Dimensionality reduction and classification for hyperspectral remote sensing data using ISOMAP

In order to address intrinsic nonlinearities of hyperspectral remote sensing data, isometric feature mapping (ISOMAP) is the most widely utilized global manifold learning approach for nonlinear dimensionality reduction. In this paper, it was employed to extract the inherent manifold of hyperspectral data and the experimental results show that ISOMAP provides a significantly more compact feature representation of hyperspectral data than the minimum noise fraction (MNF) transform. Considering the spectral information of hyperspectral data, spectral angle (SA) was applied to derive the neighborhood distances in ISOMAP algorithm, and the result was better. Extracted subspace features via ISOMAP algorithm were also implemented in conjunction with k Nearest Neighbor (kNN) classifier for classification. Experimental results show ISOMAP achieves higher classification accuracies than MNF transform, but with much smaller dimensionality. Especially, ISOMAP provides better discrimination for spectrally similar classes.