Dimensionality Reduction: State of the Art

The narrow and contiguous bands of hyperspectral imagery (HSI) contain redundant information that increases computational complexity. Additionally, all the bands are not always required for a specific application. Hence, to improve the efficiency of hyperspectral data analysis, dimensionality reduction is carried out as an important pre-processing step where information-rich bands are retained for further analysis discarding the redundant ones. The two major ways of dimensionality reduction include feature selection and feature extraction. The feature extraction methods transform the original hyperspectral data to a new subspace such that the majority of the information are embedded into a few top-ranked features that are used in subsequent analyses. The physical significance of a band is lost in this process. Alternatively, the band selection methods select a subset of bands having the most discriminating characteristics, and therefore, the physical relevance of the selected bands is maintained. This chapter discusses the state-of-the-art methods of dimensionality reduction of HSI. Specific gap areas are also analysed, and accordingly, improved methodologies are given in subsequent chapters.