Using UMAP for Dimensionality Reduction of Hyperspectral Data

Uniform Manifold Approximation and Projection (UMAP) is a recent technique for dimensionality reduction. In this paper, we study UMAP for hyperspectral data processing. For well-known hyperspectral images, we search for optimal parameters and compare UMAP to other dimensionality reduction techniques from the viewpoint of the classification quality. The list of alternative techniques includes the linear Principal Component Analysis technique as well as nonlinear methods, namely, Isomap, Locally Linear Embedding, Laplacian Eigenmaps, and Nonlinear Mapping.