A Novel Approach for Hyperspectral Image Superresolution Using Spectral Unmixing and Transfer Learning

Hyperspectral image (HSI) Super-resolution (SR) methods enhance the spatial resolution. In this paper, we propose a novel SR approach for HSIs by making use of spectral unmixing and transfer learning. We first train a deep convolutional neural network (CNN) to learn the mapping between the low-resolution (LR) and high-resolution (HR) natural images and use the same for transfer learning to get the initial estimates of the super-resolved abundances where the input corresponds to LR abundances. To get the better estimates of abundances and hence improve the SR of HSIs, we use a regularization framework in which both the LR and HR abundances are modelled as Inhomogeneous Gaussian Markov field (IGMRF) that serves as the prior. Finally, the SR HSIs are obtained by using a linear mixing model that uses the SR abundances and the endmembers estimated using an appropriate technique. Experiments on synthetic as well as on real HSIs show that the proposed method performs better when compared to other existing approaches. The advantages of the proposed approach are: 1. The method do not require auxiliary image as used in many of the existing methods, 2. Spectral details are better preserved since the SR is carried out in abundance domain, 3. Computational complexity is reduced since the SR is carried out on abundances which are few in number when compared to HSIs.