Parallel and distributed implementation on SPARK of a spectral–spatial classifier for hyperspectral images

Abstract

Hyperspectral image classification is an important task for land cover interpretation that gives input to tremendous remote sensing applications. To deal with difficult classification problems, sparse representation has shown its significant potential. However, its execution is computationally complex and its performance limits its application in time-critical scenarii. We aim to improve the performance of an existent classification algorithm based on the sparse representation of extended multiattribute profiles as spectral–spatial features. We propose a parallel and distributed implementation of this algorithm on the computing engine SPARK using the MapReduce model to ensure the scalability of the classifier over available computing resources. Experimental results obtained on real and synthetic hyperspectral datasets show that the proposed approach reveals remarkable acceleration factors while retaining the same classification accuracy with regard to the sequential version.