A hyperspectral unmixing method based on multi-domain feature perception fusion

Abstract

This invention discloses a hyperspectral unmixing method based on multi-domain feature perception fusion, comprising the following steps: acquiring hyperspectral data and extracting initial endmembers, which are used as initial weights for convolutional layers in the decoder; extracting spatial features, spectral features, and frequency domain features respectively through parallel spatial feature extraction branches, spectral feature extraction branches, and frequency domain feature extraction branches; performing semantic alignment and complementary advantages on spatial features and spectral features, as well as spatial features and frequency domain features, respectively through a cross-domain dynamic perception module to obtain spatial-spectral fusion features and spatial-frequency fusion features; further fusing the spatial-spectral fusion features and spatial-frequency fusion features through learnable channel-level weight parameters to obtain multi-domain fusion features; inputting the multi-domain fusion features into a Transformer codec for semantic encoding, and then decoding them through a semantically enhanced abundance generation module to obtain an abundance map, and reconstructing the hyperspectral image and the extracted endmember spectral curves; this invention improves the accuracy of hyperspectral unmixing.

Description

Technical Field

[0001] This invention relates to the fields of image processing and deep learning, specifically to a hyperspectral demixing method based on multi-domain feature perception fusion. Background Technology

[0002] Current unmixing research largely focuses on spectral and spatial domain feature modeling. However, due to factors such as noise, relying solely on the original spectral and spatial domains is insufficient to fully extract effective edge and texture details and is easily affected by interference. High-frequency components in the frequency domain can effectively characterize edge and texture details missing in the original domain, while low-frequency components can represent the overall structure of ground objects. Existing methods, however, do not pay enough attention to the edge and texture details contained in frequency domain features, failing to fully explore the complementary relationships between different frequency components. This results in insufficient modeling and representation of image details, and a lack of corresponding smoothing strategies, easily leading to problems such as loss of abundance map details and jitter in endmember spectral curves. Existing methods still lack effective mechanisms for the synergistic utilization of features from different domains (spectral, spatial, and frequency domains), generally exhibiting insufficient cross-domain feature fusion and inadequate utilization of information complementarity. Furthermore, methods for extracting and representing frequency domain texture details for unmixing tasks are still imperfect, urgently requiring the design of more targeted frequency domain modeling and more efficient fusion strategies. Summary of the Invention

[0003] Purpose of the invention: The purpose of this invention is to provide a hyperspectral unmixing method based on multi-domain feature sensing fusion, thereby solving the technical problems of insufficient feature representation and inadequate fusion capability in existing hyperspectral unmixing methods.

[0004] Technical solution: The present invention provides a hyperspectral unmixing method based on multi-domain feature sensing fusion, comprising the following steps:

[0005] (1) Acquire hyperspectral data and extract initial endmembers, which are used as initial weights for the convolutional layers in the decoder;

[0006] (2) Spatial features, spectral features and frequency domain features of hyperspectral data are extracted by parallel spatial feature extraction branches, spectral feature extraction branches and frequency domain feature extraction branches respectively;

[0007] (3) Through the cross-domain dynamic sensing module, semantic alignment and complementary advantages are performed on spatial features and spectral features, as well as spatial features and frequency domain features, respectively, to obtain spatial-spectral fusion features and spatial-frequency fusion features;

[0008] (4) By using learnable channel-level weight parameters to re-fuse the spatial spectrum fusion features and spatial frequency fusion features, redundancy suppression and detail enhancement are achieved, resulting in multi-domain fusion features;

[0009] (5) Input the multi-domain fusion features into the Transformer codec for semantic encoding, and then decode them through the semantic enhancement abundance generation module to obtain the abundance map, and reconstruct the hyperspectral image and extract the endmember spectral curves.

[0010] Furthermore, in step (2), the frequency domain feature extraction branch includes: performing Fourier transform and wavelet transform on the hyperspectral data simultaneously to obtain the global frequency domain structure and multi-directional texture sub-bands respectively; generating a query matrix using the low-frequency sub-band of the wavelet, generating multiple key matrices using the high-frequency sub-band of the wavelet, and generating a value matrix using the Fourier features; guiding the multi-directional texture detail information into the global frequency domain structure through the channel attention mechanism to generate enhanced frequency domain features.

[0011] Furthermore, in step (3), the cross-domain dynamic perception module is as follows: the two types of cross-domain features are split into multiple independent subspaces; a perception matrix is ​​constructed based on element-wise multiplication in each subspace, and its context perception capability is enhanced by a linear layer to generate a suitable context-aware convolution kernel for each type of feature, and neighborhood interaction aggregation is performed at each spatial location to achieve semantic alignment and complementary advantages of cross-domain features; combined with multi-scale convolution detail compensation, local details are enhanced and then fused with the residual of the original feature.

[0012] Furthermore, in step (5), the semantic enhancement abundance generation module specifically involves: dividing the semantic encoding into multiple parts according to the number of endmembers, and performing attention semantic enhancement on each part; mapping the enhanced parts to abundance responses through a linear layer, and processing them through convolutional smoothing and a temperature-controlled activation function to obtain an abundance map that satisfies the non-negativity constraint and the sum-of-one constraint.

[0013] Furthermore, in step (5), in the Transformer encoder and decoder, the encoding stage only performs query projection on the semantic vector, while performing key and value projection on all feature vectors to reduce computational complexity and enhance semantic representation; the decoding stage uses a multi-head attention mechanism to enhance semantic encoding, and then outputs abundance through linear mapping, smoothing and controllable activation.

[0014] Furthermore, the method employs a self-supervised approach during training, with the loss function including reconstruction error loss and spectral angular distance loss, which are weighted together to optimize network parameters.

[0015] The present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the methods described herein.

[0016] An electronic device according to the present invention includes a memory and a processor, wherein the memory stores a computer program, and when the program is executed by the processor, it implements any of the methods described herein.

[0017] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: On the one hand, it overcomes the limitations of relying solely on spatial-spectral original domain features, introducing frequency domain representation to enhance the ability to depict global structural information and detailed texture information, alleviating the problem that single-domain features are difficult to fully express in complex mixed ground feature scenarios; on the other hand, it achieves pairwise alignment and complementary advantages of spatial, spectral, and frequency domain features through a cross-domain dynamic sensing mechanism, and further performs learnable re-fusion to suppress redundant information, highlight effective details, and improve the discriminativeness and stability of fused features; in addition, it uses Transformer to perform semantic modeling of fused features, and combines semantically enhanced abundance generation, smoothing, and controllable temperature activation strategies to make abundance estimation more stable and meet ANC / ASC physical constraints, thereby improving the accuracy and robustness of endmember extraction and abundance estimation. Overall, this invention achieves deep collaborative fusion of multi-domain features, improves the overall performance of hyperspectral unmixing, and provides technical support for refined ground feature information extraction in remote sensing scenarios. Attached Figure Description

[0018] Figure 1 This is a diagram of the overall architecture of the present invention;

[0019] Figure 2 This is a schematic diagram of spatial spectrum feature extraction in the parallel feature extraction module of the present invention;

[0020] Figure 3 This is a schematic diagram of the frequency domain multi-branch attention in the parallel feature extraction module of the present invention;

[0021] Figure 4 This is a schematic diagram of the semantic enhancement abundance generation module of the present invention; Detailed Implementation

[0022] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0023] like Figure 1 As shown, this embodiment of the invention provides a hyperspectral unmixing method based on multi-domain feature sensing fusion, comprising the following steps:

[0024] Step 1: Data Preprocessing

[0025] The acquired hyperspectral data is represented as Where h and w are the spatial dimensions, and c represents the number of spectral bands. In step 1, the traditional VCA unmixing method is mainly used to perform preliminary coarse endmember extraction on the hyperspectral data, resulting in the unmixed endmember matrix represented as follows: And use it as the initial weight for 2D convolutions in the MDFPF-Net decoder layer.

[0026] Step 2: Feature Extraction Module

[0027] Two sets of 2D convolutions are used to extract the spatial and spectral features of the input data X, respectively. Using a basic CNN structure, three 2D convolutions with the same kernel size are employed for feature extraction (each group contains 3 convolutions, with identical kernel sizes within each group; the kernel sizes for each group are 3×3 and 1×1 respectively, and the number of output channels changes from 128 to 64 to 32; the activation function is ReLU), as detailed below:

[0028]

[0029] In the formula, and Let X be the current input data, representing the desired output feature. and Indicates the size of the convolution kernel. Represented as the first The convolution operation.

[0030] Parallel two-dimensional discrete Fourier transforms are performed on each channel of the hyperspectral data, and the amplitudes are extracted to obtain the features after the Fourier transform. Perform a two-dimensional Haar stationary wavelet transform to obtain The four filter outputs are represented as follows:

[0031]

[0032]

[0033] in This indicates the modulo length operation. and These are the spatial coordinates of pixels. and These are the coordinates of the frequency components. The imaginary unit (satisfying) ), and The symbols represent low-pass and high-pass wavelet filters, respectively. This represents the tensor outer product, used to construct two-dimensional filters.

[0034] In Frequency Domain Multi-Branch Attention (FMBA), Fourier global structural features are incorporated. Projected through a 1×1 convolution and reorganized into a value matrix :

[0035]

[0036] Low-frequency subband of wavelet decomposition Mapping to generate query matrix :

[0037]

[0038] The high-frequency subbands in the three directions are then mapped to generate bond matrices. :

[0039]

[0040] in , and Represents a 1×1 convolutional projection. This is a restructuring operation.

[0041] To improve stability, and Global average pooling (GAP) is performed in the spatial dimension to obtain channel descriptors, and attention weights are calculated in the channel domain to fuse detailed texture information from different directions, thereby adaptively optimizing the global frequency domain features to obtain the frequency domain features. :

[0042]

[0043] in This is the dynamic scaling factor. is the scale factor.

[0044] This mechanism utilizes wavelet multi-directional local frequency information to guide Fourier global structural feature enhancement and reduces computational cost with lightweight channel attention.

[0045] Step 3: Cross-domain feature perception and fusion module

[0046] The obtained spatial features and spectral characteristics Frequency domain characteristics The input features are processed in two parallel Dynamic Perception Modules (CDDP). CDDP consists of three steps: processing the two types of cross-domain features... and ,right Perform 1×1 convolution. After performing spatial pooling, a 1×1 convolution is then performed, resulting in... and The perception matrix is ​​obtained by performing simple matrix multiplication within the four subspaces. ,in This represents the number of pixels. Then the perception matrix... Enhanced through linear layer mapping and split into and Context-aware convolution kernels and Large-scale convolutional kernels are primarily used for spatial domain modeling, capturing long-range spatial dependencies and structured information of corresponding pixels and their neighborhoods through a larger receptive field; while small-scale convolutional kernels focus on characterizing subtle spectral changes and differences in ground feature distribution, ensuring accurate representation of spectral and contour details. The two convolutional kernels act on... and This design achieves directional matching between feature type and convolution kernel scale: at each spatial location, convolution operations are performed on neighborhood features using the corresponding convolution kernel, enabling deep interaction aggregation between cross-domain features and fully releasing potential discriminative information, thereby effectively promoting semantic alignment and complementary advantages of cross-domain features. Simultaneously, the input features... and The concatenation is then fed into the Multi-Scale Convolutional Detail Compensation (DSCDC) module, which consists of four sets of 2D convolutions with progressively larger kernels (kernel sizes of 3, 5, 7, and 9), yielding the detail-compensated output. ,Will Proportionally, matrix addition is performed with the aforementioned cross-domain interactive features along the channel dimension to enhance the representation ability of fine structures, preserve the differential discriminative attributes of the fused features, and incorporate multi-scale local detail compensation information from the convolution residuals. The two CDDP branches ultimately generate spatial-spectral features that emphasize endmember spectral discriminativeness. Spatial frequency characteristics that characterize differences in endmember distribution The process is as follows:

[0047]

[0048]

[0049] in, This indicates a splitting operation. This represents a linear transformation operation. and This is the proportionality coefficient. and This represents two types of convolution kernels, one large and one small, generated from spatial features and the other from spectral features. Provide corresponding detailed compensation. and This represents two types of convolution kernels, one large and one small, generated from spatial features and the other from frequency domain features. Provide corresponding detailed compensation.

[0050] Features fused through dynamic perception are used to perform spatial-spectral feature processing via a channel adaptive fusion module with learnable parameters. and spatial frequency characteristics Redundancy suppression and advantage amplification are performed, and the multi-domain fusion features are obtained by mapping the features to the channel dimensions required by the subsequent encoder through 1×1 convolution. The details are as follows:

[0051]

[0052] In the formula For convolution operations involving dimension mapping, The number of endmembers and It is a learnable bias. and These are learnable weight coefficients that can be optimized through continuous iterative updates and learning throughout the model training process to obtain fusion features.

[0053] Step 4: Transformer encoding / decoding module

[0054] The Transformer encoder-decoder architecture is used to semantically encode the fused features in the spatial-spectral-frequency domain, and semantic enhancement abundance generation is used to enhance the semantic encoding, thus completing the demixing task. The Transformer encoder is used to convert the fused features into semantic codes. First, the... Divided in spatial dimension A 5x5 space block, flattened into ,in , And add position encoding. To maintain spatial order; then, learnable semantic vectors are concatenated before the sequence. , obtain the initial input :

[0055]

[0056] in For the first A characteristic token, .

[0057] The coding phase The normalized input is then fed into a multi-head self-attention module. Unlike the standard Transformer, this method only processes semantic tokens. Query projection is performed, while key and value projections apply to the entire sequence. The algorithm calculates scaled dot product attention based on the number of attention heads; the outputs of each head are concatenated and linearly mapped, and training is stabilized using residual connections and layer normalization. Then, an MLP is used to further enhance the nonlinear representation capability. Finally, the updated semantic token is... By fusing the average of all feature tokens, a global semantic encoding is obtained. :

[0058]

[0059] This semantic encoding simultaneously integrates global semantics and local dependency information, providing a more discriminative semantic representation for subsequent decoding.

[0060] Then, endmembers and abundance are extracted from the semantic code, and the encoded data enters the Semantic Enhancement Abundance Generation (SEAG) module, which divides the semantic code into p parts to obtain... Then, attention semantic enhancement is performed on each part, and then it is mapped to the initial abundance response through a linear layer. The specific process is as follows:

[0061]

[0062] in This represents the total number of pixels, where p indicates the number of endmembers. This represents the abundance vector at the i-th pixel. Represents a linear layer operation, while Attention is a semantic enhancement for channel attention.

[0063] Next, a convolutional smoothing operation and a temperature-controlled activation function TSoftmax are applied. Optimization and lightweight constraints were applied to improve the generated abundance map. Satisfying ANC (abundance nonnegativity constraint) and ASC (abundance and uniqueness constraint), and then performing 2D convolution on the abundance matrix, we obtain the reconstructed hyperspectral data. At this point, the weights extracted from the convolutional layer are the predicted endmember spectral matrix. The specific calculation process is as follows:

[0064]

[0065]

[0066] in The temperature coefficient and , It is a convolution smoothing operation. This is an operation to retrieve the weights of the convolutional layer. Represents the natural base.

[0067] Step 5: Loss Function Design

[0068] To train the proposed unmixing model, this invention employs a self-supervised optimization method that does not rely on endmembers or ground truth abundance: the network optimizes the input hyperspectral image... The generated reconstructed image and only by measuring and The differences between the parameters are used to backpropagate and update the parameters. This is done to balance pixel reconstruction accuracy with spectral shape consistency.

[0069] The composite loss function is defined as a weighted sum of reconstruction error (RE) loss and spectral angular distance (SAD) loss:

[0070] (5.1) Reconstruction error (RE) loss:

[0071]

[0072] in and This is a scalar hyperparameter used to balance the weights of the two losses. The reconstruction error (RE) loss is as follows:

[0073]

[0074] in, and Representing the spatial location in the original image and the reconstructed image, respectively. Upper The spectral values ​​at each band are constrained to ensure that the reconstructed image approximates the original image in each band, guaranteeing the overall reconstruction accuracy and providing a stable self-supervised signal for the learning of endmembers and abundance.

[0075] (5.2) Spectral angular distance (SAD) loss:

[0076]

[0077] in, and The original image and the reconstructed image are located in spatial positions, respectively. Spectral vector on; Indicates the inner product. Represents the L2 norm, To prevent extremely small constants with a denominator of zero, this term constrains the shape consistency between the reconstructed spectrum and the original spectrum from a spectral perspective, reducing the impact of amplitude variations and enhancing the robustness of the model under complex conditions such as noise and brightness changes.

[0078] Step 6: Training Design

[0079] The experiments used publicly available datasets: the synthetic dataset, the Samson dataset, and the Jasper Ridge dataset. Both training and testing were performed on a computer configured with an Intel Core i7-12700H CPU (2.30 GHz), 16 GB of RAM, and a GEFORCE RTX4060 GPU. Network parameters were updated using the Adam optimizer to ensure stable convergence during training, with the learning rate decaying by a coefficient every 20 training epochs. The optimization function included a weight decay rate, and the network was trained using the Kaiming initialization method. The final 2D convolutions of the decoder part of the network were configured with weights in the range (0,1).

[0080] Experimental results:

[0081] Hyperparameter settings

[0082] During the experiment, the hyperparameter settings for each dataset are shown in Table 1 below.

[0083] Table 1 Hyperparameter Settings ;

[0084] Table and These are the weight parameters applied to the loss function. and It is the weight parameter between the main feature and the residual in cross-domain dynamic perception.

[0085] Evaluation indicators

[0086] To quantitatively evaluate the performance of various hyperspectral image unmixing methods, this embodiment uses root mean square error (RMSE) and spectral angular distance (SAD) as evaluation metrics for abundance estimation and endmember estimation for each land cover category, respectively. Furthermore, average root mean square error (aRMSE) and average spectral angular distance (aSAD) are used as global evaluation metrics, calculated as follows:

[0087]

[0088]

[0089]

[0090]

[0091] in and Representing the first Abundance matrix corresponding to each endmember and For the content corresponding to the spatial position (i, j) in the abundance matrix, and Representing the first The endmember curve matrix corresponding to each endmember.

[0092] Comparative experiment

[0093] As shown in Table 2, the proposed MDPF-Net is compared with traditional methods VCA-FCLS and SGSNMF, and deep learning methods DAEN, CyCU, MiSiC, DeepTrans, A2SAN, MAT, and SSAF on a synthetic dataset and two real datasets. All methods use vertex component analysis (VCA) to initialize endmembers.

[0094] Table 2 Comparison of experimental results on synthetic datasets

[0095] Table 2 shows a comparison of the unmixing results of different methods on the synthetic dataset. As can be seen, the MDFPF-Net proposed in this invention achieves the best performance in both overall average metrics: the lowest aRMSE is 0.0188 and the lowest aSAD is 0.0122, indicating that this method has advantages in both abundance estimation accuracy and endmember spectral consistency.

[0096] Based on the abundance estimation results, MDFPF-Net achieved the lowest error in the four categories of asphalt, coniferous trees, basalt, and concrete, demonstrating a stronger ability to fit different land cover mixing ratios. Although it was not the best in the limestone category, it still maintained a level close to the best, with a small difference from the optimal value. Overall, it performed stably without any obvious shortcomings.

[0097] As shown in Table 3, based on the endmember estimation results, MDFPF-Net achieved the lowest spectral angle error for concrete and limestone endmembers, indicating that it more accurately preserves the spectral shape of the endmembers. While the error for the other categories may not be the absolute minimum, it is at a suboptimal level. Overall, MDFPF-Net has significant advantages in spectral discrimination, anti-interference ability, and robustness in abundance inversion.

[0098] Table 3 shows the comparative experimental results on the Samson dataset.

[0099] Table 3 presents a comparison of the unmixing results of different methods on the Samson dataset. As can be seen, MDFPF-Net has the most significant advantage in overall average metrics: aRMSE=0.0507 and aSAD=0.0393, both of which are at the best level. This indicates that the method achieves a better overall balance between abundance inversion accuracy and endmember spectral consistency, and its overall performance is superior and more stable.

[0100] In terms of abundance estimation error, MDFPF-Net achieved the lowest RMSE for trees and soil, demonstrating more accurate inversion of the mixing ratio of major land features; it also maintained a low error level for water body categories. The comprehensive average index further indicates that this method has a stronger overall advantage in abundance estimation across the three categories.

[0101] From the perspective of endmember estimation error, MDFPF-Net achieves the lowest spectral angle error across the entire table for tree endmembers, significantly improving its ability to preserve the spectral shape of endmembers. It also maintains a low spectral angle difference for water and soil endmembers, indicating that it can more accurately characterize the essential spectral features of endmembers and has stronger anti-interference capabilities.

[0102] Table 4 shows the comparative experimental results on the Jasper Ridge dataset.

[0103] Table 4 presents a comparison of the unmixing results of different methods on the Jasper Ridge dataset. It can be seen that MDFPF-Net performs best in terms of overall average spectral angle: aSAD=0.0631, the lowest in the table, indicating that this method preserves the endmember spectral shape more accurately and has stronger spectral discrimination ability. Simultaneously, it also achieves a superior overall performance in abundance inversion, demonstrating good comprehensive unmixing ability and stability.

[0104] In terms of abundance estimation error, MDFPF-Net achieved the lowest RMSE in the three categories of trees, water bodies, and roads, indicating that it is more accurate in retrieving the mixed proportion of multiple land cover types and more adaptable to changes in the distribution of different land cover types; and it maintains a leading position in the overall average index, demonstrating more stable abundance estimation performance.

[0105] From the perspective of endmember estimation error (SAD), MDFPF-Net achieves the lowest spectral angle error in both water and soil endmembers, indicating that it more accurately characterizes the essential spectral features of key endmembers. This further verifies the advantages of the proposed multi-domain feature fusion and semantic modeling strategy in improving the reliability and anti-interference ability of endmember estimation.

Claims

1. A hyperspectral unmixing method based on multi-domain feature sensing fusion, characterized in that, Includes the following steps: (1) Acquire hyperspectral data and extract initial endmembers, which are used as initial weights for the convolutional layers in the decoder; (2) Spatial features, spectral features and frequency domain features of hyperspectral data are extracted by parallel spatial feature extraction branches, spectral feature extraction branches and frequency domain feature extraction branches respectively; (3) Through the cross-domain dynamic sensing module, semantic alignment and complementary advantages are performed on spatial features and spectral features, as well as spatial features and frequency domain features, respectively, to obtain spatial-spectral fusion features and spatial-frequency fusion features; (4) The spatial spectrum fusion feature and the spatial frequency fusion feature are re-fused by learnable channel-level weight parameters to obtain multi-domain fusion features; (5) Input the multi-domain fusion features into the Transformer codec for semantic encoding, and then decode them through the semantic enhancement abundance generation module to obtain the abundance map, and reconstruct the hyperspectral image and extract the endmember spectral curves.

2. The hyperspectral unmixing method based on multi-domain feature sensing fusion according to claim 1, characterized in that, In step (2), the frequency domain feature extraction branch includes: performing Fourier transform and wavelet transform on the hyperspectral data simultaneously to obtain the global frequency domain structure and multi-directional texture sub-bands respectively; generating a query matrix using the low-frequency sub-band of the wavelet, generating multiple key matrices using the high-frequency sub-band of the wavelet, and generating a value matrix using Fourier features; guiding the multi-directional texture detail information into the global frequency domain structure through the channel attention mechanism to generate enhanced frequency domain features.

3. The hyperspectral unmixing method based on multi-domain feature sensing fusion according to claim 1, characterized in that, In step (3), the cross-domain dynamic perception module is as follows: the two types of cross-domain features are split into multiple independent subspaces; a perception matrix is ​​constructed based on element-wise multiplication in each subspace, and its context perception capability is enhanced by a linear layer to generate an appropriate context-aware convolution kernel for each type of feature, and neighborhood interaction aggregation is performed at each spatial location to achieve semantic alignment and complementary advantages of cross-domain features; combined with multi-scale convolution detail compensation, local details are enhanced and then fused with the residual of the original feature.

4. The hyperspectral unmixing method based on multi-domain feature sensing fusion according to claim 1, characterized in that, In step (5), the semantic enhancement abundance generation module specifically involves: dividing the semantic encoding into multiple parts according to the number of endmembers, and performing attention semantic enhancement on each part; mapping the enhanced parts to abundance responses through a linear layer, and processing them through convolutional smoothing and a temperature-controlled activation function to obtain an abundance map that satisfies the non-negativity constraint and the sum-to-one constraint.

5. The hyperspectral unmixing method based on multi-domain feature sensing fusion according to claim 1, characterized in that, In step (5), in the Transformer encoder and decoder, the encoding stage only performs query projection on the semantic vector, while the key and value projection is performed on all feature vectors to reduce computational complexity and enhance semantic representation; the decoding stage uses a multi-head attention mechanism to enhance semantic encoding, and then outputs abundance through linear mapping, smoothing and controllable activation.

6. The hyperspectral demixing method based on multi-domain feature sensing fusion according to claim 5, characterized in that, The method employs a self-supervised approach during training, and the loss function includes reconstruction error loss and spectral angular distance loss, which are weighted together to optimize the network parameters.

7. A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method as described in any one of claims 1-6.

8. An electronic device comprising a memory and a processor, the memory storing a computer program that, when executed by the processor, implements the method as described in any one of claims 1-6.

Images