Self-calibration hyperspectral image super-resolution reconstruction method based on deep learning
The self-calibrated hyperspectral image super-resolution reconstruction method based on deep learning solves the problems of difficult-to-accurate characterization of prior assumptions and high computational complexity in hyperspectral image fusion, and generates high-resolution and high-fidelity hyperspectral images to meet the real-time requirements of geological exploration tasks.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING UNIV OF TECH
- Filing Date
- 2026-02-03
- Publication Date
- 2026-05-15
AI Technical Summary
Existing technologies for hyperspectral image fusion suffer from problems such as difficulty in accurately characterizing complex features based on prior assumptions, high computational complexity, bias towards low-frequency information, and spectral feature shifts, resulting in unsatisfactory fusion results that fail to meet the real-time requirements of geological exploration tasks.
A deep learning-based self-calibrated hyperspectral image super-resolution reconstruction method is adopted. A spectral distortion correction module is constructed by global spectral feature extraction and physical prior embedding. Combined with a spatial-spectral parallel super-resolution network and a spectral angle-gated fusion module, spectral distortion correction loss and fusion loss functions are designed, network parameters are optimized, and high-resolution hyperspectral images are generated.
It effectively eliminates spectral distortion noise, improves the robustness of image reconstruction and the spatial-spectral fidelity of the fusion results, achieves efficient fusion, and supports the data requirements of geological exploration tasks.
Smart Images
Figure CN122048657A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of hyperspectral remote sensing image quality enhancement technology, and is a self-calibrated hyperspectral image super-resolution reconstruction method based on deep learning. It is used to reconstruct hyperspectral images with high spatial resolution, providing a reliable data foundation for tasks such as geological exploration and resource exploration. Background Technology
[0002] Regularized modeling-based fusion methods typically incorporate one or more prior assumptions, transforming the fusion task of hyperspectral and multispectral images into a convex optimization problem with prior constraints, and then using iterative optimization algorithms to find the optimal solution. However, the spatial-spectral information in hyperspectral images is significantly complex, and manually set prior assumptions are insufficient to accurately characterize the complex features in the images, often leading to insufficient stability of the fusion results. Furthermore, the computational complexity of the iterative optimization algorithms upon which regularized modeling methods rely increases significantly with the increase in spatial resolution and the number of spectral channels in remote sensing images. In practical scenarios of geological remote sensing, such as emergency geological disaster monitoring and large-scale regional geological surveys, the real-time requirements for data processing are extremely high, and the long processing time of traditional iterative algorithms has become a key bottleneck restricting their practical application.
[0003] Currently, deep learning-based hyperspectral and multispectral image fusion techniques have achieved significant results. For example, MHF-Net combines observation models with low-rank priors for iterative fusion, while SDAGE uses graph convolution to construct spatial-spectral maps and stitch together basic features to recover spatial-spectral information. However, most existing deep learning networks inherently favor low-frequency information, tending to prioritize retaining low-frequency information while ignoring crucial high-frequency features. Consequently, the inherent spatial and spectral properties of the dual-source data are not fully explored and utilized, ultimately resulting in unsatisfactory spatial-spectral fidelity in the fusion results. Furthermore, existing fusion methods generally assume that the data to be fused does not have systematic biases, neglecting to address spectral feature shifts caused by interference from atmospheric water vapor and other media in hyperspectral data. Summary of the Invention
[0004] To ensure high resistance to physical noise and maintain the rich spectral information of hyperspectral images while preserving the fine spatial information of multispectral images, this invention proposes a deep learning-based self-calibrated hyperspectral image super-resolution reconstruction method. This method effectively eliminates spectral distortion noise in hyperspectral data while fusing hyperspectral and multispectral images, ultimately generating a high-fidelity high-resolution hyperspectral image. This provides effective data support for soil property mapping, geological structure identification, and mineral distribution location. The specific steps include:
[0005] S1 normalizes the hyperspectral and multispectral data.
[0006] S2 combines global spectral feature extraction and physical prior embedding, and constructs a spectral distortion correction module based on the attention mechanism to eliminate the spectral distortion noise that is common in hyperspectral data.
[0007] S3 proposes a spatial-spectral parallel super-resolution network to fuse normalized multispectral data and spectrally corrected hyperspectral data to generate high-resolution hyperspectral targets. This network mainly comprises a spatial super-resolution subnet, a spectral super-resolution subnet, a spectral angle-gated fusion module, and spatial and spectral degradation modules. The spatial super-resolution subnet utilizes the spatial information of the multispectral data to enhance the spatial resolution of the hyperspectral data, while the spectral super-resolution subnet uses the spectral information of the hyperspectral data to supplement the spectral resolution of the multispectral data. Both the spatial and spectral supramolecular networks are composed of dense residual units connected to Fourier domain cross-attention units. The dense residual units in the spatial supramolecular network employ two-dimensional convolution, while those in the spectral supramolecular network employ three-dimensional convolution. Based on the spectral angle similarity between the spatial super-resolution results of the hyperspectral data and the spectral super-resolution results of the multispectral data, a spectral angle-gated fusion module is designed to achieve selective fusion of the dual-path super-resolution results. Finally, the spatial and spectral degradation modules are used to perform spatial and spectral degradation on the fused high-resolution hyperspectral results, thereby achieving self-supervised fusion driven by the original dual-source input.
[0008] S4 is designed to optimize the parameters in the spectral correction module by considering spectral smoothness and loss function for spectral distortion correction.
[0009] S5 is designed to preserve spectral consistency through fusion loss, used to optimize parameters in a spatial-spectral parallel super-resolution network.
[0010] S6 processes the high- and multispectral data in the test set sequentially according to S1-S3, outputs the results of the spectral angle-gated fusion module, and performs inverse normalization on it to obtain the high-resolution hyperspectral target.
[0011] Furthermore, S1 includes the following steps:
[0012] S101: Transfer hyperspectral data Each pixel in the data is divided by the maximum value among all pixels in the hyperspectral and multispectral data, thereby assigning the pixel values of the hyperspectral data to the interval. :
[0013]
[0014] in, , , This represents the number of channels, rows, and columns of the hyperspectral data, in that order. , They represent the first The first band, the first line, number List, To perform the maximum value operation, This is normalized hyperspectral data.
[0015] S102: Transfer multispectral data Each pixel in the multispectral data is divided by the maximum value among all pixels in the high- and multispectral data, thereby assigning the pixel values of the multispectral data to the interval. :
[0016]
[0017] in, The numbers represent the number of channels, rows, and columns of the multispectral data, respectively. They represent the first The first band, the first line, number List, , This is normalized multispectral data.
[0018] Furthermore, S2 includes the following steps:
[0019] S201: Normalized hyperspectral data Spatial average pooling is performed to obtain the global spectral feature vector. :
[0020]
[0021] S202: Using dimension concatenation operations, the global spectral feature vector obtained in the previous step is... Prior vector of atmospheric absorption peak Combined into a fused feature vector .about If the hyperspectral data contains the first If there is an abnormal absorption peak in each band, then =0; if there is no distortion in this band, then The value is 1. Among them, for The first in Each element.
[0022] S203: Will The input is a gated block consisting of a concatenated convolutional layer, a ReLU function, and a sigmoid function; the output of the gated block... Dimension reset to obtain channel gain .
[0023] S204: Achieving attention through a multi-attention module guided by channel gain. Spectral distortion correction was performed to obtain corrected hyperspectral data. .
[0024] Specifically, with the first Taking the attention module as an example: the first The input to each attention module is the output of the previous attention module. With channel gain Then the first The output of each attention module is:
[0025]
[0026] in Represents matrix multiplication. This represents the Sigmoid function. This indicates the transpose operation. Represents a two-dimensional convolutional layer. The input to the first attention module... For normalized hyperspectral data The result of dimension reset This is the output of the last attention module. The result of dimension reset.
[0027] Furthermore, S3 includes the following steps:
[0028] S301: Using bilinear interpolation method to... Spatial upsampling is performed to obtain Normalized multispectral data were analyzed using cubic spline interpolation. Spectral upsampling was performed to obtain .
[0029] S302: Utilizing dense residual units based on 3D convolution to respectively... and Feature extraction is performed to obtain hyperspectral features. With multispectral features And Fourier domain cross-attention units are used to injection Get the injection result Next, respectively... indivual and The channel splicing results and the previous indivual and The channel splicing results input multiple hyperspectral data in the branch is the first Dense residual elements.
[0030] Specifically, the branches where the multispectral and hyperspectral data are located are respectively... Input of a dense residual unit For example, then
[0031]
[0032] in,
[0033]
[0034] in This indicates a dimension concatenation operation. This represents a dense residual unit based on 3D convolution. This represents a Fourier domain cross-attention unit. .also, .
[0035] S303: The features output by the last Fourier domain cross-attention unit in S302. and The super-resolution result of the multispectral image is obtained by adding them together. .
[0036] S304: Utilizing dense residual units based on 2D convolution to... and Feature extraction was performed to obtain hyperspectral features. With multispectral features And Fourier domain cross-attention units are used to injection Get the injection result Next, respectively... indivual and The channel splicing results and the previous indivual and The channel splicing result is input to the branch containing the high- and multispectral data. Dense residual elements.
[0037] Specifically, based on the branch where the high- and multispectral data are located... Input of a dense residual unit For example, then
[0038]
[0039] in,
[0040]
[0041] in This represents a dense residual unit based on two-dimensional convolution. ;also, , , ;
[0042] S305: The features output by the last Fourier domain cross-attention unit in S304. and The spatial super-resolution result of the hyperspectral image is obtained by adding them together. .
[0043] S306: Based on spatial super-resolution results and spectral super-resolution results Calculate the spectral angle similarity diagram :
[0044]
[0045] in, This represents a minimum value and is used to prevent invalid calculations.
[0046] S307: Similarity map The input consists of a gated block composed of convolutional layers, ReLU functions, and sigmoid functions, resulting in a fused weight map. .
[0047] S308: In the weighted graph Guided by this, multiplication and addition units are used to implement... and Spatial-spectral fusion yields fused high-resolution hyperspectral data. :
[0048]
[0049] in, This indicates element-wise multiplication.
[0050] S309: Spatial and spectral degradation processing is performed on the fusion results using the point diffusion model and spectral response matrix, respectively, to obtain the reconstructed hyperspectral and multispectral images. and :
[0051]
[0052] in, This is the spectral response matrix. For downsampling operation, is the blur kernel, and * represents the convolution operation.
[0053] Furthermore, S4 includes the following steps:
[0054] S401: Calculate the spectral distortion correction loss :
[0055]
[0056] in It is the Frobenius norm. For second-order spectral gradient operators, It is a scale regulation factor.
[0057] S402: Minimize the distortion correction loss of the hyperspectral data in the training set. To achieve the goal, the Adam optimizer is used to optimize the network parameters involved in S2 until... Less than the preset value.
[0058] Furthermore, S5 includes the following steps:
[0059] S501: Calculate fusion loss :
[0060]
[0061] in, For hyperspectral reconstruction loss, For multispectral reconstruction loss, For spectral angle loss, , Scale regulation factor. Specifically,
[0062]
[0063]
[0064] S502: Aiming to minimize the fusion loss of hyperspectral and multispectral data in the training set, optimize the network parameters involved in S3 using the Adam optimizer until... Less than the preset value.
[0065] Furthermore, S6 includes the following steps:
[0066] S601: Sequentially process the hyperspectral and multispectral data to be fused in the test set. and Execute S1-S3 to output the results of the spectral angle-gated fusion module. .
[0067] S602: Yes Perform inverse normalization to obtain the final fusion result. :
[0068] .
[0069] The technical effects of this invention are as follows:
[0070] 1. The hyperspectral image reconstruction method provided by this invention combines global spectral feature extraction and physical prior embedding to construct a spectral distortion correction module, which effectively eliminates the spectral distortion noise that is common in hyperspectral data, thereby improving the robustness of image reconstruction against spectral distortion.
[0071] 2. The hyperspectral image reconstruction method provided by this invention proposes a spatial-spectral parallel super-resolution network. Based on dense residual units and Fourier domain cross-attention units, a spatial-spectral parallel supramolecular network is designed. Based on a gating mechanism, a spectral angle-gated fusion module is constructed to achieve effective fusion of multi- and hyperspectral data and generate high-resolution hyperspectral targets.
[0072] 3. The hyperspectral image reconstruction method provided by this invention is designed with a spectral distortion correction loss of dual variational regularization, which effectively drives the parameter optimization of the spectral distortion correction module while ensuring the spectral smoothness of hyperspectral data. Attached Figure Description
[0073] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0074] Figure 1 This is a flowchart of the method of the present invention;
[0075] Figure 2 This is the spectral distortion correction module of the present invention;
[0076] Figure 3 This is the dense residual unit of the present invention;
[0077] Figure 4 This is the Fourier domain cross-attention unit of the present invention;
[0078] Figure 5 This is the spectral angle-gated fusion module of the present invention. Detailed Implementation
[0079] To address the demand for high-resolution hyperspectral remote sensing data in geological exploration, this invention provides a deep learning-based self-calibrated hyperspectral image super-resolution reconstruction method. This method effectively corrects spectral distortion while improving the spatial resolution of hyperspectral data. The self-calibrated super-resolution method for remote sensing images provided by this invention is as follows: Figure 1 As shown, the spectral correction module, dense residual unit, Fourier domain cross-attention unit, and spectral angle gating fusion module involved are respectively as follows: Figure 2 , Figure 3 , Figure 4 and Figure 5 As shown.
[0080] The process of the present invention will be described below with specific embodiments:
[0081] In this embodiment, the hyperspectral image has significant water vapor absorption peak characteristics, and the spatial resolution of the multispectral image is four times that of the hyperspectral image. Taking two-source remote sensing data with a spatial resolution difference of four times as an example, the super-resolution method provided by this invention will be specifically described as follows:
[0082] Step S1 involves normalizing the hyperspectral and multispectral data. This specifically includes the following steps:
[0083] S101: Transfer hyperspectral data Each pixel in the data is divided by the maximum value among all pixels in the hyperspectral and multispectral data, thereby assigning the pixel values of the hyperspectral data to the interval. :
[0084]
[0085] in, , , This represents the number of channels, rows, and columns of the hyperspectral data, in that order. , They represent the first The first band, the first line, number List, To perform the maximum value operation, This is normalized hyperspectral data.
[0086] S102: Transfer multispectral data Each pixel in the multispectral data is divided by the maximum value among all pixels in the high- and multispectral data, thereby assigning the pixel values of the multispectral data to the interval. :
[0087]
[0088] in, The numbers represent the number of channels, rows, and columns of the multispectral data, respectively. They represent the first The first band, the first line, number List, , This is normalized multispectral data.
[0089] Step S2 involves constructing a spectral distortion correction module based on an attention mechanism, combining global spectral feature extraction and physical prior embedding. This includes the following steps:
[0090] S201: Normalized hyperspectral data Spatial average pooling is performed to obtain the global spectral feature vector. :
[0091]
[0092] S202: Using dimension concatenation operations, the global spectral feature vector obtained in the previous step is... Prior vector of atmospheric absorption peak Combined into a fused feature vector .about If the hyperspectral data contains the first If there is an abnormal absorption peak in each band, then =0; if there is no distortion in this band, then The value is 1. Among them, for The first in Each element.
[0093] S203: Will The input is a gated block consisting of a concatenated convolutional layer, a ReLU function, and a sigmoid function; the output of the gated block... Dimension reset to obtain channel gain .
[0094] S204: Achieving attention through a multi-attention module guided by channel gain. Spectral distortion correction was performed to obtain corrected hyperspectral data. .
[0095] Specifically, with the first Taking the attention module as an example: the first The input to each attention module is the output of the previous attention module. With channel gain Then the first The output of each attention module is:
[0096]
[0097] in Represents matrix multiplication. This represents the Sigmoid function. This indicates the transpose operation. Represents a two-dimensional convolutional layer. The input to the first attention module... For normalized hyperspectral data The result of dimension reset This is the output of the last attention module. The result of dimension reset.
[0098] Step S3 proposes a spatial-spectral parallel super-resolution network to fuse normalized multispectral data and spectrally corrected hyperspectral data to generate high-resolution hyperspectral targets. Specifically, it includes the following steps:
[0099] S301: Using bilinear interpolation method to... Spatial upsampling is performed to obtain Normalized multispectral data were analyzed using cubic spline interpolation. Spectral upsampling was performed to obtain .
[0100] S302: Utilizing dense residual units based on 3D convolution to respectively... and Feature extraction is performed to obtain hyperspectral features. With multispectral features And Fourier domain cross-attention units are used to injection Get the injection result Next, respectively... indivual and The channel splicing results and the previous indivual and The channel splicing results input multiple hyperspectral data in the branch is the first Dense residual elements.
[0101] Specifically, the branches where the multispectral and hyperspectral data are located are respectively... Input of a dense residual unit For example, then
[0102]
[0103] in,
[0104]
[0105] in This indicates a dimension concatenation operation. This represents a dense residual unit based on 3D convolution. This represents a Fourier domain cross-attention unit. ,also, , , ;
[0106] S303: The features output by the last Fourier domain cross-attention unit in S302. and The super-resolution result of the multispectral image is obtained by adding them together. .
[0107] S304: Utilizing dense residual units based on 2D convolution to... and Feature extraction was performed to obtain hyperspectral features. With multispectral features And Fourier domain cross-attention units are used to injection Get the injection result Next, respectively... indivual and The channel splicing results and the previous indivual and The channel splicing result is input to the branch containing the high- and multispectral data. Dense residual elements.
[0108] Specifically, based on the branch where the high- and multispectral data are located... Input of a dense residual unit For example, then
[0109]
[0110] in,
[0111]
[0112] in This represents a dense residual unit based on two-dimensional convolution. ;also, , , ;
[0113] S305: The features output by the last Fourier domain cross-attention unit in S304. and The spatial super-resolution result of the hyperspectral image is obtained by adding them together. .
[0114] S306: Based on spatial super-resolution results and spectral super-resolution results Calculate the spectral angle similarity diagram :
[0115]
[0116] in, This represents a minimum value and is used to prevent invalid calculations.
[0117] S307: Similarity map The input consists of a gated block composed of convolutional layers, ReLU functions, and sigmoid functions, resulting in a fused weight map. .
[0118] S308: In the weighted graph Guided by this, multiplication and addition units are used to implement... and Spatial-spectral fusion yields fused high-resolution hyperspectral data. :
[0119]
[0120] in, This indicates element-wise multiplication.
[0121] S309: Spatial and spectral degradation processing is performed on the fusion results using the point diffusion model and spectral response matrix, respectively, to obtain the reconstructed hyperspectral and multispectral images. and :
[0122]
[0123] in, This is the spectral response matrix. For downsampling operation, is the blur kernel, and * represents the convolution operation.
[0124] Step S4 involves designing a spectral distortion correction loss function that considers spectral smoothness, used to optimize the parameters in the spectral correction module. This includes the following steps:
[0125] S401: Calculate the spectral distortion correction loss :
[0126]
[0127] in It is the Frobenius norm. For second-order spectral gradient operators, It is a scale regulation factor.
[0128] S402: Minimize the distortion correction loss of the hyperspectral data in the training set. To achieve the goal, the Adam optimizer is used to optimize the network parameters involved in S2 until... Less than the preset value.
[0129] Step S5 involves designing a fusion loss that preserves spectral consistency to optimize the parameters in the spatial-spectral parallel super-resolution network. This includes the following steps:
[0130] S501: Calculate fusion loss :
[0131]
[0132] in, For hyperspectral reconstruction loss, For multispectral reconstruction loss, For spectral angle loss, , Scale regulation factor. Specifically,
[0133]
[0134]
[0135] S502: Aiming to minimize the fusion loss of hyperspectral and multispectral data in the training set, optimize the network parameters involved in S3 using the Adam optimizer until... Less than the preset value.
[0136] Step S6 involves processing the hyperspectral and multispectral data in the test set sequentially according to S1-S3, outputting the results from the spectral angle-gated fusion module, and then performing inverse normalization to obtain the high-resolution hyperspectral target. Specifically, this includes the following steps:
[0137] S601: Sequentially process the hyperspectral and multispectral data to be fused in the test set. and Execute S1-S3 to output the results of the spectral angle-gated fusion module. .
[0138] S602: Yes Perform inverse normalization to obtain the final fusion result. :
[0139] .
Claims
1. A self-calibrated hyperspectral image super-resolution reconstruction method based on deep learning, characterized in that, Includes the following steps: S1, normalizes the hyperspectral and multispectral data; S2 combines global spectral feature extraction and physical prior embedding, and constructs a spectral distortion correction module based on the attention mechanism to eliminate the spectral distortion noise that is common in hyperspectral data; S3 proposes a spatial-spectral parallel super-resolution network to fuse normalized multispectral data and spectrally corrected hyperspectral data to generate high-resolution hyperspectral targets. The network includes a spatial super-resolution subnetwork, a spectral super-resolution subnetwork, a spectral angle-gated fusion module, and spatial and spectral degradation modules. The spatial super-resolution subnet uses the spatial information of multispectral data to enhance the spatial resolution of hyperspectral data, while the spectral super-resolution subnet uses the spectral information of hyperspectral data to supplement the spectral resolution of multispectral data. Both spatial and spectral supramolecular networks are composed of dense residual units connected by Fourier domain cross-attention units. The dense residual units in the spatial supramolecular network employ two-dimensional convolution, while those in the spectral supramolecular network employ three-dimensional convolution. Based on the spectral angle similarity between the spatial super-resolution results of hyperspectral data and the spectral super-resolution results of multispectral data, a spectral angle-gated fusion module is designed to achieve selective fusion of the dual-path super-resolution results. Finally, spatial and spectral degradation modules are used to perform spatial and spectral degradation on the fused hyperspectral results, respectively, thereby achieving self-supervised fusion driven by the original dual-source input. S4, Design a spectral distortion correction loss function that takes into account spectral smoothness, to optimize the parameters in the spectral correction module; S5, designed to preserve spectral consistency fusion loss, is used to optimize parameters in a spatial-spectral parallel super-resolution network; S6 processes the high- and multispectral data in the test set sequentially according to S1-S3, outputs the results of the spectral angle-gated fusion module, and performs inverse normalization on it to obtain the high-resolution hyperspectral target.
2. The method for self-calibrating hyperspectral image super-resolution reconstruction based on deep learning according to claim 1, characterized in that, S1 includes the following steps: S101: Transfer hyperspectral data Each pixel in the data is divided by the maximum value among all pixels in the hyperspectral and multispectral data, thereby assigning the pixel values of the hyperspectral data to the interval. : ， in, , , The numbers represent the number of channels, rows, and columns of the hyperspectral data, respectively. , They represent the first The first band, the first line, number List, To perform the maximum value operation, Normalized hyperspectral data; S102: Transfer multispectral data Each pixel in the multispectral data is divided by the maximum value among all pixels in the high- and multispectral data, thereby assigning the pixel values of the multispectral data to the interval. : ， in, The numbers represent the number of channels, rows, and columns of the multispectral data, respectively. They represent the first The first band, the first line, number List, , This is normalized multispectral data.
3. The method for self-calibrating hyperspectral image super-resolution reconstruction based on deep learning according to claim 2, characterized in that, S2 includes the following steps: S201: Normalized hyperspectral data Spatial average pooling is performed to obtain the global spectral feature vector. : ， S202: Using dimension concatenation operations, the global spectral feature vector obtained in the previous step is... Prior vector of atmospheric absorption peak Combined into a fused feature vector ; about If the hyperspectral data contains the first If there is an abnormal absorption peak in each band, then =0; if there is no distortion in this band, then =1; where, for The first in One element; S203: Will The input is a gated block consisting of a concatenated convolutional layer, a ReLU function, and a sigmoid function; the output of the gated block... Dimension reset to obtain channel gain ; S204: Achieving attention through a multi-attention module guided by channel gain. Spectral distortion correction was performed to obtain corrected hyperspectral data. ; Specifically, no. The input to each attention module is the output of the previous attention module. With channel gain Then the first The output of each attention module is: ， in Represents matrix multiplication. This represents the Sigmoid function. This indicates the transpose operation. Represents a two-dimensional convolutional layer. The input to the first attention module; For normalized hyperspectral data The result of dimension reset This is the output of the last attention module. The result of dimension reset.
4. The self-calibrating hyperspectral image super-resolution reconstruction method based on deep learning according to claim 3, characterized in that, S3 includes the following steps: S301: Using bilinear interpolation method for... Spatial upsampling is performed to obtain Normalized multispectral data were analyzed using cubic spline interpolation. Spectral upsampling was performed to obtain ; S302: Utilizing dense residual units based on 3D convolution to respectively... and Feature extraction is performed to obtain hyperspectral features. With multispectral features And Fourier domain cross-attention units are used to injection Get the injection result Next, respectively the front indivual and The channel splicing results and the previous indivual and The channel splicing results input multiple hyperspectral data in the branch is the first A dense residual element; Specifically, for the branch where multispectral and hyperspectral data are located... Input of a dense residual unit ,but ， in, ， in This indicates a dimension concatenation operation. This represents a dense residual unit based on 3D convolution. This represents a Fourier domain cross-attention unit. ,also, , , ; S303: The features output by the last Fourier domain cross-attention unit in S302. and The super-resolution result of the multispectral image is obtained by adding them together. ; S304: Utilizing dense residual units based on 2D convolution to... and Feature extraction was performed to obtain hyperspectral features. With multispectral features And Fourier domain cross-attention units are used to injection Get the injection result Next, respectively the front indivual and The channel splicing results and the previous indivual and The channel splicing result is input to the branch containing the high- and multispectral data. A dense residual element; Specifically, for the branch where the high- and multispectral data are located... Input of a dense residual unit ,but ， in, ， in This represents a dense residual unit based on two-dimensional convolution. ;also, , , ; S305: The features output by the last Fourier domain cross-attention unit in S304. and The spatial super-resolution result of the hyperspectral image is obtained by adding them together. ; S306: Based on spatial super-resolution results and spectral super-resolution results Calculate the spectral angle similarity diagram : ， in, This represents a minimum value and is used to prevent invalid calculations. S307: Similarity map The input consists of a gated block composed of convolutional layers, ReLU functions, and sigmoid functions, resulting in a fused weight map. ; S308: In the weighted graph Guided by this, multiplication and addition units are used to implement... and Spatial-spectral fusion yields fused high-resolution hyperspectral data. : ， in, Indicates element-wise multiplication; S309: Spatial and spectral degradation processing is performed on the fusion results using the point diffusion model and spectral response matrix, respectively, to obtain the reconstructed hyperspectral and multispectral images. and : ， in, This is the spectral response matrix. For downsampling operation, is the blur kernel, and * represents the convolution operation.
5. The self-calibrating hyperspectral image super-resolution reconstruction method based on deep learning according to claim 4, characterized in that, S4 includes the following steps: S401: Calculate the spectral distortion correction loss : ， in It is the Frobenius norm. For second-order spectral gradient operators, As a scale control factor; S402: Minimize the distortion correction loss of the hyperspectral data in the training set. To achieve the goal, the Adam optimizer is used to optimize the network parameters involved in S2 until... Less than the preset value.
6. The method for self-calibrating hyperspectral image super-resolution reconstruction based on deep learning according to claim 5, characterized in that, S5 includes the following steps: S501: Calculate fusion loss : ，， in, For hyperspectral reconstruction loss, For multispectral reconstruction loss, For spectral angle loss, , Scale regulation factor; specifically, ， ， S502: Aiming to minimize the fusion loss of hyperspectral and multispectral data in the training set, optimize the network parameters involved in S3 using the Adam optimizer until... Less than the preset value.
7. The method for self-calibrating hyperspectral image super-resolution reconstruction based on deep learning according to claim 6, characterized in that, S6 includes the following steps: S601: Sequentially process the hyperspectral and multispectral data to be fused in the test set. and Execute S1-S3 to output the results of the spectral angle-gated fusion module. ; S602: Yes Perform inverse normalization to obtain the final fusion result. : 。