Hyperspectral Image Super-Resolution Using a Hybrid Method Combining Deep Learning and Sparse Representation
TL;DR
The findings suggest that combining deep-learning-based spatial enhancement with sparse-representation-based spectral fusion can support lower spatial and spectral reconstruction errors under the evaluated conditions.
Abstract
Background/Aim: Hyperspectral images provide rich spectral information but often suffer from low spatial resolution due to hardware limitations and long imaging distances. This study aims to enhance the spatial resolution of low-resolution hyperspectral images while preserving their spectral information by proposing a hybrid super-resolution method that combines deep learning and sparse representation.Methods: The proposed framework comprises two components. Spatial resolution is enhanced using deep learning models, while the spectral information of the original low-resolution hyperspectral cube is reintroduced into the reconstructed image through Bayesian sparse-representation-based dictionary learning. A deep convolutional network and a deep residual network were adapted to hyperspectral images. To reduce the convolutional workload, two-dimensional convolution was applied to one band selected from every three bands. The method was evaluated on the CAVE and Harvard datasets at scale factors of ×2, ×3, and ×4 using root mean square error (RMSE) and spectral angle mapper (SAM), and was compared with bicubic interpolation, GFPCA, and SCT_SDCNN.Results: At the dataset-average level, the deep residual network-based hybrid model yielded the lowest RMSE and SAM values among the evaluated methods at all three scale factors. On CAVE, the RMSE/SAM values were 2.0146/2.8507, 2.1380/4.0742, and 2.4712/5.1211 at ×2, ×3, and ×4, respectively. The corresponding values on Harvard were 1.5753/2.1328, 1.7373/2.7135, and 2.1524/3.3576. Relative to SCT_SDCNN, the SAM reductions ranged from 19.12% to 24.80% on CAVE and from 21.93% to 26.02% on Harvard.Conclusion: The findings suggest that combining deep-learning-based spatial enhancement with sparse-representation-based spectral fusion can support lower spatial and spectral reconstruction errors under the evaluated conditions. The need for training data with characteristics comparable to the target data remains a limitation.