Spectral super-resolution reconstruction of multispectral images based on low-rank coupled dictionary learning

Spectral super-resolution reconstruction uses auxiliary information and sample learning to mine the spectral mapping relationship from multispectral images to hyperspectral images. Estimating the regression matrix from pairs of multispectral and hyperspectral images is an underdetermined problem, and prior information is often beneficial for the model to seek a more accurate spectral mapping relationship. Therefore, based on the spectral low-rank of hyperspectral images, the spectral super-resolution reconstruction is regarded as the problem of image low-rank reconstruction, and a spectral super-resolution reconstruction method based on low-rank coupled dictionary learning is proposed. Firstly, the method creates a coupling dictionary for multispectral and hyperspectral images with different spectral resolutions in the overlapping region, integrates the minimization of dictionary rank into the sparse representation of dictionary learning, and derives and constructs the optimized learning process of spectral dictionary and sparse coefficient based on ADMM algorithm, thereby reducing sparse error propagation and redundant information in the images. The obtained low-rank coupled dictionary ensures stable reconstruction of the images. Subsequently, the sparse coefficient of the multispectral images of the reconstruction region are utilized, combined with the low-rank dictionary of the hyperspectral images, to achieve spectral super-resolution reconstruction. To investigate the accuracy of the proposed algorithm, experiments were conducted using two sets of real datasets, ZY1-02D and GF5. The experimental results indicate that, compared to the contrast methods, the reconstruction accuracy of the proposed method has improved from the perspectives of element reconstruction quality (RMSE and ERGAS), spatial reconstruction quality (PSNR), spectral reconstruction quality (SAM), and spatial structural reconstruction quality (SSIM). To explore the application value of the proposed method, multispectral images from environments similar to the dataset used in this paper were selected as experimental subjects. Using the optimized dictionary from this paper, high-quality spectral super-resolution products were reconstructed more conveniently through transfer learning. The experimental results confirm the feasibility of the proposed method in practical application environments, effectively reducing the cost of obtaining hyperspectral images.