Hyperspectral Data Processing: Algorithm Design and Analysis

6.9 Conclusions

Dimensionality reduction (DR) is a commonly used data preprocessing technique to cope with vast amount of data volumes. This chapter presents two types of DR techniques: DR by Transform (DRT) and DR by band selection (DRBS). Two transformed-based techniques are developed for DR. One is statistics-based component transforms that represent image data by a set of component images that are statistically uncorrelated in terms of the used criterion. These include some popular and well-known PCA, MNF, and ICA transforms. The other is feature transforms that represent image data in a space specified by a set of feature vectors that can be obtained by a feature extraction-based criterion. Two such feature transforms are of interest: Fisher's linear discriminant analysis (FLDA) and orthogonal subspace projection (OSP). These DR transforms will be used as an initial pre-processing step in endmember extraction in PART II and exploitation-based hyperspectral information compression in PART V where a new concept of dimensionality prioritization (DP) extended from DRT will be introduced in Chapter 20 to perform progressive spectral dimensionality process (PSDP). In parallel to the development of DRT, a similar treatment can be also carried out for DRBS and will be discussed in detail in Chapter 21. Specifically, a concept similar to DP, referred to as band prioritization (BP), will be introduced in Chapter 21 as a counterpart of DP to perform progressive band dimensionality process ...