Understanding spatial-spectral domain interactions in hyperspectral unmixing using exploratory data analysis

This paper presents a visual exploratory analysis of an AVIRIS hyperspectral image to understand the interactions between the spatial and spectral domains in hyperspectral unmixing. We show how the global data cloud may not be convex due to spatial constraints on the distribution of the materials in the scene. Furthermore, we show that by segmenting the data cloud in feature space into piecewise convex segments, we can analyze individual segments and extract endmembers that better capture local structures compared to methods that look at the global cloud. Challenges remain as to how to do the cloud segmentation using machine-based approaches. However, experimental results point to the use of segmentation as a way to address the problem.