A decision fusion approach for clustering of hyperspectral data using spectral unmixing methods

This paper aims at a decision fusion approach for combining three spectral unmixing methods to cluster hyperspectral data. Unlike standard image clustering techniques, analyzing hyperspectral data on a pure pixel basis may not be a true assumption. Meanwhile, multiple classifier systems often show better performance than each of the constituent classifiers. This is due to the fact that each classifier makes errors on different regions of the input space. With these facts in mind, this paper distills these two approaches into a single approach and exploits the advantages of both spectral unmixing algorithms and decision fusion methods. In this paper, three unmixing methods namely, Fully Constrained Least Squares (FCLS), Nonnegatively Constrained Least Squares (NCLS) and Sum-to-one Constrained Least Squares (SCLS) are employed as the ensemble classifiers and their results are combined at two different fusion levels: the abstract level and the measurement level. Experimental results on a real-world hyperspectral data proved that the proposed approach shows better clustering results compared to those of K-Means and Fuzzy c-Means in terms of the Adjusted Random Index (ARI) measure.