Developing unsupervised learning techniques to detect mineral and geological structures using hyperspectral data.

In mineralogical studies, the use of hyperspectral images helps to detect different minerals in a faster and more accurate manner. However, it can be problematic to process such data, as the acquisition of sufficient ground truth information is challenging. Therefore, the main goal of this research is to develop unsupervised learning techniques (sparse subspace-based clustering algorithms) to distinguish different mineralogical features from hyperspectral images. As a first milestone for the proposed research topic, a new sparse subspace-based clustering algorithm was developed to cluster the mineralogical features into meaningful groups. The proposed algorithm is able to process highly mixed and complex data in a robust and fast manner. As a follow-up, spatial information was exploited in a newly developed clustering algorithm. Such information helps to take into account spatial structures of mineralogical samples. By including the spatial information, the precision of the proposed algorithm increased. In the period to come, we will explore the potential of deep clustering methods, based on state-of-the-art sparse autoencoder-based algorithms, or the concepts of deep image priors. As a result of the study, fast and robust deep learning-based clustering algorithms will be developed, to analyze complex hyperspectral images from mineralogical and geological features.

Keywords:HYPERSPECTRAL DATA ANALYSIS, HYPERSPECTRAL REMOTE SENSING

Disciplines:Remote sensing, Photogrammetry and remote sensing