A Biclassifier Network With Intermediate Domain for Unsupervised Domain Adaptation PolSAR Image Classification

Due to significant data distribution differences among polarimetric synthetic aperture radar (PolSAR) images and the expensive and time-consuming nature of data labeling, existing methods are challenged to classify newly acquired unlabeled data. To address these issues, this letter proposes an unsupervised domain adaptation (UDA) network that leverages an intermediate domain-assisted biclassifier. An adversarial UDA network incorporating a biclassifier is introduced as the fundamental structure. Then, by interacting with the semantic information of the features extracted from the source and target domains, the cross-domain feature enhancement module (CFEM) is integrated to improve intraclass cohesion and interclass separation. In addition, to achieve a more stable domain alignment process, an intermediate domain is created through the weighted fusion of features from both the source and target domains, serving as a conduit for cross-domain knowledge transfer. Experimental results conducted on three datasets captured with different systems or regions show that the proposed method achieves an average overall accuracy (OA) improvement of over 1.5% compared with other state-of-the-art methods.