Analysis on Hyperspectral Signature Coding (Paperback)

This dissertation addresses issues of hyperspectral signature coding, where three coding techniques from various applications, Arithmetic Coding (AC) in source coding, Texture Feature Coding Method (TFCM) in texture analysis and Block Truncation Coding (BTC) in image coding are investigated and further explored for hyperspectral signature characterization. Current coding-based approaches to spectral signature characterization include SPectral Analysis Manager (SPAM) developed by Mazer et al. and its extension Spectral Feature-based Binary Coding (SFBC) by Qian et al. where spectral signatures are encoded as code words and the spectral analysis is conducted by using Hamming distance as a spectral similarity measure. Unfortunately, due to the fact that the Hamming distance is a bit-wise (referred to as memoryless) measure, its performance is completely determined by spectral variation with individual bits without taking into account adjacent bits within a code word. Accordingly, one way to improve the performance of such bit memoryless spectral coding is to introduce bit-memory into its used distance measure. In order to address this issue two new coding techniques are developed, called Spectral Feature Probabilistic Coding (SFPC) which is based on the AC and Spectral Derivative Feature Coding (SDFC) which is based on TFCM. These two techniques extend memoryless signature coding to memory-signature coding and have shown to yield better performance than the current memoryless coding methods. Another alternative way to improve the performance of current memoryless coding schemes is called Block Truncation Signature Coding (BTSC) which is based on BTC and can process the signatures adaptively. This dissertation then culminates in a new approach, called Orthogonal Subspace Projection-based Band Signature for Signature Coding (OSP-BSSC) that can be used to perform signature dimensionality reduction for signature coding. It introduces a new concept, called Joint Dimensionality (JD) of a signature pair for discrimination and then further develops criteria and techniques to find a image of the JD for a signature to retain sufficient information for its discrimination from other signatures. Finally, extensive experiments are conducted to substantiate the techniques proposed in this dissertation for performance evaluation and analysis.