A new committee-based active learning (CBAL) approach to hyperspectral remote sensing data classification

Abstract

This letter presents a new committee-based active learning (AL) model that queries the contention samples between spectral–spatial-description-based classifier-patch-based support vector machine (PTSVM) and spectral-description-based classifier-pixel-based support vector machine (PXSVM) for classification of hyperspectral remote sensing images. The proposed model consists of three main steps. Firstly, a given image is partitioned into overlapping patches. Then, a PTSVM classifier and a PXSVM classifier are trained using the initial patch and corresponding pixel training sets, respectively. Secondly, a set of unlabelled pixels from pixel candidate pool whose prediction labels disagree by two classifiers are added to the contention pool (CTP). Lastly, a margin sampling (MS)-based AL method is employed to select the most informative pixels from CTP. These pixels are labelled by annotator and added to the pixel training set. At the same time, the patches that contain at least one of these pixels will be added to the patch training set. This process will be repeated until a predefined convergence condition is satisfied. Experimental results show good performance on two hyperspectral data-sets as compared to the state-of-the-art MS and entropy query-by-bagging-based AL models.