Multiple Classifier Ensembles with Band Clustering for Hyperspectral Image Classification

Abstract Due to the high dimensionality of a hyperspectral image, classification accuracy of a single classifier may be limited when the size of the training set is small. A divide-and-conquer approach has been proposed, where a classifier is applied to each group of bands and the final output will be the fused result of multiple classifiers. Since the dimensionality in each band group is much lower, classification accuracy of the overall system can be improved even when training samples are limited. In this paper, we proposed a new multiple classifier ensembles which using SKMd-based band clustering features as input. We also investigate the impact of band partition for this approach. We find out that band partition based on spectral clustering (resulting in band groups composed of non-consecutive bands) can outperform the partition based on spectral correlation coefficient (resulting in band groups composed of consecutive bands only), in particular when the number of training samples is small.