Low-dimensional Representations of Hyperspectral Data for Use in CRF-based Classification

Probabilistic graphical models have strong potential for use in hyperspectral image classification. One important class of probabilisitic graphical models is the Conditional Random Field (CRF), which has distinct advantages over traditional Markov Random Fields (MRF), including: no independence assumption is made over the observation, and local and pairwise potential features can be defined with flexibility. Conventional methods for hyperspectral image classification utilize all spectral bands and assign the corresponding raw intensity values into the feature functions in CRFs. These methods, however, require significant computational efforts and yield an ambiguous summary from the data. To mitigate these problems, we propose a novel processing method for hyperspectral image classification by incorporating a lower dimensional representation into the CRFs. In this paper, we use representations based on three types of graph-based dimensionality reduction algorithms: Laplacian Eigemaps (LE), Spatial-Spectral Schroedinger Eigenmaps (SSSE), and Local Linear Embedding (LLE), and we investigate the impact of choice of representation on the subsequent CRF-based classifications.