Hyperspectral Image Restoration, Unmixing, and Classification by Low-Rank Tensor Recovery, Nonlocal Self-Similarity, and Deep Priors

Hyperspectral remote sensing images, recording electromagnetic information of subjects on Earth surface, have played an important role in countless applications, such as earth observation, environmental protection, and natural disaster monitoring. New-generation hyperspectral imagers tend to have a higher spectral resolution (i.e., to acquire images with more color channels), which enables precise material identification with higher accuracy. For example, a Chinese GaoFen-5 satellite, equipped with the Advanced Hyperspectral Imager (AHSI) and launched in May 2018, has 330 color channels covering the spectrum from visible to short-wave infrared. However, the ever-increasing spectral resolution of hyperspectral images (HSIs) is often obtained at the cost of noisy images, thus calling for effective denoising methods and data processing techniques. To address the complex noise and to improve the accuracy of information extracted from HSIs, this project will use China’s newly launched Gaofen-5 satellite image as the main data source, combine the advantages of cooperation between the mainland and Hong Kong teams, and study theoretical problems of hyperspectral image restoration, spectral unmixing, and classification by combining traditional machine learning and deep learning techniques. Our main focus will be proposing new mathematical models and algorithms for three typical inverse problems (restoration, spectral unmixing, and classification) in hyperspectral images by exploiting the data structure of hyperspectral images. We propose the following directions of research. (i) Hyperspectral image restoration by low-rank tensor recovery, nonlocal self-similarity, and deep priors. (ii) Hyperspectral image unmixing by low-rank tensor decomposition and training deep prior for abundance maps. (iii) Low-rank tensor decomposition and deep learning-based hyperspectral image classification. The results of this project are expected to greatly improve the quality of information extracted from HSIs. Therefore, the results of this project will be beneficial to earth observation research by improving the accuracy of information extracted from HSIs. Furthermore, it can expand the applications of HSIs and discover new remote sensing applications, which had been precluded by the poor quality of images.