Batch-Mode Active-Learning Methods for the Interactive Classification of Remote Sensing Images

摘要: This paper investigates different batch-mode active-learning (AL) techniques for the classification of remote sensing (RS) images with support vector machines. is done by generalizing to multiclass problem defined binary classifiers. The investigated exploit query functions, which are based on evaluation two criteria: uncertainty and diversity. criterion associated confidence supervised algorithm in correctly classifying considered sample, while diversity aims at selecting a set unlabeled samples that as more diverse (distant one another) possible, thus reducing redundancy among selected samples. combination criteria results selection potentially most informative each iteration AL process. Moreover, we propose novel function kernel-clustering technique assessing new strategy representative sample from cluster. proposed theoretically experimentally compared state-of-the-art methods adopted RS applications. accomplished considering very high resolution multispectral hyperspectral images. By this comparison, observed method resulted better accuracy respect other state-of-the art both data sets. Furthermore, derived some guidelines design systems types