Accelerating MCR-ALS decomposition of hyperspectral images using k-means clustering

Abstract: A hyperspectral image may be decomposed into component spectra and their distri- bution in the image to simplify analysis by revealing underlying patterns and reducing the dimensionality of the image; this may be achieved by the algorithm MCR-ALS. However, the algorithm is time consuming, but could be accelerated by a data re- duction. Data reduction can be done by using a clustering method. In this project, the aim is to determine how clustering, k-means in particular, can be incorporated with MCR-ALS to achieve an accelerated decomposition. We measured how different losses and time consumption were influenced by different parameter choices, e.g, initialization of the k-means. Clustering can result in a re- duction in the time-consumption independently of the choice of parameters, but the choices altered the decomposition substantially. From the results, we concluded that k-means can be incorporated into MCR-ALS, and that the method for selection of centroids is the most crucial step. Accordingly, an optimal set of parameters could be determined.