level dot multiplication on the global spatial semantic information and the global spectral perception information to obtain spatial-spectral coupling information; Average pooling and maximum indexing are performed based on the spatial-spectral coupling information to obtain a classification result of the hyperspectral image to be classified.

2. The hyperspectral image classification method based on spectral-spatial decoupling of the state-space model according to claim 1 is characterized in that: The step of inputting the shallow features into a spatial semantic modeling module based on a state space model to obtain global spatial semantic information output by the spatial semantic modeling module includes: Performing spatial information aggregation on the shallow features to obtain multi-scale spatial information; Performing local correlation on the multi-scale spatial information to obtain local correlation information; Performing semantic extraction on the local association information to obtain global spatial semantic information; The step of inputting the shallow features into a spectral perception module based on a state space model to obtain global spectral perception information output by the spectral perception module includes: Extracting spectral information of the shallow features to obtain multi-scale spectral information; Spectrally grouping the multi-scale spectral information to obtain a band group sequence; Spectral information analysis is performed on the band group sequence to obtain global spectral perception information.

3. The hyperspectral image classification method based on spectral-spatial decoupling of state-space model according to claim 2 is characterized in that: The aggregating spatial information of the shallow features to obtain multi-scale spatial information includes: Extract the shallow features through 3×3 convolution to obtain adjacent pixel information; Extract the shallow features through 5×5 average pooling to obtain mid-scale rotation invariant information; Extract the shallow features through 1×1 convolution to obtain pixel-level information; The adjacent pixel information, the mesoscale rotation invariant information, the pixel level information and the shallow features are cascaded and aggregated to obtain multi-scale spatial information, wherein the adjacent pixel information, the mesoscale rotation invariant information, the pixel level information and the shallow features have the same spatial scale.

4. The hyperspectral image classification method based on spectral-spatial decoupling of state-space model according to claim 2, characterized in that: The extracting spectral information of the shallow features to obtain multi-scale spectral information includes: The shallow features are convolved, normalized, and activated by a 1×1 convolution with 7 channels to obtain first-scale spectral information; The shallow features are convolved, normalized, and activated by a 1×1 convolution with 3 channels to obtain second-scale spectral information; The first-scale spectral information, the second-scale spectral information and the shallow features are added and globally pooled to obtain multi-scale spectral information.

5. The hyperspectral image classification method based on spectral-spatial decoupling of state-space model according to claim 2, characterized in that: The spectral grouping of the multi-scale spectral information to obtain a band group sequence includes: The number of channels of the multi-scale spectral information is converted into the target number of channels according to the following formula to obtain a band group sequence: ； in, represents the dimension of the band group sequence, represents the number of bands in the band group sequence, represents the interval number of the starting bands of adjacent band group sequences of the band group sequence, represents the number of channels of the multi-scale spectral information, represents the remainder operation, and Indicates rounding down and rounding up respectively.

6. The hyperspectral image classification method based on spectral-spatial decoupling of state-space model according to claim 1, characterized in that: The performing average pooling and maximum value indexing based on the spatial-spectral coupling information to obtain the classification result of the hyperspectral image to be classified includes: Performing 3×3 average pooling and 1×1 convolution on the spatial-spectral coupling information to obtain an output result; The output result is indexed by maximum value to determine the probability that each pixel in the hyperspectral image to be classified belongs to a target category among multiple preset categories.

7. A hyperspectral image classification device based on spectral spatial decoupling of a state-space model, characterized in that: include: An extraction module is used to extract features from the hyperspectral image to be classified and obtain shallow features; A spatial semantic module, used for inputting the shallow features into a spatial semantic modeling module based on a state space model, to obtain global spatial semantic information output by the spatial semantic modeling module; A spectral perception module, used for inputting the shallow features into the spectral perception module based on the state space model to obtain the global spectral perception information output by the spectral perception module; A channel dot multiplication module, used for performing channel-level dot multiplication based on the global spatial semantic information and the global spectral perception information to obtain spatial-spectral coupling information; The classification module is used to perform average pooling and maximum indexing based on the spatial-spectral coupling information to obtain a classification result of the hyperspectral image to be classified.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the hyperspectral image classification method based on spectral-spatial decoupling of the state-space model as claimed in any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the hyperspectral image classification method based on spectral-spatial decoupling of a state-space model as claimed in any one of claims 1 to 6 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the hyperspectral image classification method based on spectral-spatial decoupling of a state-space model as claimed in any one of claims 1 to 6 is implemented.

Images