A hyperspectral blind unmixing method

Abstract

The application discloses a hyperspectral blind unmixing method, which comprises the following steps: constructing a hyperspectral blind unmixing model, inputting hyperspectral image data into the hyperspectral blind unmixing model, encoding the spatial feature and the spectral feature of each pixel in the hyperspectral image data through a shallow encoder to generate joint features, and synchronously transmitting the joint features to an end member extraction branch and an abundance estimation branch through a feature sharing mechanism; the end member extraction branch receives hidden end member features extracted from the joint features through a feature extraction module, and then extracts pure end members from the joint features by using a Kolmogorov-Arnold network; the abundance estimation branch extracts an abundance map corresponding to the pure end members from the joint features by using a windowed multi-head self-attention mechanism, and obtains an unmixing result of the mixed spectrum. The method does not need to depend on prior end member information, reduces the dependence of a traditional method on a known end member library, and improves the unmixing efficiency, stability and accuracy of the system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of hyperspectral image processing, in particular to a hyperspectral blind unmixing method, and more particularly to a hyperspectral blind unmixing method based on Kolmogorov-Arnold network driving. BACKGROUND

[0002] Hyperspectral imaging technology has been widely used in remote sensing, agriculture, environmental monitoring and other fields, because it can accurately capture the composition information of different substances in the scene through multiple waveband spectral information. Each pixel point contained in the hyperspectral image is often a mixture of multiple substance spectra, and how to decompose the mixed spectrum into the pure spectrum (endmember) of each substance and its corresponding abundance map becomes an important task of hyperspectral unmixing.

[0003] Existing hyperspectral unmixing methods usually rely on known pure endmember spectral information, or need to construct an endmember library. However, in practical applications, it is extremely challenging to obtain complete and accurate endmember information. Therefore, the traditional blind hyperspectral unmixing method has the following defects: 1) it relies on a complex model structure, has a large amount of calculation, and has low unmixing efficiency; 2) the stability and accuracy of the unmixing result also need to be improved. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a hyperspectral blind unmixing method to solve the technical problems mentioned in the prior art.

[0005] A hyperspectral blind unmixing method, characterized in that the method comprises the following steps:

[0006] Constructing a hyperspectral blind unmixing model, the hyperspectral blind unmixing model having a shallow encoder, an endmember extraction branch and an abundance estimation branch;

[0007] Inputting the hyperspectral image data into the hyperspectral blind unmixing model, encoding the spatial features and spectral features of each pixel in the hyperspectral image data through the shallow encoder to generate joint features, and synchronously transmitting the joint features to the endmember extraction branch and the abundance estimation branch through a feature sharing mechanism;

[0008] The endmember extraction branch receives the joint features output by the shallow encoder, extracts hidden endmember features from the joint features through a feature extraction module, and then extracts pure endmembers from the joint features using a Kolmogorov-Arnold network;

[0009] The abundance estimation branch receives the joint features output by the shallow encoder, and extracts the abundance map corresponding to the pure endmember from the joint features using a windowed multi-head self-attention mechanism to obtain the unmixing result of the mixed spectrum.

[0010] Optionally, the method further comprises linearly combining the extracted pure endmembers and the abundance map to reconstruct a hyperspectral image, and evaluating the accuracy of the unmixing result of the hyperspectral blind unmixing model by comparing the consistency between the hyperspectral image and the hyperspectral image data input to the hyperspectral blind unmixing model.

[0011] If the hyperspectral image is consistent with the hyperspectral image data input to the hyperspectral blind unmixing model, it indicates that the accuracy of the unmixing result of the hyperspectral blind unmixing model is high.

[0012] If the hyperspectral image is inconsistent with the hyperspectral image data input to the hyperspectral blind unmixing model, it indicates that the accuracy of the unmixing result of the hyperspectral blind unmixing model is low.

[0013] Optionally, the method further comprises using a joint loss function to constrain the endmember extraction branch and the abundance estimation branch of the hyperspectral blind unmixing model respectively, so as to iteratively optimize the hyperspectral blind unmixing model during the training process, wherein:

[0014] The joint loss function comprises a spectral angle distance loss function, a mean square error loss function, a total variation loss function and a regularization loss function.

[0015] The spectral angle distance loss and the mean square error loss are used to ensure that the hyperspectral image reconstructed according to the extracted pure endmembers and the abundance map can accurately approximate the hyperspectral image data input to the hyperspectral blind unmixing model.

[0016] The total variation loss and the regularization loss are used to enhance the perception of subtle texture information and structural information in the hyperspectral image data, so as to ensure that the abundance map can capture edge features and local features in the hyperspectral image data.

[0017] Optionally, each pixel of the hyperspectral image data corresponds to a mixed spectrum of different substances, and the mixed spectrum contains spatial features and spectral features.

[0018] The spectral features have a plurality of wavebands, and the wavelength range of each waveband is different.

[0019] Optionally, the convolutional neural network structure of the shallow encoder is 1x1 convolution.

[0020] Optionally, the method for the Kolmogorov-Arnold network to extract pure endmembers from the joint features is:

[0021] The joint features are separated by a nonlinear multivariate function, and a clustering technique is used to cluster the spectral features separated from the joint features to obtain the pure endmembers.

[0022] Optionally, the neural network of the endmember extraction branch has an activation function layer to constrain the pure endmembers extracted by the K-A network, so that the pure endmembers output by the endmember extraction branch are non-negative values.

[0023] Optionally, the abundance estimation branch is a Swin Transformer module.

[0024] Optionally, the output features of the Swin Transformer module are smoothed, and the smoothing is realized by a smoothing layer to constrain the sum of the abundance values of all pixels in the target abundance matrix output by the abundance estimation branch to be 1.

[0025] The smoothing layer includes a 3x3 convolution layer and a softmax function.

[0026] The beneficial effects that can be produced by the present application include:

[0027] The hyperspectral blind unmixing method provided by the present application does not need to rely on prior endmember information, reduces the dependence of traditional methods on known endmember library, and improves the flexibility and applicability of the system. And using a double-branch deep learning architecture, endmember extraction and abundance estimation can be performed in parallel, not only improving the unmixing efficiency, but also ensuring the accuracy and stability of the unmixing results. In addition, the introduction of the feature sharing mechanism enables the spatial features and spectral features extracted by the shallow encoder to be utilized simultaneously, enhancing the information consistency between endmember extraction and abundance estimation, thereby optimizing the unmixing performance. Among them, in terms of abundance estimation, the Swin Transformer module used effectively improves the perception ability of the spatial structure of the hyperspectral image, and through the windowed self-attention mechanism and window offset mechanism, it better captures the long-distance dependency relationship in the image. In order to further optimize the unmixing results, the present application designs a joint loss function, combining spectral angle distance loss, mean square error loss, total variation loss and regularization loss, to ensure the dual advantages of precision and detail preservation of the unmixing results. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 The flowchart of the hyperspectral blind unmixing method of the present application;

[0029] Figure 2 The unmixing result diagram of the Jasper Ridge data set in the present application;

[0030] Figure 3A schematic diagram of the unmixing result of the Urban4 data set in the present application;

[0031] Figure 4 A schematic diagram of the unmixing result of the Urban5 data set in the present application;

[0032] Figure 5 A schematic diagram of the unmixing result of the synthetic data set in the present application. DETAILED DESCRIPTION

[0033] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0034] As Figure 1 shown, the present application provides a hyperspectral blind unmixing method, which comprises the following steps: constructing a hyperspectral blind unmixing model, the hyperspectral blind unmixing model having a shallow encoder, an endmember extraction branch and an abundance estimation branch; inputting hyperspectral image data into the hyperspectral blind unmixing model, encoding the mixed spectrum of each pixel in the hyperspectral image data through the shallow encoder to generate joint features, and synchronously transmitting the joint features to the endmember extraction branch and the abundance estimation branch through a feature sharing mechanism; wherein the joint features comprise spatial features and spectral features, so as to share the encoded spatial features and spectral features to the endmember extraction branch and the abundance estimation branch, which can ensure that the endmember extraction branch and the abundance estimation branch jointly utilize low-level features, guarantee the consistency of information, thereby optimizing the learning effect of the double-branch deep learning architecture and improving the accuracy and stability of the unmixing result of the hyperspectral blind unmixing model; the endmember extraction branch receives the joint features output by the shallow encoder, extracts hidden endmember features from the joint features through a feature extraction module, and then extracts pure endmembers from the hyperspectral image data using a Kolmogorov-Arnold network; it should be noted that the present application replaces the traditional multilayer perceptron with the Kolmogorov-Arnold network, which can directly extract pure endmember spectra from the hyperspectral image data, avoids the dependence on the known endmember library in the traditional method, simplifies the unmixing process, and can adapt to more applications in actual scenarios; the abundance estimation branch receives the joint features output by the shallow encoder, and extracts an abundance map corresponding to the pure endmembers from the joint features using a windowed multi-head self-attention mechanism, to obtain the unmixing result of the mixed spectrum, wherein the abundance map represents the relative proportion of each material in each pixel.

[0035] In the above, hyperspectral image data is collected by spectral imaging equipment. Each pixel of the hyperspectral image data corresponds to a mixed spectrum of different substances. The mixed spectrum includes spatial features and spectral features. The spectral features have several bands, and each band has a different wavelength range.

[0036] In the above, the convolutional neural network structure of the shallow encoder is a 1×1 convolution, which is used to reduce the dimensionality of the input high-dimensional hyperspectral image data while retaining its key spatial and spectral features. Specifically, the joint feature F extracted by the shallow encoder is:

[0037] F = Encoder(Conv(Y));

[0038] Among them, Y represents the input hyperspectral image data, Conv is the 1×1 convolution operation for dimensionality reduction, and Encoder is the encoder.

[0039] Furthermore, the Kolmogorov-Arnold network extracts pure endmembers from the joint features by separating the joint features through a nonlinear multivariable function and clustering the spectral features separated from the joint features using clustering technology to obtain pure endmembers. The nonlinear multivariable function f(x) is:

[0040]

[0041] Among them, φ i and φ i,j are continuous single variable functions, x j Represents the jth component of the input vector, and the i-th continuous single variable function can be based on x j It should be noted that through this nonlinear multivariable function decomposition method, the Kolmogorov-Arnold network can decompose the mixed spectrum into pure end members of different substances and use E hid Extract the separated pure end members;

[0042]

[0043] Among them, E ext represents the pure end member extracted from the Kolmogorov-Arnold network, KAN represents the operation of the Kolmogorov-Arnold network, E i Indicates E hid The i-th end member curve, e mrepresents spectral characteristics, P is the number of endmembers; in the endmember extraction branch, the branch layer number of the Kolmogorov-Arnold network is [N, 2N, P]. In this embodiment, the neural network of the endmember extraction branch has an activation function layer to constrain the pure endmembers extracted by the Kolmogorov-Arnold network, so that the pure endmembers output by the endmember extraction branch are non-negative values.

[0044] Further, the abundance estimation branch is a Swin Transformer module. The Swin Transformer module performs local feature extraction through a windowed multi-head self-attention mechanism (W-MSA) to realize smooth conversion between local windows and global windows, where the multi-head self-attention mechanism (W-MSA) is represented as:

[0045]

[0046] where Q is a query matrix, K is a key matrix, V is a value matrix, B is a bias matrix, d is the dimension of the query matrix or the key matrix, and T is a transpose. The Swin Transformer module (a window movement-based Transformer module) shares information between local and global windows through a self-attention mechanism to capture long-distance dependencies in joint features, thereby generating an accurate abundance map. In order to further improve the calculation efficiency, a window offset mechanism (SW-MAS) is introduced in the Swin Transformer module to realize information interaction across windows by performing window offset of a number of pixels between two consecutive windows in each layer of self-attention calculation, thereby improving the global perception ability of the abundance estimation branch.

[0047] In the above, by utilizing the nonlinear function decomposition capability of the Kolmogorov-Arnold network, the hyperspectral blind unmixing model can capture more complex spectral characteristics in the abundance estimation process, thereby improving the generation accuracy of the abundance map. As shown in Figure 1 The calculation process of the Swin Transformer module for pixel window offset between two consecutive windows is as follows:

[0048]

[0049] where z l is the encoding information of the lth layer, is the attention encoding of the lth layer, LN is layer normalization, KAN is the linear layer of the Kolmogorov-Arnold network, W-MSA is the multi-head self-attention mechanism, and SW-MSA is the multi-head self-attention mechanism with sliding windows.

[0050] In the above, in order to constrain the sum of the abundance map to 1, the present invention smoothes the output features of the Swin Transformer module. The smoothing process is implemented by a smoothing layer to constrain the sum of the abundance values ​​of all pixels in the target abundance matrix output by the abundance estimation branch to 1; the smoothing layer includes a 3×3 convolutional layer and a softmax function.

[0051] Furthermore, in order to evaluate the accuracy of the unmixing results of the hyperspectral blind unmixing model, the extracted pure endmembers and abundance maps are linearly combined to reconstruct the hyperspectral image, and the consistency of the hyperspectral image and the hyperspectral image data input into the hyperspectral blind unmixing model is compared; if the two are consistent, it means that the accuracy of the unmixing results of the hyperspectral blind unmixing model is high; if the two are inconsistent, it means that the accuracy of the unmixing results of the hyperspectral blind unmixing model is low.

[0052] Furthermore, a joint loss function is used to constrain the endmember extraction branch and abundance estimation branch of the hyperspectral blind unmixing model respectively, so as to iteratively optimize the hyperspectral blind unmixing model during the training process. The joint loss function includes the spectral angular distance (SAD) loss function, the mean square error (MSE) loss function, the total variation (TV) loss function and the l 1 / 2 Regularization loss function; specifically, spectral angle distance loss and mean square error loss are used to ensure that the hyperspectral image reconstructed based on the extracted pure endmembers and abundance map can accurately approximate the hyperspectral image data input by the hyperspectral blind unmixing model; total variation loss and regularization loss are used to enhance the perception of subtle texture information and structural information in the hyperspectral image data to ensure that the abundance map can capture the edge features and local features in the hyperspectral image data.

[0053] It should be noted that in the unmixing task based on reconstructed hyperspectral images, the commonly used loss functions are the spectral angular distance (SAD) loss function SADLoss and the mean square error (MSE) loss function MSELoss, which are used together to evaluate the reconstructed hyperspectral image Y re The performance of its combined loss function Loss m The expression is:

[0054] Loss m (Y, Y re )=α·SADLoss(Y,Y re )+β·MSELoss(Y,Y re );

[0055] The above combined loss function Loss mWithout considering the texture information in a complex scene, in order to make the abundance matrix A more sensitive to target edge and structure information, while further reducing noise interference and smoothing the abundance matrix, the total variation (TV) loss is introduced to encourage the optimization of the abundance matrix under the condition of less noise and edge transition smoothing, and the expression of the TV loss function TVLoss is as follows:

[0056]

[0057] Wherein, is the horizontal gradient of A, is the vertical gradient of A.

[0058] In the above, since the degrees of freedom are too high when solving the abundance matrix A and the endmember matrix E in the unmixing task, in order to enhance the robustness and feature extraction ability of the model, the l 1 / 2 regularization loss function is introduced to stabilize the abundance matrix, the l 1 / 2 regularization loss function Loss 1 / 2 The expression of the l

[0059]

[0060] The final joint loss function Loss H U (Y, Y re , A) is:

[0061] Loss HU (Y, Y re , A) = alpha * SADLoss(Y, Y re ) + beta * MSELoss(Y, Y re ) + gamma * TVLoss(A) + lambda * Loss 1 / 2 (A);

[0062] Wherein, alpha = 1, beta = 0.95, gamma = 10 -7 , lambda = 10 _8 .

[0063] It is worth noting that the hyperspectral blind unmixing method provided by the present application does not rely on prior endmember information, reduces the dependence of traditional methods on known endmember library, and improves the flexibility and applicability of the system. And using a double-branch deep learning architecture, endmember extraction and abundance estimation can be performed in parallel, not only improving the unmixing efficiency, but also ensuring the accuracy and stability of the unmixing results. In addition, the introduction of the feature sharing mechanism enables the spatial and spectral features extracted by the shallow encoder to be utilized simultaneously, enhancing the information consistency between endmember extraction and abundance estimation, thereby optimizing the unmixing performance. Among them, in terms of abundance estimation, the Swin Transformer module used effectively improves the perception of the spatial structure of the hyperspectral image, and through the windowed self-attention mechanism and window offset mechanism, it better captures the long-distance dependence relationship in the image. In order to further optimize the unmixing results, the present application designs a joint loss function, combining the spectral angle distance (SAD) loss function, mean square error (MSE) loss function, total variation (TV) loss function and l 1 / 2 regularization loss function, ensuring the dual advantages of precision and detail preservation of the unmixing results.

[0064] Based on the above embodiments, in order to verify the unmixing results of the hyperspectral blind unmixing model, the present application carries out a plurality of experiments to evaluate the effectiveness and accuracy of the method proposed by the present application in hyperspectral unmixing. Specifically, the experimental method includes the following steps:

[0065] 1) Select the data set: In this embodiment, a plurality of public hyperspectral data sets are selected, including the Jasper Ridge data set, the Urban data set and the synthetic data set, which are widely used for algorithm verification in the field of hyperspectral unmixing.

[0066] In the above, the Jasper Ridge data set is collected in central California, USA, and uses the AVIRIS sensor for data acquisition. The acquisition scene includes four materials: water, soil, trees and roads. The data cube is composed of 100x100 pixels, and the original data contains 224 wavebands. The spectral range is 0.38-2.5. After removing the influence of water vapor and atmospheric interference, the data retains 198 wavebands.

[0067] In the above, the Urban dataset was collected in the urban area of Copperas Cove, Texas, USA, using a HYDICE HSI sensor. The EMs (endmembers) in the dataset were divided into 4 classes and 5 classes. The 4 classes (Urban4) included asphalt, grass, trees, and rooftops; and the 5 classes (Urban5) added soil to the 4 classes (Urban4). The datacube contained 307 x 307 pixels. After removing dense water vapor and atmospheric effects, the dataset retained 162 bands, with a spectral range of 400-2500 and a spectral resolution of 10.

[0068] In the above, the synthetic dataset is a synthetic dataset that contains various mineral electromagnetic curves from the U.S. Geological Survey spectral library. The dataset can generate datasets of different sizes using its built-in multilinear mixing model and has various mixtures of pure EMs. In this method, the present application selects 10 mineral spectra, creates a dataset with a spatial dimension of 128 x 128 pixels and 216 bands for experiments.

[0069] 2) Set evaluation index: In quantitative analysis, the present application uses spectral angle distance (SAD) and root mean square error (RMSE) to evaluate the extracted endmember matrix and abundance matrix. Among them, the spectral angle distance (SAD) is used to evaluate the estimated endmember matrix, which measures the similarity between the estimated endmember of each material and the pure endmember. The expression of SAD is:

[0070]

[0071] wherein, represents the i-th estimated endmember, E i represents the i-th pure endmember, P is the number of endmembers in the hyperspectral image. The root mean square error (RMSE) is used to evaluate the abundance matrix, and the expression of RMSE is:

[0072]

[0073] wherein, represents the i-th estimated abundance matrix, A i represents the i-th abundance map.

[0074] It should be noted that the ideal values of SAD and RMSE are both 0.

[0075] 3) Selection of experimental methods: This embodiment selects 8 methods for comparison, namely fully constrained least squares linear unmixing algorithm (FCLSU), robust non-negative matrix factorization (rNMF), extended linear mixture model (ELMM), linear mixture model (PLMM), three-dimensional convolutional neural network (3DCNN), window sliding based Transformer (swt), spatial-spectral attention bilateral hyperspectral unmixing network (SSABN) and abundance-guided hyperspectral unmixing network (A2SAN). Among them, FCLSU, rNMF, ELMM, PLMM belong to traditional machine learning methods, while 3D-CNN, SwinT, SSABN, A2SAN belong to deep learning methods. In order to ensure a fair comparison, the present invention adopts the VCA method to obtain a priori end members, thereby initializing the method that requires a priori end members. In addition, the parameters of the hyperspectral blind unmixing model are all set to default values.

[0076] 4) Experimental results and analysis: Figure 2 Figure 2 shows the performance of the experimental methods described above on the Jasper Ridge dataset. For abundance estimation, rNMF and ELMM incorrectly classified "road" areas as "water," while FCLSU, ELMM, and PLMM also failed to accurately estimate "road" areas in their abundance estimates. In contrast, deep learning methods outperformed traditional machine learning methods overall. However, 3D-CNN significantly overestimated the abundance of "road" compared to the ground truth, SSABN had significant deficiencies in estimating "soil" area, and Swin Transformer and A2SAN also faced difficulties distinguishing "road" from "soil" in their abundance estimates, resulting in some "road" areas being incorrectly included in the "soil" estimates. Compared with traditional methods, the proposed method demonstrated higher accuracy in abundance estimation. In the comparison of endmember extraction, the extraction results for "road" showed the greatest discrepancy, indicating that most methods performed poorly in extracting "road," a situation also reflected in abundance estimation. Furthermore, some experimental methods produced non-positive endmember outputs. For example, the PLMM produced negative values ​​in the EM curve for "trees," while the rNMF produced zero values ​​in the EM curve for "soil." Overall, the method proposed in this paper achieved more accurate results in both abundance estimation and endmember extraction. Table 1 lists the performance metrics of all experimental methods on the Jasper Ridge dataset.

[0077] Table 1:

[0078]

[0079] From Table 1, it can be seen that the SAD values of the traditional machine learning methods are relatively high. This is because most ML methods need to perform an endmember extraction algorithm first, and then further refine the endmember and abundance estimation according to the extraction results. In contrast, the data-driven deep learning method performs better in the measurement results. Compared with the suboptimal method A2SAN, the method proposed in the present application significantly reduces the SAD and RMSE, further proving its effectiveness.

[0080] As Figure 3 shown, the performance of the above experimental method on the Urban4 dataset is shown. Obviously, compared with the JasperRidge dataset, the abundance estimation in GT is more complex; FCLSU, rNMF, ELMM and PLMM not only have low accuracy in abundance estimation, but also have various misestimation problems, such as PLMM mistakenly including the area that should be attributed to "asphalt" into the estimation of "grass"; rNMF mistakenly includes "asphalt" in the estimation of "roof"; and FCLSU has significant overlap between the abundance estimation areas of "trees" and "grass", while underestimating the abundance of "asphalt". Compared with the four traditional machine learning methods, 3D-CNN has certain improvement in abundance estimation, but still faces the problem of misestimation; the Swin Transformer module lacks accuracy in area estimation, A2SAN underestimates the abundance of "roof", and SSABN overestimates the abundance of "grass". Referring to Figure 5 the electromagnetic curves shown in the figure, although all experimental methods perform relatively well in the extraction of "trees", there are significant deviations in the extraction of "asphalt", "grass" and "roof", and even some experimental methods produce negative values. As shown in Table 2, the quantitative results of all experimental methods on the Urban4 dataset are shown, and obviously the method proposed in the present application has achieved satisfactory results in terms of SAD and RMSE.

[0081] Table 2:

[0082]

[0083] As Figure 4As shown in FIG. 6, the abundance estimation and endmember extraction of all experimental methods on the Urban5 dataset are shown. Among them, with the increase of the number of endmembers, the results of the first four experimental methods cannot achieve ideal abundance estimation; FCLSU cannot produce the abundance estimation of "trees" and "roof", and rNMF incorrectly estimates the abundance of "roof". 3D-CNN successfully estimates the abundance of five endmembers, but still makes mistakes in the estimation of "grass" and "trees". SwinT and A2SAN are insufficient in estimating the abundance of "asphalt" and "soil". Although SSABN and the method proposed in the present application perform better in abundance estimation, both of them underestimate the newly added endmember "soil", compared with the deep learning method which shows significant improvement. In the endmember extraction curve, the pure endmembers extracted by most experimental methods have different degrees of deviation, among which the difference of "soil" is the most significant, but the method proposed in the present application can accurately extract the curve of each endmember. As shown in Table 3, the quantitative results of all experimental methods under different noise intensity levels of the Urban5 dataset are shown. Compared with the results of the Urban4 dataset, the accuracy of all experimental methods has decreased, which indicates that the more detailed the distinction between endmembers in the same scene is, the more difficult it is to extract. Nevertheless, among all experimental methods, the method proposed in the present application still maintains a high accuracy rate, and its results are largely unaffected by the increase of noise level, proving its robustness; the signal-to-noise ratio (SNR) used in the experimental process is 20 dB, 40 dB and 60 dB, respectively.

[0084] Table 3:

[0085]

[0086] As Figure 5 shown in FIG. 6, the abundance estimation and endmember extraction of all experimental methods on the Urban5 dataset are shown. Among them, with the increase of the number of endmembers, the results of the first four experimental methods cannot achieve ideal abundance estimation; FCLSU cannot produce the abundance estimation of "trees" and "roof", and rNMF incorrectly estimates the abundance of "roof". 3D-CNN successfully estimates the abundance of five endmembers, but still makes mistakes in the estimation of "grass" and "trees". SwinT and A2SAN are insufficient in estimating the abundance of "asphalt" and "soil". Although SSABN and the method proposed in the present application perform better in abundance estimation, both of them underestimate the newly added endmember "soil", compared with the deep learning method which shows significant improvement. In the endmember extraction curve, the pure endmembers extracted by most experimental methods have different degrees of deviation, among which the difference of "soil" is the most significant, but the method proposed in the present application can accurately extract the curve of each endmember. As shown in Table 3, the quantitative results of all experimental methods under different noise intensity levels of the Urban5 dataset are shown. Compared with the results of the Urban4 dataset, the accuracy of all experimental methods has decreased, which indicates that the more detailed the distinction between endmembers in the same scene is, the more difficult it is to extract. Nevertheless, among all experimental methods, the method proposed in the present application still maintains a high accuracy rate, and its results are largely unaffected by the increase of noise level, proving its robustness; the signal-to-noise ratio (SNR) used in the experimental process is 20 dB, 40 dB and 60 dB, respectively.

[0087] In the above, although the method proposed by the present application has some deviations in the abundance estimation of "cedar" and "paperboard", compared with other experimental methods, it can more accurately estimate the abundance estimation of most substances, especially in the electromagnetic curve of end member extraction, the method proposed by the present application shows higher accuracy, while the spectral curve of other experimental methods is mostly distorted. As shown in Table 4, the quantitative results of all experimental methods on the Synthesis data set are shown. Under low noise intensity, the SAD value of most machine learning methods is low, and the performance of end member extraction is better. However, as the noise intensity increases, the SAD value of the machine learning method deteriorates more significantly. In contrast, the performance of the deep learning method remains relatively stable, reflecting its strong robustness. In particular, the method proposed by the present application maintains a high quantitative result under all noise intensities, further proving its ability to perform high-precision unmixing in complex scenes with multiple end members.

[0088] Table 4:

[0089]

[0090] In summary, the experimental results of the method proposed by the present application on a variety of hyperspectral image data sets show that the method has significant advantages in effectiveness and accuracy in hyperspectral unmixing, and at the same time demonstrates its significant advantages in unmixing precision and robustness, especially suitable for multiple practical application scenarios such as remote sensing, environmental monitoring and agricultural resource assessment, and shows a wide application prospect and good market potential.

Claims

1. A hyperspectral blind unmixing method, characterized in that, The method comprises the following steps: constructing a hyperspectral blind unmixing model having a shallow encoder, an endmember extraction branch and an abundance estimation branch; inputting the hyperspectral image data into the hyperspectral blind unmixing model, encoding the mixed spectrum of each pixel in the hyperspectral image data through the shallow encoder to generate joint features, and synchronously transmitting the joint features to the endmember extraction branch and the abundance estimation branch through a feature sharing mechanism; the endmember extraction branch receives the joint features output by the shallow encoder, extracts hidden endmember features from the joint features through a feature extraction module, and then extracts pure endmembers from the joint features using a Kolmogorov-Arnold network; wherein the joint features are separated through a nonlinear multivariate function, and the separated spectral features in the joint features are clustered using a clustering technique to obtain the pure endmembers; the neural network of the endmember extraction branch has an activation function layer to constrain the pure endmembers extracted by the Kolmogorov-Arnold network, so that the pure endmembers output by the endmember extraction branch are non-negative; the abundance estimation branch receives the joint features output by the shallow encoder, and extracts an abundance map corresponding to the pure endmembers from the joint features using a windowed multi-head self-attention mechanism to obtain an unmixing result of the mixed spectrum. 2.The hyperspectral blind unmixing method according to claim 1, characterized in that, The method further comprises linearly combining the extracted pure endmembers and the abundance map to reconstruct a hyperspectral image, and evaluating the accuracy of the unmixing result of the hyperspectral blind unmixing model by comparing the consistency of the hyperspectral image with the hyperspectral image data input into the hyperspectral blind unmixing model; if the hyperspectral image and the hyperspectral image data input into the hyperspectral blind unmixing model remain consistent, it indicates that the accuracy of the unmixing result of the hyperspectral blind unmixing model is high; if the hyperspectral image and the hyperspectral image data input into the hyperspectral blind unmixing model are inconsistent, it indicates that the accuracy of the unmixing result of the hyperspectral blind unmixing model is low. 3.The hyperspectral blind unmixing method of claim 1, wherein, The method further comprises using a joint loss function to constrain the endmember extraction branch and the abundance estimation branch of the hyperspectral blind unmixing model respectively to iteratively optimize the hyperspectral blind unmixing model during the training process, wherein the joint loss function comprises a spectral angle distance loss function, a mean square error loss function, a total variation loss function and a regularization loss function; the spectral angle distance loss and the mean square error loss are used to ensure that the hyperspectral image reconstructed from the extracted pure endmembers and the abundance map can accurately approximate the hyperspectral image data input into the hyperspectral blind unmixing model; the total variation loss and the regularization loss are used to enhance the perception of subtle texture information and structural information in the hyperspectral image data to ensure that the abundance map can capture edge features and local features in the hyperspectral image data.

4. The hyperspectral blind unmixing method of claim 1, wherein, Each pixel of the hyperspectral image data corresponds to a mixed spectrum of different substances, and the mixed spectrum contains spatial features and spectral features. The spectral features have a plurality of wave bands, and the wavelength range of each wave band is different.

5. The hyperspectral blind unmixing method of claim 1, wherein, The convolutional neural network structure of the shallow encoder is 1x1 convolution.

6. The hyperspectral blind unmixing method of claim 1, wherein, The abundance estimation branch is a Swin Transformer module.

7. The hyperspectral blind unmixing method of claim 6, wherein, The output features of the Swin Transformer module are smoothed, and the smoothing is realized by a smoothing layer to constrain the sum of the abundance values of all pixels in the target abundance matrix output by the abundance estimation branch to be 1. The smoothing layer includes a 3x3 convolution layer and a softmax function.

Images