Spatial-spectral local discriminant projection for dimensionality reduction of hyperspectral image

Abstract

Dimensionality reduction (DR) technology plays an important role in hyperspectral image (HSI) classification. However, many existing DR algorithms ignore the complex intrinsic structure in spatial domain and spectral domain of HSI. To address this issue, we put forward a spatial-spectral local discriminant projection (SSLDP) method based on the manifold learning theory and spatial consistency in HSI. In SSLDP, hyperspectral pixels are reconstructed by minimizing the weighted reconstruction errors to preserve the local geometric structure. Then, two weighted scatter matrices are designed to maintain the neighborhood structure in spatial domain and two reconstruction graphs are constructed to discover the local discriminant relationship in spectral domain. Finally, an objective function is designed for obtaining an optimal projection by compacting the spatial-spectral local intraclass points while separating the spatial-spectral local interclass points. The experiments performed on some real hyperspectral images, including the Indian Pines, PaviaU and Washington DC, demonstrate that the presented SSLDP algorithm is significantly superior to some state-of-the-art DR algorithms.