Expressive Spectral Error Visualization for Enhanced Spectral Unmixing

Abstract

A major issue in multispectral data analysis stems from the concept of spectral mixture analysis, i.e. the fact that a pixel does not cover only one material but corresponds to a mixture of materials. Even though many automatic methods for spectral unmixing exist, in many practical applications, domain experts have to verify the result and sometimes have to manually adjust the set of determined materials to achieve proper spectral reconstructions. In this paper, we propose an approach to enhance the very tedious and time-consuming task of manual verification of the unmixing and optional refinement of the materials. Our visual analysis approach comprises different techniques for an expressive spectral error visualization, efficiently guiding the user towards spectra in the dataset which are potentially missing materials. Here, combined views allow comprehensive, local and global error inspections in parallel. We present results of our proposed approach for two domains.