Accelerating support vector machines for remote platforms by increasing sparsity

Abstract

The support vector machine (SVM) classification algorithm often achieves quite high accuracy on hyperspectral images, even when trained on small amounts of data. However, SVMs can still be computationally expensive relative to the desired throughput and available resources on remote imaging platforms.In this paper, the possibility of decreasing the computational costs of SVMs by increasing their sparsity is explored on a few simple hyperspectral scenes. The number of bands is reduced by a factor of up to 20, which roughly corresponds with a