Superpixel-based spatial-spectral dimension reduction for hyperspectral imagery classification

摘要

Dimension reduction (DR) is a useful preprocessing technology for hyperspectral image (HSI) classification. This paper presents an HSI DR method named superpixel-based spatial-spectral dimension reduction (SSDR), which integrates the spatial and spectral similarity. The HSI is first segmented into non-overlapping superpixels, where pixels belonging to the same superpixel have strong correlations, and should be preserved after DR. We then apply the superpixel-based linear discriminant analysis (SPLDA) method, which learns a superpixel-guided graph to capture the spatial similarity. Pixels from the same label also have strong spectral correlations; thereby, we also construct a label-guided graph to explore the spectral similarity. These two graphs are finally integrated to learn the discriminant projection. The classification results on two widely used HSIs demonstrate the advantage of the proposed algorithms compared to the other state-of-the-art DR methods. (C) 2019 Elsevier B.V. All rights reserved.