Hyperspectral Image Superresolution by Transfer Learning

Hyperspectral image superresolution is a highly attractive topic in computer vision and has attracted many researchers' attention. However, nearly all the existing methods assume that multiple observations of the same scene are required with the observed low-resolution hyperspectral image. This limits the application of superresolution. In this paper, we propose a new framework to enhance the resolution of hyperspectral images by exploiting the knowledge from natural images: The relationship between low/high-resolution images is the same as that between low/high-resolution hyperspectral images. In the proposed framework, the mapping between low- A nd high-resolution images can be learned by deep convolutional neural network and be transferred to hyperspectral image by borrowing the idea of transfer learning. In addition, to study the spectral characteristic between low- A nd high-resolution hyperspectral image, collaborative nonnegative matrix factorization (CNMF) is proposed to enforce collaborations between the low- A nd high-resolution hyperspectral images, which encourages the estimated solution to extract the same endmembers with low-resolution hyperspectral image. The experimental results on ground based and remote sensing data suggest that the proposed method achieves comparable performance without requiring any auxiliary images of the same scene.