Hyperspectral image classification based on union of subspaces

Characterizing mixed pixels is an important topic in the analysis of hyperspectral data. Recently, a subspace-based technique in a multinomial logistic regression (MLR) framework called MLRsub has been developed to address this issue. MLRsub assumes that the training samples of each class live in a single low-dimensional subspace. However, having in mind that materials in a given class tend to appear in groups and the (possible) presence on nonlinear mixing phenomena, a more powerfull model is a union of subspaces. This paper presents a new approach based on union of subspaces for hyperspectral images. The proposed method integrates subspace clustering with MLR method for supervised classification. Our experimental results with an urban hyperspectral image collected by the NSF-funded Center for Airborne Laser Mapping (NCALM) over the University of Houston campus indicate that the proposed method exhibits state-of-the-art classification performance.