A multi-path evolution-based few-shot hyperspectral remote sensing image classification method and system

Abstract

The present application relates to the technical field of remote sensing image processing, in particular to a multi-path evolution-based few-sample hyperspectral remote sensing image classification method and system. The method comprises generating spectral attention weights by using spectral compression reconstruction according to the obtained multi-scale deep feature representation, performing three-dimensional convolution evolution on the weighted features based on adaptive multi-path feature evolution to obtain information fusion features; the information fusion features of the three paths are weighted and fused through a dynamic gating mechanism; and the classification result of the hyperspectral image is obtained according to the fused features. The present application realizes adaptive evolution and cross-path selective fusion of multi-scale features; the deep features coded by the above multi-level features are used for final classification reasoning, so that high-precision hyperspectral image classification is realized under the condition of few samples.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and in particular to a method and system for classifying few-sample hyperspectral remote sensing images based on multi-path evolution. Background Technology

[0002] Hyperspectral remote sensing images, with their dozens to hundreds of spectral bands, can provide detailed spectral difference information for ground object identification. However, the acquisition and manual annotation of hyperspectral data are costly, often resulting in insufficient available labeled samples in practical applications. Under limited sample conditions, hyperspectral images exhibit characteristics such as high spectral dimensionality, uneven sample distribution, large intra-class spectral variations, and subtle inter-class differences, making it difficult for traditional deep learning methods, which rely on large-scale training samples, to fully learn the joint representation of spectral and spatial features. Existing few-sample classification techniques still have shortcomings in handling multi-scale correlations, spectral-spatial coupling characteristics, and feature generalization capabilities of hyperspectral data, making it difficult to achieve stable and reliable classification performance under limited sample conditions. Therefore, how to effectively extract spectral-spatial representations and improve classification accuracy with a small number of labeled samples has become a significant technical challenge in the field of hyperspectral remote sensing image classification.

[0003] Hyperspectral remote sensing images, possessing rich spectral and spatial structural information, are widely used in precision agriculture, resource surveys, environmental monitoring, mineral identification, and urban feature classification. Unlike traditional monochrome or RGB images, hyperspectral images typically contain dozens to hundreds of continuous spectral bands, with each pixel corresponding to a complete spectral curve, effectively distinguishing features with similar textures but different material compositions. However, the high dimensionality, spectral redundancy, and spatial correlation of hyperspectral data also present multiple challenges for model design and feature extraction.

[0004] In real-world production scenarios, hyperspectral remote sensing data typically requires manual pixel-by-pixel annotation. This process is not only time-consuming and labor-intensive but also extremely costly, resulting in a limited number of labels for most hyperspectral datasets, often exhibiting a distinct long-tailed class distribution. With very few labeled samples, traditional deep learning models struggle to train effectively, are prone to overfitting, and severely limit their generalization ability.

[0005] To alleviate the problem of insufficient annotation, few-shot learning is introduced into hyperspectral image classification. By constructing a matching method between the support set and the query set, the model can complete the classification task with very few samples. However, most existing few-shot learning methods are geared towards the natural image domain. When directly transferred to hyperspectral images, they encounter three major difficulties: (1) The discriminative information of hyperspectral images not only comes from hundreds of dimensions of spectral curves, but also includes multi-level features such as spatial structure, texture, and neighborhood shape. Existing methods focus on spectral features or two-dimensional spatial features, which cannot maintain the effective coordination of spectral and spatial dimensions in the case of few samples, resulting in incomplete feature representation. (2) Land cover categories in hyperspectral images have different spatial structures at different scales. For example, small areas of vegetation and large areas of farmland have significant differences in spatial range and texture variation. Single-scale convolution is difficult to capture local and global structural information at the same time. Existing few-shot hyperspectral methods are relatively weak in multi-scale feature extraction and do not make sufficient use of cross-scale spectral-spatial dependencies. (3) Hyperspectral data have obvious differences in spectral distribution across scenes. Different sensors, acquisition time, land cover humidity, and atmospheric conditions will all lead to inconsistent data distribution. With fewer samples, this distributional difference is even more difficult to overcome, causing a significant drop in the performance of traditional methods when migrating between different scenarios.

[0006] In summary, existing few-sample hyperspectral classification methods have significant limitations in joint spectral-spatial feature extraction and multi-scale feature fusion, making it difficult to meet the hyperspectral classification requirements of real-world scenarios characterized by "few labels, large differences, and high dimensionality." Therefore, there is an urgent need for a novel hyperspectral classification method that can fully extract spectral-spatial information under few-sample conditions and possesses cross-scenario generalization capabilities. Summary of the Invention

[0007] To address the aforementioned problems, this invention provides a method and system for classifying few-sample hyperspectral remote sensing images based on multi-path evolution.

[0008] Firstly, the present invention provides a few-sample hyperspectral remote sensing image classification method based on multi-path evolution, which adopts the following technical solution:

[0009] A few-sample hyperspectral remote sensing image classification method based on multi-path evolution includes:

[0010] Acquire hyperspectral three-dimensional data;

[0011] The acquired hyperspectral three-dimensional data are preprocessed;

[0012] Preliminary spectral-spatial features of preprocessed data were extracted based on spectral-spatial decoupled convolution.

[0013] Based on preliminary spectral-spatial characteristics, multi-scale deep feature representations are obtained using an asymmetric multi-scale residual mechanism.

[0014] Based on the obtained multi-scale deep feature representation, spectral attention weights are generated by spectral compression reconstruction. The weighted features are then subjected to three-dimensional convolution evolution based on adaptive multi-path feature evolution to obtain information fusion features.

[0015] The information fusion features of the three paths are weighted and fused using a dynamic gating mechanism;

[0016] The hyperspectral image classification results are obtained based on the fused features.

[0017] Furthermore, the preprocessing of the acquired hyperspectral 3D data includes spectral dimension normalization of the source domain data and the target domain dataset, linearly normalizing the pixel values ​​of each band to a uniform range to reduce numerical deviations caused by different sensors and acquisition environments. Subsequently, a corresponding spectral-spatial cube is constructed at each labeled pixel position using a fixed 9x9 spatial neighborhood clipping method for subsequent feature extraction. To achieve cross-domain feature alignment, the hyperspectral cubes of each dataset are uniformly mapped to the same feature dimension space. The mapping layer consists of a 1×1 two-dimensional convolution with the number of input channels being the original number of spectral channels c and the number of output channels fixed at 100, thereby projecting the 9×9×c source domain or target domain spectral cube into a 9×9×100 data structure with a uniform spectral dimension.

[0018] Furthermore, the preliminary spectral-spatial features extraction based on spectral-spatial decoupling convolution of preprocessed data includes first performing spectral-spatial decoupling convolution on the preprocessed hyperspectral data to extract spectral and spatial features respectively. The spectral-spatial decoupling convolution module contains two branches: a spectral feature extraction branch and a spatial feature extraction branch. The spectral branch performs convolution operations in the spectral dimension with a kernel size of 3×1×1 to capture spectral variation information between continuous bands. The spatial branch performs convolution operations in the spatial dimension with a kernel size of 1×3×3 to extract spatial structure information of local regions. The calculation formula is as follows:

[0019]

[0020]

[0021] in, Indicates spectral characteristics, Indicates spatial characteristics, The input is hyperspectral data; then the two types of features are concatenated along the channel dimension to form a joint feature representation containing complementary information. The concatenated features are batch normalized to stabilize the training process, and the nonlinearity of feature expression is enhanced by the ReLU activation function, thus obtaining the preliminary spectral-spatial feature representation of the output.

[0022] Furthermore, the acquisition of multi-scale deep feature representations based on the asymmetric multi-scale residual mechanism includes obtaining preliminary spectral-spatial fusion features. The input is fed into an asymmetric multi-scale residual module, which fuses deep features at three scales—global, spectral, and spatial—through a multi-level, asymmetric convolutional structure. Residual connections are then used to improve feature transfer efficiency. Specifically, the input features are first subjected to a 3D convolution operation. Subsequently, the output undergoes batch normalization and non-linear activation to stabilize training and enhance feature representation capabilities. This is represented as follows:

[0023] ,

[0024] Then The data is fed into a second convolutional unit, which extends the receptive field in the spectral dimension while minimizing it in the spatial dimension. This allows for a focus on capturing the changing trends in spectral direction and cross-band correlations. Batch normalization and activation operations are also performed to improve the discriminability of spectral direction features, as shown below:

[0025] ,

[0026] Next The input is a third convolutional unit, which expands the spatial dimension using its kernel while maintaining a compact spectral dimension. This is used to capture local spatial structure, texture details, and spatial context information, and is represented as follows:

[0027] ,

[0028] Finally, residual connection paths are introduced and processed by an activation function to obtain the final multi-scale deep spectral-spatial feature representation, expressed as: ,

[0029] in, This represents the activation function. This indicates a batch normalization operation. These are the deep features output by the asymmetric multi-scale residual module.

[0030] Furthermore, the step of generating spectral attention weights by spectral compression reconstruction based on the acquired multi-scale deep feature representation includes the deep feature tensor output by the asymmetric multi-scale residual module. For input, where It refers to the batch size. It is the number of channels. It is the spectral dimension. and The spectral dimensions are represented by the spatial height and width, respectively. A 1×1 convolutional layer is used to convert the spectral dimensions from... Compress to Meanwhile, the ReLU activation function is used to enhance the nonlinear representation capability; then the compressed features are reconstructed back to the original spectral dimension through a 1×1 convolutional layer. Normalized adaptive spectral weights are generated using the Sigmoid activation function, and the formula for spectral compression-reconstruction is as follows:

[0031] ,

[0032] in, and These represent the weight matrices for compressing and reconstructing the convolution, respectively. This represents the Sigmoid activation function. This is the generated adaptive spectral weights; finally, the generated weights are then... With original features Element-wise multiplication achieves adaptive weighting along the spectral dimension, represented as: .

[0033] Furthermore, the step of performing 3D convolution evolution on the weighted features based on adaptive multi-path feature evolution includes converting the weighted features... The input is an adaptive multi-path feature evolution module, which utilizes three feature evolution paths—short-term, medium-term, and long-term—containing 3D convolutional branches with different dilation rates to extract local, mesoscale, and global correlation information. Specifically, the short-term path is used to extract local features, including the dilation rate. The convolution output is denoted as: The intermediate path is used to extract features from the mesoscale region, with an inflation rate d=2. The convolution output is denoted as: ; Utilize long-term paths to capture global correlation information, expansion rate The convolution output is denoted as: .

[0034] Furthermore, the three-dimensional convolutional evolution of weighted features based on adaptive multi-path feature evolution also includes, to enhance the discriminative power and spatial-channel representation power of each path feature, outputting short-term, medium-term, and long-term paths respectively. , Channel attention and spatial attention are applied sequentially, and the convolution output of each path is processed. The input channel attention module generates channel weight mappings through adaptive learning. Then, the channel weights are multiplied by the path features channel by channel to obtain the channel-enhanced features, represented as: In channel enhancement features Based on this, a spatial attention module is introduced to generate spatial weight mappings. This is used to emphasize feature regions that are important in both spatial and spectral dimensions. The spatial weights are multiplied element-wise with the channel-enhanced features to obtain the multipath features after joint channel and spatial attention enhancement, expressed as: After channel and spatial attention weighting, the three paths generate multi-scale evolutionary features for the short, medium, and long term, respectively. .

[0035] Furthermore, the weighted fusion of information fusion features from the three paths via a dynamic gating mechanism includes obtaining multi-scale feature representations for the short-term, medium-term, and long-term paths. Subsequently, a dynamic gating fusion mechanism is introduced. Path weights are generated through learnable gating mappings, thereby dynamically adjusting the contribution of different paths to the fusion result. First, the features of the three paths are combined along the channel dimension or feature dimension to form a joint feature representation for gating decision-making, expressed as:

[0036] To learn the importance of each path, The input is a dynamic gating network with a backbone structure of 1×1×1 3D convolutional maps as the main structure, used to learn the weights of different paths contributing to the final decision. The gating network first outputs the unnormalized path weights. , is represented as: in, The weights are 1×1×1 three-dimensional convolutional kernels used to extract path discriminative information. As a bias term; then, on the path dimension... After performing Softmax normalization, it is converted into a dynamic weight vector that can be used for adaptive fusion, represented as: in The weight mappings correspond to the short-term, medium-term, and long-term paths, respectively. Based on the weights, the features of the three paths are fused element-wise to achieve adaptive integration of cross-scale features, as shown below:

[0037] ,

[0038] Furthermore, the weighted fusion of information fusion features from the three paths via a dynamic gating mechanism also includes, to improve the stability and expressiveness of the fused features, inputting the fused features into a channel projection layer with a 1×1×1 three-dimensional convolution as its core. The final enhanced fused representation is generated through batch normalization and activation functions. To ensure gradient stability and feature consistency, residual connections are introduced during the process, as follows:

[0039]

[0040] in, This is either an identity mapping or a residual branch after necessary dimensionality reduction. Using the activation function, the final fused features are obtained. The input is fed into the classification prediction module to generate the final category determination result for the hyperspectral image.

[0041] Secondly, a few-sample hyperspectral remote sensing image classification system based on multi-path evolution includes:

[0042] The data acquisition module is configured to acquire hyperspectral three-dimensional data;

[0043] The preprocessing module is configured to preprocess the acquired hyperspectral three-dimensional data;

[0044] The feature extraction module is configured to extract preliminary spectral-spatial features from the preprocessed data based on spectral-spatial decoupled convolution.

[0045] The feature representation module is configured to obtain multi-scale deep feature representations based on the preliminary spectral-spatial features and an asymmetric multi-scale residual mechanism.

[0046] The feature fusion module is configured to generate spectral attention weights by spectral compression reconstruction based on the acquired multi-scale deep feature representations, and perform three-dimensional convolution evolution on the weighted features based on adaptive multi-path feature evolution to obtain information fusion features.

[0047] The weighted fusion module is configured to perform weighted fusion of the information fusion features of the three paths through a dynamic gating mechanism.

[0048] The classification module is configured to obtain hyperspectral image classification results based on the fused features.

[0049] Thirdly, the present invention provides a computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the described method for classifying few-sample hyperspectral remote sensing images based on multi-path evolution.

[0050] Fourthly, the present invention provides a terminal device, including a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to provide a few-sample hyperspectral remote sensing image classification method based on multi-path evolution.

[0051] In summary, the present invention has the following beneficial technical effects:

[0052] This invention proposes a few-sample hyperspectral image classification method based on multi-path feature evolution. It achieves refined preliminary feature extraction of hyperspectral information by performing separate modeling of spectral and spatial directions on the input hyperspectral 3D cubic data. An asymmetric multi-scale residual structure is introduced into the extracted features to obtain deep feature representations across multiple scales and receptive fields. Addressing the problems of high dimensionality, high redundancy, and easy overfitting under few-sample conditions in hyperspectral data, a spectral compression and reconstruction module is designed to compress and reconstruct the spectral dimensions, generating adaptive spectral attention weights to suppress redundant bands and highlight key spectral information.

[0053] Through the adaptive multi-path feature evolution module, local regions, regional contexts, and global dependency information are modeled from three paths: short-term, medium-term, and long-term. Combined with channel attention, spatial attention, and soft fusion dynamic gating, adaptive evolution and cross-path selective fusion of multi-scale features are achieved. The deep features after the above multi-level feature encoding are used for final classification inference, thereby achieving high-precision hyperspectral image classification under conditions of very few samples. Attached Figure Description

[0054] Figure 1 This is a schematic diagram of a few-sample hyperspectral remote sensing image classification method based on multi-path evolution according to Embodiment 1 of the present invention;

[0055] Figure 2 This is a schematic diagram of the specific structure of the spectral-spatial decoupling convolution module in Embodiment 1 of the present invention;

[0056] Figure 3 This is a schematic diagram of the specific structure of the spectral compression and reconstruction module in Embodiment 1 of the present invention;

[0057] Figure 4 This is a schematic diagram of the specific structure of the adaptive multi-path feature evolution module in Embodiment 1 of the present invention;

[0058] Figure 5 This is a schematic diagram illustrating the visualization of the classification results of this invention in the Indian Pines dataset;

[0059] Figure 6 This is a schematic diagram of the visualization classification results of this invention in the Pavia University dataset;

[0060] Figure 7 This is a schematic diagram of the t-SNE visualization mapping results of the deep features extracted from the Pavia University and Indian Pines datasets in this invention. Detailed Implementation

[0061] The present invention will be further described in detail below with reference to the accompanying drawings.

[0062] Example 1

[0063] Reference Figure 1 This embodiment of a few-sample hyperspectral remote sensing image classification method based on multi-path evolution includes:

[0064] S1. Data Acquisition

[0065] This invention utilizes hyperspectral remote sensing data in a cross-domain, few-shot learning scenario, where the data source consists of a source domain and a target domain. The source domain provides sufficient and diverse labeled samples to support the model in learning generalizable spectral-spatial features; the target domain, with very few labeled samples, is used to complete the final hyperspectral image classification task. This invention selects publicly available hyperspectral datasets as research objects, with the Chikusei dataset as the source domain and the Indian Pines and Pavia University datasets as the target domains. These datasets were all acquired by airborne hyperspectral imaging sensors, covering agricultural, urban, and various mixed land cover scenarios, possessing tens to hundreds of continuous spectral channels, with a spectral range of approximately 400–2500 nm, comprehensively reflecting the spectral response characteristics of different land cover types.

[0066] S2. Data Preprocessing

[0067] In cross-domain few-shot learning tasks, this invention performs preprocessing steps such as standardization, cropping, and feature dimension mapping on the original hyperspectral data to ensure that the source and target domain data can be aligned and compared in a unified feature space.

[0068] S2.1 Data Standardization and Sample Cube Construction

[0069] First, spectral dimension normalization was performed on both the source domain Chikusei data and the target domain datasets (Indian Pines, Pavia University) to linearly normalize the pixel values ​​of each band to a uniform range, thereby reducing numerical biases caused by different sensors or acquisition environments. Then, a corresponding spectral-spatial cube was constructed at each labeled pixel location using a fixed window (e.g., 9×9) spatial neighborhood cropping method for subsequent feature extraction.

[0070] Specifically, firstly, for each spectral band in the original hyperspectral image... (A total of B bands) Band-by-band normalization is performed to eliminate differences in response scale between different sensors. The calculation formula for this process is:

[0071] ,

[0072] in Indicates the source or target domain at the th The average value of each band, The corresponding standard deviation is... These are the standardized band pixels. This operation effectively improves cross-domain spectral alignment capabilities.

[0073] In actual hyperspectral data, certain bands may be affected by water vapor absorption, sensor noise, or saturation. This invention automatically removes noisy bands based on the signal-to-noise ratio (SNR) or variance threshold. The calculation formula for this process is as follows:

[0074] ,

[0075] like If the variance is less than the variance threshold, the band is removed. This yields the effective spectral dimensions.

[0076] The source domain (e.g., Chikusei) and the target domain (e.g., Indian Pines) have different numbers of spectral bands, requiring spectral dimension mapping to bring them into a unified feature space. The calculation formula for aligning the source domain spectrum to the target domain spectrum is as follows:

[0077] ,

[0078] in for The original spectrum, For convolution kernel, For target domain wavelength sampling points, This is the new aligned spectrum. This step ensures that the cross-domain spectral correspondences are consistent.

[0079] At each labeled pixel location Extract its spatial neighborhood and stack the spectral dimensions to form a cube input. Assume the window size is... The present invention adopts For image data ,

[0080] ,

[0081] in Final cube size Use mirror padding or zero padding for edge pixels to ensure consistent dimensions.

[0082] S2.2 Unified Dimension Mapping

[0083] Due to the differences in the number of spectral channels across different datasets, to achieve cross-domain feature alignment, the hyperspectral cubes of each dataset are first uniformly mapped to the same feature dimension space. The mapping layer consists of 1×1 two-dimensional convolutions, with the input channel number being the original spectral channel number *c*, and the output channel number fixed at 100. This projects the 9×9×c source or target domain spectral cube into a 9×9×100 data structure with a unified spectral dimension. This mapping process effectively mitigates the structural differences caused by the inconsistency in spectral dimensions between different datasets, laying the foundation for subsequent cross-domain feature learning.

[0084] Specifically, the mapping layer uses a 1×1 two-dimensional convolution operation, which operates only on the spectral dimension channels and does not change the spatial dimension. The input cube is defined as... ,in , The original number of channels, and the output dimension after unified mapping are: ,in The unified spectral dimension is defined for this invention.

[0085] For each spatial location (i,j) of the input cube, the numerical calculations for the k output spectral channels are as follows:

[0086] ,

[0087] in As a bias, it is projected onto a unified dimension after a linear transformation. .

[0088] S3. Extract preliminary features

[0089] This invention first performs spectral-spatial decoupled convolution on the preprocessed hyperspectral data to extract spectral and spatial features separately. Specifically, the spectral-spatial decoupled convolution module includes two branches: a spectral feature extraction branch and a spatial feature extraction branch. The spectral branch performs convolution in the spectral dimension with a kernel size of 3×1×1 to capture spectral variation information between consecutive bands; the spatial branch performs convolution in the spatial dimension with a kernel size of 1×3×3 to extract spatial structure information of local regions. The calculation formula for this process is as follows:

[0090]

[0091]

[0092] in, Indicates spectral characteristics, Indicates spatial characteristics, The input is hyperspectral data.

[0093] Subsequently, the two types of features are concatenated along the channel dimension to form a joint feature representation containing complementary information. The concatenated features are then batch normalized to stabilize the training process, and the nonlinearity of feature representation is enhanced using the ReLU activation function, thus obtaining the preliminary spectral-spatial feature representation output for this stage. The calculation formula for this process is as follows:

[0094]

[0095] in, For the extracted features, This indicates normalization.

[0096] Reference Figure 2 After preprocessing and unified dimensionality mapping, a 9×9×100 spectral-spatial cube is input into the spectral-spatial decoupled convolution module to achieve joint modeling of spectral dependencies and spatial structure information, thereby obtaining a preliminary explicit representation suitable for cross-domain few-shot learning.

[0097] S3.1 Spectral Correlation Modeling

[0098] Adjacent bands in hyperspectral data exhibit high correlation, while distant bands show weaker correlation. To explicitly learn the dependency structure between bands, this invention constructs a spectral map convolutional layer, modeling relationships along the spectral dimension through learnable linear mappings and normalization operations. Given input features... ,in The batch size, For the number of channels, For spectral dimensions, For spatial dimensions.

[0099] First, expand the spectral dimensions:

[0100]

[0101] The spectral image convolution operation is as follows:

[0102]

[0103] in , For learnable spectral mapping, For LayerNorm, The GELU activation function is used. The result is then restored to 3D feature form. .

[0104] S3.2 Spatial Structure Modeling

[0105] To further extract spatial structure features, depthwise separable convolution is performed on the features enhanced by spectral correlation:

[0106] Depth-wise convolution (channel-independent):

[0107] ,

[0108] By performing 3×3 spatial convolution on each channel independently, local spatial texture and neighborhood structure information are captured.

[0109] Pointwise convolution (channel blending)

[0110] ,

[0111] Spatial features are fused along the channel dimension to generate a spectral-spatial feature map with higher expressive power.

[0112] S3.3 Dynamic Attention Recalibration

[0113] To improve the model's adaptability and robustness to cross-domain data, a dynamic recalibration mechanism similar to channel attention is introduced.

[0114] First, global statistics are obtained through global pooling. The calculation formula for this process is as follows:

[0115] ,

[0116] Dynamic weights are then generated through two convolutional layers. The calculation formula for this process is as follows:

[0117] ,

[0118] Finally, the features are recalibrated, and the calculation formula for this process is as follows:

[0119] ,

[0120] in It is the Sigmoid activation function. This is for element-wise multiplication.

[0121] Finally, a preliminary spectral-spatial feature representation is obtained through BatchNorm and GELU activation. The calculation formula for this process is as follows:

[0122] .

[0123] S4. Input the extracted preliminary features into the asymmetric multi-scale residual module to obtain multi-scale deep feature representations across different receptive fields.

[0124] To further enhance the model's ability to represent cross-receptive field spectral-spatial information, the preliminary spectral-spatial fusion features obtained in step S3 will be used... The input is fed into the asymmetric multi-scale residual module. This module fuses deep features at three scales—global, spectral, and spatial—through a multi-level, asymmetric convolutional structure, and combines residual connections to improve feature transfer efficiency.

[0125] First, a standard 3D convolution operation is performed on the input features. This convolution kernel covers both spectral and spatial dimensions, extracting global fundamental features across dimensions. Subsequently, the output undergoes batch normalization and non-linear activation to stabilize training and enhance feature representation capabilities. The formula for this process is:

[0126] ,

[0127] Will The data is fed into a second convolutional unit. This convolution's receptive field is lengthened in the spectral dimension while being minimized in the spatial dimension, thus focusing more on capturing trends in spectral orientation and cross-band correlations. This stage also includes batch normalization and activation operations to improve the discriminability of spectral orientation features. The formula for this process is:

[0128] ,

[0129] Next, Input the third convolutional unit. This convolutional kernel expands in the spatial dimension but remains compact in the spectral dimension, thus capturing local spatial structure, texture details, and spatial context information. The formula for this process is:

[0130] ,

[0131] To avoid the deep feature degradation problem caused by sequential convolution and to ensure that the input and output maintain dimensionality consistency, a residual connection path is introduced and processed by an activation function to obtain the final multi-scale deep spectral-spatial feature representation. The formula for this process is:

[0132] ,

[0133] in, This represents the activation function. This indicates a batch normalization operation. These are the deep features output by the asymmetric multi-scale residual module.

[0134] S5: Deep features are input into the spectral compression and reconstruction module. Spectral attention weights are generated through spectral compression and reconstruction, and the features are then spectrally weighted. The weighted features are then fed into the cross-scale feature pyramid module to enhance classification and representation capabilities. Next, the adaptive multi-path feature evolution module performs 3D convolutional evolution with different dilation rates in the short-term, medium-term, and long-term paths, respectively, and introduces channel attention and spatial attention to enhance the features of each path, resulting in information fusion features with local, mesoscale, and global correlations.

[0135] S5.1 Spectral Compression-Reconstruction

[0136] Deep feature tensors output by asymmetric multi-scale residual modules For input, where It refers to the batch size. It is the number of channels. It refers to the spectral dimension (number of bands). and These refer to the space's height and width, respectively.

[0137] Reference Figure 3 ,

[0138] Spectral compression: The spectral dimensions are reduced from [previous dimension] to [current dimension] using a 1×1 convolutional layer. Compress to Meanwhile, the ReLU activation function is used to enhance the nonlinear representation capability.

[0139] Spectral reconstruction: The compressed features are then reconstructed back to the original spectral dimensions using a 1×1 convolutional layer. Normalized adaptive spectral weights are generated using the Sigmoid activation function. The formula for spectral compression-reconstruction is:

[0140] ,

[0141] in, and These represent the weight matrices for compressing and reconstructing the convolution, respectively. This represents the Sigmoid activation function. It is the generated adaptive spectral weight.

[0142] The generated weights will then be used With original features Element-wise multiplication achieves adaptive weighting along the spectral dimension. The formula for this process is:

[0143] ,

[0144] S5.2 Cross-Scale Feature Pyramid Module

[0145] Features after spectral compression and reconstruction The system performs bidirectional propagation of downsampling and upsampling, adaptively fuses high / low resolution features through a gating mechanism, and finally obtains enhanced output features through refined convolution and residual short-circuiting.

[0146] First, a 3D convolution pair with a stride of 2 is used. Downsampling, normalization, and nonlinearity are performed to obtain low-resolution semantic features. The calculation formula for this process is as follows:

[0147] ,

[0148] in , , , .

[0149] The downsampling result is then transposed and convolved to restore the resolution, and normalization and activation are also performed. The calculation formula for this process is as follows:

[0150] ,

[0151] like Spatial / spectral size and original input If they are not completely consistent, they are adjusted to match using bitrilinear interpolation. The dimensions are consistent, and the calculation formula for this process is:

[0152] ,

[0153] Original resolution features Features recovered by upsampling The input is stitched together along the channel dimension to form a fused input. The calculation formula for this process is as follows:

[0154] ,

[0155] The fusion weights are generated through a gating network (mapped to [0,1] channel by channel). The calculation formula for this process is as follows:

[0156] ,

[0157] in Indicates the core size is convolution, Activated by Sigmoid, making The value is located in (0,1).

[0158] Based on gating weights Spectral compression reconstruction features With upsampling features The weighted fusion process is calculated using the following formula:

[0159] ,

[0160] Next to The representation is further improved by refining convolutions; the calculation formula for this process is as follows:

[0161] ,

[0162] Finally, a residual join is used to connect the refined result with the original input. The features are fused and then output as a final feature through nonlinear activation. The calculation formula for this process is as follows:

[0163] ,

[0164] S5.3 Adaptive Multi-Path Feature Evolution

[0165] Enhanced features The input is an adaptive multi-path feature evolution module, which includes three feature evolution paths: short-term, medium-term, and long-term. Each path consists of three-dimensional convolutional branches with different dilation rates, which are used to extract local, mesoscale, and global correlation information, respectively.

[0166] Reference Figure 4 The convolutional evolution forms of the three paths are as follows:

[0167] Short-run path (local): expansion rate Used to extract local features, the convolution output is denoted as:

[0168]

[0169] Mid-scale path: dilation rate d=2, used to extract features from mid-scale regions, convolution output denoted as:

[0170]

[0171] Long-term path (global): Expansion rate Used to capture global correlation information, the convolution output is denoted as:

[0172] .

[0173] S5.4 Attention Enhancement

[0174] To further enhance the discriminative power and spatial-channel representation ability of each path feature, the outputs for short-term, medium-term, and long-term paths are respectively... , Channel attention and spatial attention processing are applied sequentially:

[0175] First, output the convolution of each path. The input channel attention module generates channel weight mappings through adaptive learning. This is done to highlight the importance of each channel in the feature representation. Then, the channel weights are multiplied by the path features channel by channel to obtain the channel-enhanced features. The calculation formula for this process is:

[0176]

[0177] Channel enhancement features Based on this, a spatial attention module is introduced to generate spatial weight mappings. This is used to emphasize feature regions that are important in both spatial and spectral dimensions. The spatial weights are multiplied element-wise with the channel-enhanced features to obtain the multipath features after joint channel and spatial attention enhancement. The calculation formula for this process is:

[0178]

[0179] After channel and spatial attention weighting, the three paths generate multi-scale evolutionary features for the short, medium, and long term, respectively. These features not only preserve local, mesoscale, and global information, but also enhance the discriminativeness and robustness of the features through a joint attention mechanism, providing high-quality representations for subsequent feature fusion and classification.

[0180] S6: The features of the three paths are weighted and fused through a dynamic gating mechanism to achieve adaptive integration of multi-path features.

[0181] To obtain multi-scale feature representations of short-term, medium-term, and long-term paths. Subsequently, a dynamic gating fusion module is introduced to achieve adaptive integration of cross-path features. This module generates path weights through learnable gating mappings, thereby dynamically adjusting the contribution of different paths to the final fusion result.

[0182] First, the features of the three paths are combined along either the channel dimension or the feature dimension to form a joint feature representation for gating decisions. The calculation formula for this process is as follows:

[0183]

[0184] To learn the importance of each path, Input a dynamic gating mapping network. This gating network uses 3D convolutional mappings with a kernel size of 1×1×1 as its backbone to learn the weights of different paths' contributions to the final decision. The gating network first outputs unnormalized path weights. The calculation formula for this process is:

[0185]

[0186] in, The weights are 1×1×1 three-dimensional convolutional kernels used to extract path discriminative information. This is a bias term.

[0187] Next, on the path dimension... Softmax normalization is performed to transform it into a dynamic weight vector that can be used for adaptive fusion. The calculation formula for this process is as follows:

[0188]

[0189] in These correspond to the weight mappings for short-term, medium-term, and long-term paths, respectively.

[0190] Based on this weight, the features of the three paths are fused element-wise to achieve adaptive integration of cross-scale features. The calculation formula for this process is as follows:

[0191] ,

[0192] To further improve the stability and expressiveness of the fused features, the fused features are input into a channel projection layer with a 1×1×1 3D convolution as its core. The final enhanced fused representation is generated through batch normalization and activation functions. Simultaneously, residual connections are introduced during this process to ensure gradient stability and feature consistency. The calculation formula for this process is as follows:

[0193]

[0194] in, This is either an identity mapping or a residual branch after necessary dimensionality reduction. The activation function is used. Finally, the fused features are obtained. The input is fed into the classification prediction module to generate the final category determination result for the hyperspectral image.

[0195] S7: Input the fused features into the classification prediction module and output the final hyperspectral image classification result.

[0196] Obtained fusion features The input classification prediction module extracts the global representation of the spectral space through global average pooling, and feeds the extracted global feature vector into a fully connected classifier to perform linear mapping to obtain the prediction score for each category. Then, the prediction score is normalized to obtain the probability distribution of the pixel in each category, and the final classification result is output according to the category corresponding to the maximum probability.

[0197] Specifically, the classification process includes four steps: global feature aggregation, linear mapping, probability calculation, and category decision.

[0198] First, regarding fusion features In spatial dimension ( , ) and spectral dimensions Perform a global average to obtain the global feature vector. The calculation formula for this process is:

[0199]

[0200] global feature vectors Input fully connected layer:

[0201]

[0202] in This is the classification weight matrix. For bias, Predicted score for each category.

[0203] The scores are then mapped to predicted probabilities for each category:

[0204]

[0205] The category with the highest probability is ultimately selected as the final classification result.

[0206] ,

[0207] in This is the final predicted label.

[0208] Experimental verification

[0209] During the meta-training phase, the number of classes in the support set and query set is consistent with the number of classes in the target domain dataset (e.g., 9 classes in the Pavia University dataset and 16 classes in the Indian Pines dataset). Each class in the support set contains 1 sample, and each class in the query set contains 19 samples. Theoretically, the more samples per class in the query set, the stronger the model's generalization ability and the more stable the training results. The methods compared are all state-of-the-art methods for few-shot classification of hyperspectral images. Evaluation metrics include Overall Accuracy (OA), Average Accuracy (AA), and the Kappa coefficient.

[0210] Table 1. Classification results of different methods

[0211]

[0212] Table 1 Figure 5 and Figure 6 The classification results of different hyperspectral remote sensing image classification models on the Indian Pines dataset and the Pavia University dataset are shown, where: (a) is the false-color image of the dataset; (b) is the ground truth label image of the dataset; (c) is the classification result of the Gia-CFSL model; (d) is the model result of the DCFSL model; (e) is the classification result of the SGPL model; (f) is the classification result of the MLPA model; and (g) is the classification result of the model proposed in this invention.

[0213] The Gia-CFSL model is a hyperspectral image classification model that combines graph information aggregation and cross-domain few-shot learning. By introducing intra-domain relation modeling and cross-domain similarity alignment mechanisms, it can effectively improve feature transfer capability and classification accuracy under conditions of scarce samples and large domain differences. However, graph structure construction and similarity calculation increase time overhead, which is not conducive to real-time application in large-scale scenarios.

[0214] The DCFSL model is a cross-domain hyperspectral image classification framework that combines domain adaptation and few-shot learning. However, this model is heavily reliant on adversarial training, its optimization is unstable, and its domain alignment performance is limited when there are large class differences or very few samples in the target domain, which may lead to feature confusion and decreased classification accuracy.

[0215] The SGPL model is a semantically guided prototype learning method that achieves cross-domain, few-sample hyperspectral image classification by incorporating category semantic information into the prototype generation process. However, it relies heavily on semantic encoding, and its performance may decline when generalizing to scenarios without semantic priors or with large category differences.

[0216] The MLPA model is a cross-domain few-shot hyperspectral image classification method that achieves fine-grained matching between the source and target domains through multi-layer feature alignment and prototype-based adversarial training. However, its structure is relatively complex and it is sensitive to parameter settings and alignment levels.

[0217] The proposed model achieves independent modeling of spectral and spatial features through spectral-spatial decoupled convolution, effectively separating spectral correlations from spatial texture structures in the early stages. Subsequently, an asymmetric multi-scale residual structure enhances the model's comprehensive representation of global spectral patterns, local spatial details, and cross-band relationships across different receptive fields. Building upon this, a spectral compression-reconstruction module is introduced to generate spectral attention for adaptive enhancement of the spectral dimension. Furthermore, a cross-scale feature pyramid module enables bidirectional interactive fusion of multi-resolution features, balancing high-resolution details with low-resolution semantic information. Finally, the model utilizes an adaptive multi-path feature evolution mechanism to model local, mesoscale, and global dependencies in short-term, medium-term, and long-term paths, respectively, and achieves dynamic optimization and deep fusion of features through the combined effect of channel and spatial attention. Thanks to the aforementioned multi-stage, cross-scale, and cross-dimensional refined feature modeling strategy, the proposed model achieves significantly higher classification accuracy than traditional methods even on the Indian Pines dataset, which is characterized by high classification difficulty, strong spectral mixing, and weak inter-class differences, fully demonstrating its robustness and generalization ability in complex hyperspectral scenes.

[0218] To intuitively compare the classification performance of the proposed model, this invention further employs t-SNE (t-Distributed Stochastic Neighbor Embedding) to perform a two-dimensional visualization mapping of the deep features extracted by the model. Specifically, the high-dimensional feature representations obtained from the source and target domain samples in the layer preceding the classifier are input into the t-SNE algorithm. By minimizing the difference in probability distribution between the high-dimensional and low-dimensional spaces, the samples maintain their proximity relationships in the original feature space within the two-dimensional plane. The resulting feature distribution map intuitively reflects the cluster compactness and inter-class isolation between different categories. Figure 7 As shown, the model of this invention has significant advantages in enhancing inter-class separability, maintaining intra-class compactness, and aligning cross-domain features.

[0219] Example 2

[0220] This embodiment provides a few-sample hyperspectral remote sensing image classification system based on multi-path evolution, including:

[0221] The data acquisition module is configured to acquire hyperspectral three-dimensional data.

[0222] The preprocessing module is configured to preprocess the acquired hyperspectral three-dimensional data;

[0223] The feature extraction module is configured to extract preliminary spectral-spatial features from the preprocessed data based on spectral-spatial decoupled convolution.

[0224] The feature representation module is configured to obtain multi-scale deep feature representations based on the preliminary spectral-spatial features and an asymmetric multi-scale residual mechanism.

[0225] The feature fusion module is configured to generate spectral attention weights by spectral compression reconstruction based on the acquired multi-scale deep feature representations, and perform three-dimensional convolution evolution on the weighted features based on adaptive multi-path feature evolution to obtain information fusion features.

[0226] The weighted fusion module is configured to perform weighted fusion of the information fusion features of the three paths through a dynamic gating mechanism.

[0227] The classification module is configured to obtain hyperspectral image classification results based on the fused features.

[0228] A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the described method for classifying few-sample hyperspectral remote sensing images based on multipath evolution.

[0229] A terminal device includes a processor and a computer-readable storage medium, the processor being configured to implement various instructions; the computer-readable storage medium being configured to store multiple instructions adapted for loading and execution by the processor of the described method for classifying few-sample hyperspectral remote sensing images based on multi-path evolution.

[0230] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for few-shot hyperspectral remote sensing image classification based on multi-path evolution, characterized in that, The method comprises the following steps: acquiring hyperspectral three-dimensional data; preprocessing the acquired hyperspectral three-dimensional data; extracting preliminary spectral-spatial features of the preprocessed data based on spectral-spatial decoupling convolution; obtaining multi-scale deep feature representation based on an asymmetric multi-scale residual mechanism according to the preliminary spectral-spatial features; generating spectral attention weights by spectral compression reconstruction according to the obtained multi-scale deep feature representation, and performing three-dimensional convolution evolution on the weighted features based on adaptive multi-path feature evolution to obtain information fusion features; performing weighted fusion on the information fusion features of the three paths through a dynamic gating mechanism; obtaining a hyperspectral image classification result according to the fused features. The multi-scale deep feature representation based on the asymmetric multi-scale residual mechanism includes the obtained preliminary spectral-spatial fusion features The input is input into the asymmetric multi-scale residual module. The deep features of three scales of global, spectral and spatial are fused through a multi-level and asymmetric convolution structure, and the feature transmission efficiency is improved by combining residual connection. Specifically, first, a three-dimensional convolution operation is performed on the input features, then the output is normalized by batch and activated by nonlinearity, which is used to stabilize the training and enhance the feature expression ability, represented as: ， Then the is sent into the second convolution unit, which elongates the receptive field in the spectral dimension while keeping it minimal in the spatial dimension, so as to focus on capturing the trend of spectral direction and the cross-band correlation. Batch normalization and activation are also performed to improve the distinguishability of the spectral direction features, which is expressed as: ， Then the The third convolutional unit is inputted, which expands in the spatial dimension with the convolution kernel of the third convolutional unit, but keeps the spectral dimension compact, for capturing local spatial structure, texture details and spatial context information, denoted as: ， Finally, the residual connection path is introduced and processed by the activation function to obtain the final multi-scale deep spectral-spatial feature representation, denoted as: , wherein, denotes an activation function, denotes a batch normalization operation, is a deep feature output by the asymmetric multi-scale residual module; The spectral attention weight is generated by using spectral compression reconstruction according to the obtained multi-scale deep feature representation, including a deep feature tensor output by an asymmetric multi-scale residual module is an input, where is a batch size, is a number of channels, is a spectral dimension, and are a spatial height and a spatial width, respectively, the spectral dimension is compressed from to by a 1x1 convolution layer, and a ReLU activation function is used to enhance a nonlinear representation capability; then the compressed feature is reconstructed to the original spectral dimension by a 1x1 convolution layer, and a Sigmoid activation function is used to generate a normalized spectral adaptive weight, and a formula of the spectral compression-reconstruction is: ， where, and denote the weight matrix of compressed and reconstructed convolution, respectively, denotes the Sigmoid activation function, is the generated spectral adaptive weight; the generated weight is then element-wise multiplied with the original feature to achieve adaptive weighting in the spectral dimension, denoted as: .

2. The method of claim 1, wherein, The preprocessing of the acquired hyperspectral three-dimensional data comprises the following steps: performing spectral dimension normalization processing on source domain data and target domain data sets respectively, linearly normalizing each band pixel value to a unified range, so as to weaken the numerical deviation caused by different sensors and collection environments, and then performing spatial neighborhood clipping in a fixed window 9x9, constructing a corresponding spectral-spatial cube at each labeled pixel position, and using the spectral-spatial cube for subsequent feature extraction; in order to realize cross-domain feature alignment, the hyperspectral cubes of each data set are uniformly mapped to the same feature dimension space, the mapping layer is composed of a two-dimensional convolution with a size of 1x1, the input channel number is the original spectral channel number c, and the output channel number is fixed as 100, so that the source domain or target domain spectral cube with a size of 9x9xc is projected into a 9x9x100 data structure with a unified spectral dimension.

3. The method of claim 2, wherein, The preliminary spectral-spatial features of the preprocessed data are extracted based on spectral-spatial decoupling convolution, which comprises the following steps: first, performing spectral-spatial decoupling convolution on the preprocessed hyperspectral data, and extracting spectral features and spatial features respectively; the spectral-spatial decoupling convolution module comprises two branches: a spectral feature extraction branch and a spatial feature extraction branch, wherein the spectral branch performs convolution operation in the spectral dimension, and the convolution kernel size is set to 3x1x1, which is used to capture the spectral change information between consecutive bands; the spatial branch performs convolution operation in the spatial dimension, and the convolution kernel size is set to 1x3x3, which is used to extract the spatial structure information of the local area, and the calculation formula is: ， ， wherein, denotes spectral features, denotes spatial features, is the input hyperspectral data; subsequently, the two types of features are concatenated along the channel dimension to form a joint feature representation containing complementary information. The concatenated features are batch-normalized to stabilize the training process and activated by a ReLU activation function to enhance the non-linear capability of feature representation, thus obtaining the output preliminary spectral-spatial feature representation.

4. The method of claim 3, wherein, The adaptive multi-path feature evolution is used for three-dimensional convolution evolution of the weighted features, including evolution of the weighted features The input adaptive multi-path feature evolution module uses three short-term, medium-term and long-term feature evolution paths containing three-dimensional convolution branches with different expansion rates to extract local, mesoscale and global correlation information, wherein the short-term path is used to extract local features, the expansion rate The convolution output is recorded as: The medium-term path is used to extract mesoscale regional features, the expansion rate The convolution output is recorded as: The long-term path is used to capture global correlation information, the expansion rate The convolution output is recorded as: .

5. The method of claim 4, wherein, The method of performing 3D convolution evolution on weighted features based on adaptive multi-path feature evolution also includes outputting short-term, medium-term, and long-term paths respectively to enhance the discriminative ability and spatial-channel representation ability of each path feature. , Channel attention and spatial attention are applied sequentially, and the convolution output of each path is processed. The input channel attention module generates channel weight mappings through adaptive learning. Then, the channel weights are multiplied by the path features channel by channel to obtain the channel-enhanced features, represented as: In channel enhancement features Based on this, a spatial attention module is introduced to generate spatial weight mappings. This is used to emphasize feature regions that are important in both spatial and spectral dimensions. The spatial weights are multiplied element-wise with the channel-enhanced features to obtain the multipath features after joint channel and spatial attention enhancement, expressed as: After channel and spatial attention weighting, the three paths generate multi-scale evolutionary features for the short, medium, and long term, respectively. .

6. The method of claim 5, wherein, The information fusion features of the three paths are weighted and fused through a dynamic gating mechanism, including obtaining multi-scale feature representations of the short-term, medium-term and long-term three paths After that, a dynamic gating fusion mechanism is introduced to generate path weights through a learnable gating mapping, so as to dynamically adjust the contribution of different paths to the fusion result. First, the features of the three paths are combined in the channel dimension or the feature dimension to form a joint feature representation for gating decision, denoted as: To learn the importance of each path, The input is a dynamic gating network with a backbone structure of 1×1×1 3D convolutional mappings to learn the weights of different paths contributing to the final decision. The gating network first outputs the unnormalized path weights, represented as: in, The weights are 1×1×1 three-dimensional convolutional kernels used to extract path discriminative information. As a bias term; then, on the path dimension... After performing Softmax normalization, it is converted into a dynamic weight vector that can be used for adaptive fusion, represented as: wherein The weight mapping corresponding to the short-term, medium-term and long-term paths respectively, the characteristics of the three paths are element-wise weighted fused based on the weights, realizing adaptive integration of cross-scale features, expressed as: 。 7. The method of claim 6, wherein, The dynamic gating mechanism for weighting and fusing the information fusion features of the three paths further includes inputting the fusion features into a channel projection layer with a 1×1×1 three-dimensional convolution as a core to improve stability and expressiveness of the fusion features And through batch normalization and an activation function, a final enhanced fusion representation is generated, and in order to ensure gradient stability and feature consistency, a residual connection is introduced in the process, and is expressed as: wherein, is an identity mapping or a residual branch with necessary dimension reduction, is an activation function, and finally, the fused feature The input classification prediction module is used to generate the final class discrimination result of the hyperspectral image.

8. A multi-path evolution based few-shot hyperspectral remote sensing image classification system, which executes a multi-path evolution based few-shot hyperspectral remote sensing image classification method according to claim 1. The method comprises the following steps: The data acquisition module is configured to acquire hyperspectral three-dimensional data; The preprocessing module is configured to preprocess the acquired hyperspectral three-dimensional data; The feature extraction module is configured to extract preliminary spectral-spatial features of the preprocessed data based on spectral-spatial decoupling convolution; The feature representation module is configured to obtain multi-scale deep feature representation based on an asymmetric multi-scale residual mechanism according to the preliminary spectral-spatial features; The feature fusion module is configured to generate spectral attention weights by spectral compression reconstruction according to the obtained multi-scale deep feature representation, and perform three-dimensional convolution evolution on the weighted features based on adaptive multi-path feature evolution to obtain information fusion features; The weighted fusion module is configured to perform weighted fusion on the information fusion features of the three paths through a dynamic gating mechanism; The classification module is configured to obtain a hyperspectral image classification result according to the fused features.

Images