Hierarchical Subspace Learning Based Unsupervised Domain Adaptation for Cross-Domain Classification of Remote Sensing Images

We address the problem of automatic updating of land-cover maps by using remote sensing images under the notion of domain adaptation (DA) in this paper. Essentially, unsupervised DA techniques aim at adapting a classifier modeled on the source domain by considering the available ground truth and evaluate the same on a related yet diverse target domain consisting only of test samples. Traditional subspace learning based strategies in this respect inherently assume the existence of a single subspace spanning the data from both the domains. However, such a constraint becomes rigid in many scenarios considering the diversity in the statistical properties of the underlying semantic classes and problem due to data overlapping in the feature space. As a remedy, we propose an automated binary-tree based hierarchical organization of the semantic classes and subsequently introduce the notion of node-specific subspace learning from the learned tree. We validate the method on hyperspectral, medium-resolution, and very high resolution datasets, which exhibits a consistently improved performance in comparison to standard single subspace learning based strategies as well as other representative techniques from the literature.