Exploration of utility of data dimensionality in multispectral and hyperspectral image classification

Multispectral image classification has been widely used in land cover/land use in remote sensing community for various applications. Hyperspectral image classification is recently developed for target detection and classification with particular interest in man-made rather than natural targets. Due to different focuses in applications, their utilities are also different. Unfortunately, a common consensus seems to mislead to a misconception that hyperspectral image classification seems a natural extension of multispectral image classification. As a result, it is expected that many multispectral image processing techniques are readily applied to hyperspectral image processing via simple straightforward extensions. This thesis shows otherwise by investigating and exploring the basic design philosophy in processing multispectral and hyperspectral images in how to effectively use data dimensionality. Due to the use of an insufficient number of spectral channels for data collection multispectral image processing generally requires dimensionality expansion for data accommodation and analysis. Two approaches are developed for this purpose, Band Generation Process (BGP) for spectral band dimensionality expansion and kernel-based Principal Components Analysis (KPCA) for feature dimensionality expansion. To the contrary, due to overwhelming spectral information provided by hundreds of contiguous spectral channels hyperspectral image processing generally requires data dimensionality reduction in which case determining number of dimensions, q required to be retained becomes crucial prior to hyperspectral image processing. To address this issue a recently developed concept, called Virtual Dimensionality (VD) is used to estimate the value of q. Once the q is determined, two approaches to dimensionality reduction can be derived, band selection and components analysis which can be viewed as counterparts of BGP and KPCA respectively developed for multispectral image processing, band expansion and feature dimensionality expansion. Finally, this thesis concludes by conducting extensive experiments-based study and analysis to demonstrate that multispectral imaging and hyperspectral imaging are actually two different techniques.