Domain Adaptation for Multilabel Remote Sensing Image Annotation With Contrastive Pseudo-Label Generation

Abstract

Deep-learning-based multilabel remote sensing image annotation (MLRSIA) is receiving increasing attention in recent years. MLRSIA needs a large volume of labeled samples for effective training of the deep models. However, the scarcity of labeled samples is a common challenge in this field. Domain adaptation (DA), aiming to transfer knowledge from label-rich datasets (source domains) to label-scarce datasets (target domains), has become an effective means to address this problem of limited labeled samples. But most of the existing DA models are primarily designed for single-label annotation tasks, leaving the application of DA to multilabel annotation tasks as an open issue. In this article, a DA method for MLRSIA, named contrastive pseudo-label generation (CPLG), is proposed. CPLG mainly consists of two parts: generating and selecting pseudo-labels for the samples in the target domain, and enhancing the cross-domain feature consistency through contrastive learning. Specifically, the soft predictions (or posterior probabilities) and the corresponding pseudo-labels of the target samples are first generated using neighborhood aggregation. Then, a positive and negative pseudo-label selection strategy is designed to refine these pseudo-label. Finally, a contrastive loss is introduced to align the similar sample features between the source and target domains to avoid the pseudo-labels of the target samples being overly biased toward the source domain, further improving the precision of these pseudo-labels. The MLRSIA experiments, conducted across four different DA scenarios on three benchmark datasets, demonstrate the advantages of the proposed CPLG compared to other state-of-the-art methods.