STUDY OF FEW DIMENSIONALITY REDUCTION ALGORITHMS FOR HYPERSPECTRAL DATA

Abstract

Recent advances in sensor technology have led to the development of hyperspectral sensors capable of collecting remote sensing imagery at several hundreds of narrow spectral bands over the spectrum. While these developments hold great promise for earth observation, they create new processing challenges. Therefore, processing hyperspectral data using new efficient techniques is a must. One such approach for effectively processing hyperspectral data is through dimension reduction prior to the use of data in applications such as land use/land cover classifications. Hyperspectral sensors provide high-resolution spatial and spectrally rich information. Though hyperspectral data is highly nonlinear, dimensionality reduction is usually carried out using simple linear techniques like PCA. As the spatial and spectral resolution hyperspectral sensors keep on increasing every year, nonlinearity of information content will keep on increasing and linear techniques will become more inappropriate to process hyperspectral data. So manifold learning dimensionality reduction algorithms which harness the inherent nonlinearity of hyperspectral data are needed in the near future. Non-linear dimensionality reduction algorithms are applied in various machine learning domains like automatic pose recognition from a set of face images etc but their applicability in hyperspectral remote sensing is less explored mainly because they are not yet implemented in popular remote sensing software. In this research, various nonlinear dimensionality reduction algorithms, namely Isomap, LLE, KPCA, LTSA and Laplacian Eigenmaps are studied and their performance evaluated based on runtime and classification accuracy on hyperspectral data. Keywords: Hyperspectral remote sensing, intrinsic dimensionality, nonlinear dimensionality reduction, manifold learning, Isomap, LLE, KPCA, Laplacian Eigenmaps, LTSA.