Active Learning for Domain Adaptation in Classification of Remote Sensing Data by Minimizing Expected Error with Diversity Maximization

摘要

This paper investigates a method addressing remote sensing image classification based on the domain adaptation (DA) and active learning (AL). The key idea of our method is to retrain the classifier by using the available labeled samples from source domain, and adding minimum number of the most informative samples with active queries in the target domain. The active learning approaches provide pool of candidate samples based on different query functions with the assessment of uncertainty and diversity. The uncertainty criterion evaluates the algorithm' confidence in order to correctly classify the considered samples, while the diversity criterion decreasing the redundancy between the selected samples by choosing a set of unlabeled samples with more diversity. So, based on this model, we exploit both two criteria in order to choose the most informative samples are selected iteratively at the active learning procedure. In this work, we have proposed a novel uncertainty criteria based minimum expected error (MEE), and also use angular based diversity (ABD) as diversity criteria. The experimental outcomes prove the superiority of the proposed approach to counterpart methods.