Spectral-Angle-Based Discriminant Analysis of Hyperspectral Data for Robustness to Varying Illumination

In this paper, we assert that a recently developed approach to classify hyperspectral images provides substantial robustness to illumination effects for hyperspectral image analysis. The approach [angular discriminant analysis (ADA) and its variants] seeks to find lower dimensional subspaces, where class-specific information is angularly well-separated and where appropriate classifiers can then be applied. We demonstrate the ADA, and its local and kernel variants provide substantial robustness to data which is affected by such illumination differences (commonly encountered in geospatial images). We validated with two different hyperspectral datasets, both of which contain classes in well-illuminated regions and in shadows. We note that the proposed framework is suitable to the general problem of illumination variation in scenes-we validate with a specifically harsh and commonly encountered scenario where classes in an image are present in regions where they are well illuminated, and regions where they are occluded by shadows. Experimental results show that this dimensionality reduction not only enables robustness to sample size by virtue of feature reduction but also is very effective at capturing class-specific information irrespective of changes in illumination content, in the lower dimensional subspace.