| Remote sensing instruments have been traditionally multispectral, with observations collected at a few spectral bands. However, hyperspectral instruments capable of collecting observations at hundreds of bands are becoming very popular. Furthermore, research and development efforts continue towards ultraspectral instruments that can produce observations at thousands of spectral bands. Although these remote sensing technology developments hold a great promise for new findings in Earth and space science, they also present many challenges. For example, conventional methods for land use/land cover classifications may not be applicable, due to the large volumes of the hyperspectral data cubes. Therefore, such conventional methods require a preprocessing step, namely dimension reduction. Dimension Reduction is a spectral transformation, aimed at concentrating the vital information and discarding redundant data in order to avoid what has been named the Hughes phenomenon. The Hughes phenomenon refers to the loss of classifier performance with increased data dimensionality. The dimension reduction process, however, is computationally intensive. This is specially true with the large amounts of data produced by hyperspectral instruments. Thus, efficient sequential and parallel dimension reduction techniques were pursued in this work.; In this work, we proposed and investigated efficient dimension reduction algorithms, both sequential and parallel, based on three techniques: (1) Adaptive Principal Component Analysis, (2) Wavelet Reduction Analysis, and (3) A hybrid approach that attempts to combine the desirable features of both wavelet and PCA methods by trading the weaknesses of one with the strengths of the other. |
