LOW RANK AND SPARSE SPACE DECOMPOSITION APPROACHES TO HYPERSPECTRAL DETECTION

Abstract

Hyperspectral imaging (HSI) has become an emerging remote sensing technology in recent years. In particular, HSI has shown its particular strength in target detection and anomaly detection due to its ability in detecting targets at mixed pixel and subpixel scale. However, such capability is also traded for detecting many unknown material substances from unknown background (BKG). This dissertations develops three approaches to resolving this issue. One is data sphering (DS) which can remove data statistics of the first two orders, specifically Gaussian BKG. Another is to use a popular low-rank and sparse matrix decomposition (LRaSMD) model to decompose a data space into low rank, sparse and noise subspace where the low rank and sparse subspaces can be specified by BKG and targets respectively. As a result, BKG suppression can be achieved in the low rank subspace, while performing target detection in the sparse subspace. A third approach combines DS and the LRaSMD model to further improve target detection and BKG suppression. In this dissertations we investigate these three approaches by exploring four well-known detectors, Orthogonal Subspace Projection (OSP)-based target detector, Constrained Energy Minimization (CEM) target detector, Target Constrained Interference-Minimized Filter (TCIMF) target detector and Reed-Xiaoli Anomaly Detector (RX-AD) for hyperspectral target detection (HTD) and hyperspectral anomaly detection (HAD) and further develop their BKG-annihilated counterparts, DS-OSP, BA-OSP, LRaSMD-OSP, BA-TCIMF.