Progressive band selection and prioritization for linear spectral mixture analysis

Linear Spectral Mixture Analysis (LSMA) has been widely used in the remote sensing community. It assumes that a data sample vector is linearly mixed by a set of distinct signatures as a linear mixture from which it can be further unmixed as abundance fractions in terms of these signatures. While LSMA has shown to be a promising spectral unmixing technique in remote sensing image analysis, it also suffers from an issue of nonlinear separability encountered in both multispectral and hyperspectral image processing. To resolve this dilemma, a kernel-based LSMA (KLSMA) is proposed in this dissertation, which projects data samples into a high dimensional feature space to solve linearly non-separable problems. Similar techniques are further used to extend Fisher's LSMA (KFLSMA), Weighted Abundance Constrained LSMA (KWAC-LSMA) to their kernel-based versions, referred to as kernel-based FLSMA (KFLSMA), and kernel-based WAC-LSMA (KWAC-LSMA). Since hyperspectral imagery is generally acquired in hundreds of contiguous spectral channels with very high spectral resolution such high inter-band correlation provides high redundant information and can be removed with no significant loss of information. So, it has been a great interest in hyperspectral image analysis to seek a means of how to effectively reduce dimensionality without significantly compromising performance. Band Selection (BS) is one of commonly used approaches for this purpose. However, there are several crucial issues arising in BS which must be addressed, such as the number of bands required for BS to select, p, and what criterion needed to be used to select bands. In order to solve these issues, Band Prioritization (BP) is introduced in this dissertation from which Progressive Band Dimensionality Process (PBDP) is derived to rank bands according to BP which paves the way for a subsequently developed Progressive Band Selection (PBS). Due to practical issues the number of bands to be selected must adapt to various applications instead of being fixed at a constant as BS does. To further mitigate this problem, a new concept of Band Dimensionality Allocation (BDA) is introduced, which allows users to determine band dimensionality dynamically according to specific applications. When PBDP is implemented, one issue may arise in the fact that two highly prioritized bands may also share significant information in common. As a result, if one band is selected, the other band should be considered as a redundant band and must be removed by BS. To address this issue, the Band De-correlation (BD) is further proposed for this purpose. Finally, by implementing the PBDP in conjunction with the BDA and the BD a Progressive Band Selection (PBS) can be derived as an alternative to the traditional BS. The experiments conducted in this dissertation show that the PBS provides significant performance and advantages which could not be achieved by traditional BS.