Comparison of fully constrained and weakly constrained unmixing through mine-tailing composition mapping

Abstract

In recent years, remote-sensing scientists have been developing applications for hyperspectral remote sensing. Analysis techniques, such as linear spectral unmixing, have been used increasingly to solve real-world problems related to vegetation stress detection, mineral prospecting, and environmental monitoring. Information products such as end member fraction maps can be generated by interpreting unmixing results. However, the validity of these maps has not been fully examined. In this paper, two types of constrained linear spectral unmixing techniques are investigated, namely fully constrained and weakly constrained unmixing. Overall, this study revealed that the weakly constrained approach provides more realistic results than the fully constrained technique. From a data-analysis perspective, this study demonstrates that the weakly constrained unmixing can be used successfully when an incomplete list of end members is used in the unmixing, which is almost always the case in reality. In addition, this paper also addresses the scaling issue. When using laboratory spectra as end members for unmixing, proper scaling of the spectra amplitude will reduce errors and, hence, improve unmixing results. The tests were conducted in a mine-tailing site, but the findings derived from this study also apply to other applications.