Mixed pixel decomposition method and system for hyperspectral image

Abstract

The invention discloses a mixed pixel decomposition method and system for a hyperspectral image, and relates to the technical field of remote sensing image processing, and the mixed pixel decomposition method comprises the steps: obtaining a hyperspectral image and a ground feature end member spectrum corresponding to the hyperspectral image, and generating a training data set based on the hyperspectral image; the training data set and the ground feature end member spectrum are input into a pixel decomposition model to optimize model parameters of the pixel decomposition model, and the pixel decomposition model comprises a data loader, a three-dimensional convolution encoder, a multi-head cross attention module, a gating guide module and a linear decoder. And reasoning and splicing the training data set by using the optimized pixel decomposition model to obtain an end member abundance image after pixel decomposition. According to the invention, on the basis of the convolution auto-encoder, the multi-head cross attention module and the gating unit are introduced, so that adaptive sparse correction based on scene characteristics is realized, and then the precision of hyperspectral image de-mixing is improved.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, specifically to a method and system for hybrid pixel decomposition of hyperspectral images. Background Technology

[0002] Hyperspectral remote sensing technology is a remote sensing method that uses continuous spectral data with nanometer-level spectral resolution to image ground objects. It can break through the spatial-spectral dimension limitations of traditional multispectral remote sensing and achieve diagnosis of the spatial texture and fine spectrum of ground objects.

[0003] However, due to spatial resolution limitations, many pixels correspond to surface areas containing multiple land features. The spectral reflectance of these features mixes, forming mixed pixels, making it impossible to definitively classify the spectral characteristics of these pixels into a single category. The existence of mixed pixels significantly limits the classification accuracy and land feature identification capabilities of hyperspectral data.

[0004] Therefore, there is an urgent need for a ground feature identification method that can effectively decompose mixed pixels, thereby improving the application capabilities of hyperspectral remote sensing and providing more refined ground feature information support for scientific research. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for decomposing mixed pixels in hyperspectral images. By introducing an attention gating mechanism, it effectively identifies and separates different land cover components in mixed pixels.

[0006] This invention discloses a hybrid pixel decomposition method for hyperspectral images, comprising: acquiring hyperspectral images and land cover endmember spectra corresponding to the hyperspectral images, wherein the hyperspectral images contain at least two land cover types, and the type and number of land cover endmember spectra are matched with the land cover types;

[0007] Training dataset generated from hyperspectral images;

[0008] The training dataset and the spectra of ground features endmembers are input into the pixel decomposition model to optimize the model parameters; the pixel decomposition model includes a data loader, a 3D convolutional encoder, a multi-head cross-attention module, a gating guidance module, and a linear decoder;

[0009] The optimized pixel decomposition model is used to infer and stitch together the training dataset to obtain the endmember abundance image after pixel decomposition.

[0010] Preferably, the spectral resolution of the hyperspectral image is greater than or equal to 10 nanometers, and the spectral resolution of the ground object end-member spectrum is greater than or equal to the spectral resolution of the hyperspectral image.

[0011] The wavelength range of the end-member spectrum of ground features is greater than or equal to the wavelength range of the hyperspectral image.

[0012] Preferably, after acquiring the hyperspectral image and the corresponding endmember spectra of the ground features, preprocessing is performed on the hyperspectral image and the endmember spectra of the ground features, including:

[0013] Obtain the image description file of the hyperspectral image, and perform radiometric calibration on the hyperspectral image according to the image description file to obtain the radiometrically calibrated surface reflectance;

[0014] Atmospheric correction is applied to the surface reflectance to obtain the atmospherically corrected surface reflectance;

[0015] The end-member spectra of ground features are resampled based on the bands of the hyperspectral image to align the end-member spectra with the hyperspectral image in the spectral dimension.

[0016] Preferably, a training dataset is generated based on hyperspectral imagery, including:

[0017] Perform specular reflection filling on the hyperspectral image to obtain the filled image;

[0018] The filled image is cut into a preset number of image blocks, and each image block is a square of the same size.

[0019] A training dataset is generated based on image patches.

[0020] Preferably, the training dataset and the spectra of ground cover endmembers are input into the pixel decomposition model to optimize the model parameters of the pixel decomposition model, including:

[0021] Load the training dataset and ground feature endmember spectra using a data loader;

[0022] A three-dimensional convolutional encoder is used to calculate the abundance prior estimate of image patches;

[0023] The dynamic relationship between image patches and ground object endmember spectra is calculated using a multi-head cross-attention module;

[0024] The abundance prior estimate is optimized using an attention score through a gating guidance module to obtain the abundance vector;

[0025] A linear decoder is used to generate reconstructed images, and the similarity between the reconstructed images and the endmember spectra of ground features is calculated based on a loss function. The model parameters of the pixel decomposition model are then updated based on the similarity.

[0026] Preferably, the 3D convolutional encoder includes multiple 3D convolutional modules, which acquire the spatial and spectral features of the training dataset and establish the correlation between the spatial and spectral features.

[0027] Preferably, the multi-head cross-attention module includes multiple attention modules, which are used to calculate cross-attention.

[0028] Preferably, the loss function includes spectral reconstruction loss and a sparsity regularization term;

[0029] Spectral reconstruction loss is used to characterize the difference in the angle between the spectral vectors of the reconstructed image and the endmember spectra of ground features;

[0030] The sparse regularization term is used to regularize the sparsity of the abundance vector.

[0031] Preferably, after obtaining the endmember abundance image after pixel decomposition, the method further includes:

[0032] Accuracy evaluation of endmember abundance images is performed based on the overall root mean square error of abundance and the abundance signal reconstruction error.

[0033] This invention discloses a hybrid pixel decomposition system for hyperspectral images, used to perform a hybrid pixel decomposition method for hyperspectral images. The hybrid pixel decomposition system includes:

[0034] The data acquisition module is configured to acquire hyperspectral images and the corresponding land cover end-member spectra. The hyperspectral images contain at least two land cover types, and the type and number of land cover end-member spectra match the land cover types.

[0035] The training set generation module is configured to generate a training dataset based on hyperspectral images;

[0036] The model training module is configured to input the training dataset and the spectra of ground features endmembers into the pixel decomposition model to optimize the model parameters of the pixel decomposition model; the pixel decomposition model includes a data loader, a 3D convolutional encoder, a multi-head cross-attention module, a gating guidance module, and a linear decoder;

[0037] The model prediction module is configured to use the optimized pixel decomposition model to infer and stitch together the training dataset to obtain the endmember abundance image after pixel decomposition.

[0038] Compared with existing technologies, the beneficial effects of this invention are as follows: By introducing a three-dimensional convolutional encoder to extract spatial and spectral features of images, and combining a multi-head cross-attention module and a gating guidance module, accurate decomposition of mixed pixels is achieved. By inputting the training dataset and the endmember spectra of ground features into the pixel decomposition model, the model's ability to capture complex surface textures and fine spectral information is enhanced. By optimizing the sparsity of the abundance vector through a sparsity regularization term, the accuracy and robustness of the decomposition results are further improved, thus providing a technical foundation for intelligent ground feature classification and recognition of hyperspectral images. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating the hybrid pixel decomposition method for hyperspectral images provided by the present invention.

[0040] Figure 2 This is a schematic diagram of the structure of the hybrid pixel decomposition system for hyperspectral images provided by the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] The present invention will now be described in further detail with reference to the accompanying drawings.

[0043] like Figure 1 As shown, this embodiment of the invention provides a method for hybrid pixel decomposition of hyperspectral images, including the following steps.

[0044] S1. Obtain hyperspectral images and the corresponding end-member spectra of ground features.

[0045] In this embodiment of the invention, hyperspectral imagery is remote sensing data that acquires the reflectance or radiation characteristics of ground features through continuous, narrow-band imaging. The hyperspectral imagery contains at least two ground feature types, and the type and number of end-member spectra of the ground features match the ground feature types. The end-member spectra of the ground features are obtained from a spectral library.

[0046] Specifically, the spectral resolution of the hyperspectral image is greater than or equal to 10 nanometers, the spectral resolution of the ground object endmember spectrum is greater than or equal to that of the hyperspectral image, and the wavelength range of the ground object endmember spectrum is greater than or equal to that of the hyperspectral image. This improves the data matching between the ground object endmember spectrum and the hyperspectral image, and enhances the accuracy of mixed pixel decomposition.

[0047] In practical applications, hyperspectral imagery often has high spectral resolution, and the spectral resolution of the end-member spectra of ground features needs to be greater than or equal to that of the hyperspectral imagery to meet the requirements of hybrid pixel decomposition. Simultaneously, the wavelength range of the end-member spectra of ground features also needs to cover the wavelength range of the hyperspectral imagery to ensure accurate reflection of ground feature information in the hyperspectral imagery and to provide reliable basic data for subsequent hybrid pixel decomposition.

[0048] In this embodiment of the invention, after acquiring the hyperspectral image and the corresponding ground object end-member spectrum, preprocessing is performed on the hyperspectral image and the ground object end-member spectrum to improve the quality and usability of the data.

[0049] The preprocessing steps include performing radiometric calibration on the hyperspectral image to obtain the radiometrically calibrated surface reflectance; performing atmospheric correction on the surface reflectance to obtain the atmospherically corrected surface reflectance; and resampling the end-member spectrum of the ground object according to the band correspondence between the end-member spectrum of the ground object and the hyperspectral image to align the end-member spectrum of the ground object with the hyperspectral image in the spectral dimension.

[0050] These preprocessing steps can reduce the impact of atmospheric and illumination factors on hyperspectral imagery and improve the matching degree between the end-member spectra of ground features and hyperspectral imagery. Specifically, radiometric calibration converts the digital values ​​(DN) recorded by the hyperspectral multiband sensor into absolute radiance (i.e., surface reflectance). Atmospheric correction is then used to transfer the surface reflectance through the atmosphere to the ground, eliminating errors caused by atmospheric scattering, atmospheric reflection, and absorption, thereby obtaining the ground reflectance.

[0051] If the number of bands and the center wavelength of each band in the endmember spectrum of a ground feature do not match those in the hyperspectral image, the band settings of the hyperspectral image need to be used as the target for resampling the endmember spectrum to align it with the hyperspectral image in the spectral dimension. This improves the accuracy and reliability of mixed pixel decomposition. Let the center wavelength of the endmember spectrum be x. i The reflectance is y i For the target wavelength x'∈[x i ,x i+1 The differential reflectance is expressed using formula (1):

[0052] .

[0053] S2. Generate a training dataset based on hyperspectral images.

[0054] To train the pixel decomposition model, a training dataset needs to be generated from the hyperspectral image. In this embodiment of the invention, specular reflection filling is performed on the hyperspectral image to obtain a filled image. The filled image is then cut into a predetermined number of image blocks, each image block being a square of uniform size, and a training dataset is generated based on these image blocks.

[0055] Because hyperspectral imagery can contain various complex surface features and lighting conditions, specular fill can effectively reduce image distortion caused by specular reflection, making the training dataset closer to reality. The image is divided into uniformly sized square image patches. The side length of each patch depends on the image's dimensions and the GPU configuration used for training, and can be adjusted as needed. By applying specular fill to the hyperspectral image, both its length and width are divisible by the patch's side length, facilitating subsequent model training and classification.

[0056] S3. Input the training dataset and the endmember spectra of ground features into the pixel decomposition model to optimize the model parameters.

[0057] In this embodiment of the invention, the pixel decomposition model is a deep learning network comprising multiple modules, including a data loader, a 3D convolutional encoder, a multi-head cross-attention module, a gating guidance module, and a linear decoder. By inputting the training dataset and the spectra of ground feature endmembers into the model and optimizing the parameters using a loss function, the weights and biases within the model can be gradually adjusted, enabling it to decompose mixed pixels more accurately.

[0058] Specifically, a data loader is used to load the training dataset and ground feature endmember spectra. A 3D convolutional encoder is used to calculate the abundance prior estimates of image patches. A multi-head cross-attention module is used to calculate the dynamic relationship between image patches and ground feature endmember spectra. A gating guidance module is used to optimize the abundance prior estimates using attention scores to obtain the abundance vector. A linear decoder is used to generate reconstructed images, and the similarity between the reconstructed images and ground feature endmember spectra is calculated based on a loss function. The model parameters of the pixel decomposition model are then updated based on the similarity.

[0059] The 3D convolutional encoder includes multiple 3D convolutional modules. These modules acquire the spatial and spectral features of the training dataset and establish the correlation between the spatial and spectral features.

[0060] For example, given an input tensor Each 3D convolution module sequentially processes the input tensor using formulas (2) and (3):

[0061]

[0062] The Conv3D() function represents the use of a 3×3×3 convolution kernel, and uses mirror reflection to fill in the gaps to ensure that the input and output maintain the same shape. The IN() function represents instance normalization, forcibly pulling the input distribution back to the range of a relatively standard normal distribution with a mean of 0 and a variance of 1. This ensures that the input value of the nonlinear transformation function falls into the region that the activation function is sensitive to, thereby improving the gradient and avoiding the gradient vanishing problem. The ReLU() function represents the activation function, used to introduce nonlinear relationships.

[0063] Furthermore, to prevent network degradation, this invention also incorporates a residual structure, adding the original input to IN(Conv3D(F1)). Then, activation is performed using an activation function. Stacked convolutional modules capture hierarchical features, where each layer aggregates information from the local spectral neighborhood (in the depth direction) and spatial context (in the height and width directions). Finally, the feature map is flattened and reshaped by projecting it onto m channels through fully connected layers, and softmax normalization is performed along the channel dimensions to form an abundance prior estimate A0∈R. p×m .

[0064] In mixed spectra, when a certain component reaches a certain abundance, it makes a significant contribution to the overall spectral morphology, causing the spectrum to exhibit similarity to that endmember spectrum. Based on this, embodiments of the present invention utilize a multi-head cross-attention mechanism to capture the dynamic dependencies between mixed pixels and endmembers. The multi-head cross-attention module includes multiple attention modules, which are used to calculate cross-attention.

[0065] For example, let E∈R m×b Given an endmember spectrum, the pixel spectrum X is obtained by flattening the original image. flat ∈R p×b X flat E is linearly projected onto the embedding dimension Q∈R p×b , K∈R m×b Q and K are calculated using formulas (4) and (5) respectively.

[0066] Q = X flat W Q (4);

[0067] K = EW K (5);

[0068] Among them, W Q and W K It is a learnable weight matrix, W K ∈R b×d .

[0069] Furthermore, the attention score for the scaled dot product is defined using formula (6):

[0070]

[0071] Where T represents matrix transpose and d represents vector dimension. The attention score represents the similarity between each pixel and the endmember spectrum. Since the attention module is multi-headed, there are multiple independent subspaces for calculating cross-attention. The multiple attention scores are used as confidence weights for gating. The final attention score used for gating adjustment is the average of the results from multiple attention modules. When the attention module has n heads, the attention score used for gating adjustment is expressed by formula (7):

[0072] .

[0073] In this embodiment, the gating guidance module uses formula (8) to optimize the prior abundance by multiplying and normalizing element by element and using attention scores.

[0074]

[0075] Here, ⊙ represents the Hadamard product. This application utilizes a gating guidance module to enhance the abundance of endmembers similar to hyperspectral images while suppressing the abundance of irrelevant endmembers, thus introducing sparsity guidance.

[0076] In this embodiment of the invention, the linear decoder reconstructs the image through matrix multiplication and calculates the similarity between the reconstructed image and the endmember spectra of ground features based on a loss function. The loss function includes two loss terms: a spectral reconstruction loss and a sparsity regularization term.

[0077] The spectral reconstruction loss uses the spectral angular distance (SAD) between the reconstructed pixel and the input pixel to measure the reconstruction quality. SAD measures the angle between the spectral vectors, focusing more on describing the overall shape of the spectral curve rather than absolute numerical differences, thus avoiding the influence of amplitude variations. The spectral reconstruction loss is expressed by formula (9):

[0078]

[0079] The sparsity regularization term is expressed using formula (10):

[0080]

[0081] The loss function is expressed using formula (11):

[0082]

[0083] in, Indicates spectral reconstruction loss, Let represent the sparsity regularization term, and λ be used to balance the weights of the two terms. Here, represents the loss function, p represents the number of samples, and arccos() represents the inverse cosine function.

[0084] Based on the above process, the pixel decomposition model forms a self-supervised training based on the hyperspectral image dataset to be unmixed and the corresponding endmember spectra, and promotes the sparsity of the abundance vector, thereby updating the model parameters to output a reconstructed image with high similarity to the hyperspectral image.

[0085] S4. Use the optimized pixel decomposition model to infer and stitch together the training dataset to obtain the endmember abundance image after pixel decomposition.

[0086] After the model training is completed, the optimized pixel decomposition model can be used to infer the image patches in the training dataset and stitch the patches together to form the original image sheet, thereby generating the abundance images of each endmember. After obtaining the endmember abundance images after pixel decomposition, the accuracy of the endmember abundance images is evaluated based on the overall root mean square error of abundance (aRMSE) and the abundance signal reconstruction error (aSRE).

[0087] For example, the Samson dataset was processed using the hybrid pixel decomposition method (AGC-Net) for hyperspectral images provided by this invention, along with other deep learning methods, to obtain endmember abundance maps. The accuracy of the endmember abundance images was evaluated based on the overall root mean square error of abundance and the abundance signal reconstruction error, yielding the evaluation results shown in Table 1. The other deep learning methods include: Non-negative Matrix Factorization (NMF), Fully Constrained Least Squares (FCLS), Variable Splitting and Augmented Lagrange Operator Sparse Unmixing (SUnSAL), Deep Convolutional Autoencoder (DCAE), Endmember-Guided Unmixing Network (EGU-Net), and Abundance-Guided Attention Network (A2SAN).

[0088] Table 1. Evaluation of Samson dataset unmixing accuracy

[0089] NMF FCLS SUnSAL DCAE EGU-Net A2SAN AGC-Net aRMSE 0.3102 0.3698 0.2698 0.1025 0.0497 0.1001 0.0431 aSRE 3.4144 2.4775 4.9051 13.4057 19.7431 13.5526 20.9813

[0090] The lower the value of the total root mean square error of abundance (aRMSE) and the higher the value of the abundance signal reconstruction error (aSRE), the better the unmixing performance is. According to Table 1, it can be concluded that the unmixing performance of the hyperspectral image hybrid pixel decomposition method provided in this application is better than other deep learning methods.

[0091] For example, the JasperRidge dataset is processed using the hybrid pixel decomposition method (AGC-Net) for hyperspectral images provided by this invention and the aforementioned deep learning method, respectively, to obtain endmember abundance maps. The accuracy of the endmember abundance images is evaluated based on the overall root mean square error of abundance and the abundance signal reconstruction error, and the evaluation results shown in Table 2 are obtained.

[0092] Table 2. JasperRidge Dataset Dehazing Accuracy Evaluation

[0093] NMF FCLS SUnSAL DCAE EGU-Net A2SAN AGC-Net aRMSE 0.3903 0.2474 0.1887 0.0954 0.0810 0.0769 0.0733 aSRE 0.4927 4.3059 6.8711 12.9690 13.8781 14.7746 15.0054

[0094] As can be seen, the hybrid pixel decomposition method for hyperspectral images provided in this application outperforms other deep learning methods in unmixing the JasperRidge dataset.

[0095] like Figure 2As shown, this embodiment of the invention also provides a hybrid pixel decomposition system for hyperspectral images, used to perform a hybrid pixel decomposition method for hyperspectral images. The hybrid pixel decomposition system for hyperspectral images includes: a data acquisition module 201, a training set generation module 202, a model training module 203, and a model prediction module 204.

[0096] The data analysis module 201 is configured to acquire hyperspectral images and corresponding land cover endmember spectra. The hyperspectral images contain at least two land cover types, and the type and number of land cover endmember spectra match the land cover types. The training set generation module 202 is configured to generate a training dataset based on the hyperspectral images. The model training module 203 is configured to input the training dataset and land cover endmember spectra into the pixel decomposition model to optimize the model parameters. The pixel decomposition model includes a data loader, a 3D convolutional encoder, a multi-head cross-attention module, a gating guidance module, and a linear decoder. The model prediction module 204 is configured to use the optimized pixel decomposition model to infer and stitch the training dataset to obtain the endmember abundance image after pixel decomposition.

[0097] As described above, this invention discloses a method and system for hybrid pixel decomposition of hyperspectral images. By introducing a three-dimensional convolutional encoder to extract the spatial and spectral features of the image, and combining a multi-head cross-attention module and a gating guidance module, accurate decomposition of hybrid pixels is achieved. By inputting the training dataset and the endmember spectra of ground features into the pixel decomposition model, the model's ability to capture complex surface textures and fine spectral information is enhanced. Furthermore, by optimizing the sparsity of the abundance vector through a sparsity regularization term, the accuracy and robustness of the decomposition results are improved, thus providing a technical foundation for intelligent ground feature classification and recognition of hyperspectral images.

[0098] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for hybrid pixel decomposition of hyperspectral images, characterized in that, include: Acquire hyperspectral images and the corresponding land cover endmember spectra, wherein the hyperspectral images contain at least two land cover types, and the type and number of the land cover endmember spectra match the land cover types. A training dataset was generated based on the hyperspectral images; The training dataset and the endmember spectra of the ground features are input into the pixel decomposition model to optimize the model parameters of the pixel decomposition model; The pixel decomposition model includes a data loader, a 3D convolutional encoder, a multi-head cross-attention module, a gating guidance module, and a linear decoder; The optimized pixel decomposition model is used to infer and stitch together the training dataset to obtain the endmember abundance image after pixel decomposition.

2. The hybrid pixel decomposition method according to claim 1, characterized in that, The spectral resolution of the hyperspectral image is greater than or equal to 10 nanometers, and the spectral resolution of the end-member spectrum of the ground features is greater than or equal to the spectral resolution of the hyperspectral image. The wavelength range of the end-member spectrum of the ground feature is greater than or equal to the wavelength range of the hyperspectral image.

3. The hybrid pixel decomposition method according to claim 1, characterized in that, After acquiring the hyperspectral image and the corresponding land cover endmember spectra, preprocessing is performed on the hyperspectral image and the land cover endmember spectra, including: Obtain the image description file of the hyperspectral image, and perform radiometric calibration on the hyperspectral image according to the image description file to obtain the radiometrically calibrated surface reflectance; The surface reflectance is atmospherically corrected to obtain the atmospherically corrected surface reflectance. The end-member spectra of the ground features are resampled according to the bands of the hyperspectral image, so that the end-member spectra of the ground features are aligned with the hyperspectral image in the spectral dimension.

4. The hybrid pixel decomposition method according to claim 1, characterized in that, The generation of the training dataset based on the hyperspectral image includes: The hyperspectral image is subjected to specular reflection filling to obtain a filled image; The filled image is cut into a preset number of image blocks, and the image blocks are squares of the same size; The training dataset is generated based on the image patches.

5. The hybrid pixel decomposition method according to claim 4, characterized in that, The step of inputting the training dataset and the endmember spectra of the ground features into the pixel decomposition model to optimize the model parameters of the pixel decomposition model includes: The training dataset and the ground feature endmember spectra are loaded using the data loader. The abundance prior estimate of the image block is calculated using the three-dimensional convolutional encoder; The dynamic relationship between the image patch and the endmember spectrum of the ground feature is calculated using a multi-head cross-attention module; The abundance prior estimate is optimized using an attention score through a gating guidance module to obtain an abundance vector; A reconstructed image is generated using a linear decoder, and the similarity between the reconstructed image and the endmember spectrum of the ground feature is calculated based on a loss function. The model parameters of the pixel decomposition model are then updated based on the similarity.

6. The hybrid pixel decomposition method according to claim 5, characterized in that, The 3D convolutional encoder includes multiple 3D convolutional modules, which acquire the spatial and spectral features of the training dataset and establish the correlation between the spatial features and the spectral features.

7. The hybrid pixel decomposition method according to claim 6, characterized in that, The multi-head cross-attention module includes multiple attention modules, which are used to calculate cross-attention.

8. The hybrid pixel decomposition method according to claim 7, characterized in that, The loss function includes spectral reconstruction loss and a sparsity regularization term; The spectral reconstruction loss is used to characterize the difference in the angle between the spectral vectors of the reconstructed image and the endmember spectrum of the ground feature; The sparse regularization term is used to regularize the sparsity of the abundance vector.

9. The hybrid pixel decomposition method according to claim 1, characterized in that, After obtaining the endmember abundance image after pixel decomposition, the following is also included: The accuracy of the endmember abundance image is evaluated based on the overall root mean square error of abundance and the abundance signal reconstruction error.

10. A hybrid pixel decomposition system for hyperspectral images, used to perform the hybrid pixel decomposition method for hyperspectral images as described in any one of claims 1 to 9, characterized in that, include: The data acquisition module is configured to acquire hyperspectral images and the end-member spectra of land cover corresponding to the hyperspectral images, wherein the hyperspectral images contain at least two types of land cover, and the type and number of the end-member spectra of the land cover match the types of land cover. The training set generation module is configured to generate a training dataset based on the hyperspectral image. The model training module is configured to input the training dataset and the endmember spectrum of the ground features into the pixel decomposition model to optimize the model parameters of the pixel decomposition model; the pixel decomposition model includes a data loader, a three-dimensional convolutional encoder, a multi-head cross-attention module, a gating guidance module, and a linear decoder; The model prediction module is configured to: use the optimized pixel decomposition model to infer and stitch together the training dataset to obtain the endmember abundance image after pixel decomposition.

Images