Complex-Valued Source-Free Domain Adaptation for PolSAR Image Classification

The discrepancies of sensors, acquisition conditions and terrain class distributions in polarimetric synthetic aperture radar (PolSAR) data has aroused the cross domain PolSAR image classification problem. Domain adaptation, which aims to improve the classification performance of target domain with knowledge from the source domain, is a promising approach to address this issue. However, conventional domain adaptation methods typically assume access to both source and target domain data, which may be infeasible in real-world settings due to data privacy and confidentiality concerns. Therefore, it is crucial to develop source-free domain adaptation PolSAR image classification methods which rely only on target domain data. In this scenario, how to make full use of the label sparse target domain data is a consequent challenge to overcome. To this end, we propose CVSFDA, a complex-valued source-free domain adaptation framework tailored for PolSAR image classification. CVSFDA incorporates a complex-valued multiscale prototypical matching module (CVMP) and a complex-valued relation network (CVRN). Given the pretrained source domain model and the limited labeled samples in target domain, CVMP captures the similarities between samples leveraging the hierarchical spatial information across multiple encoder layers. A similarity convolutional network is further devised in CVRN, to comprehensively model class-specific information in the target domain. Extensive experiments on four benchmark PolSAR datasets demonstrate that CVSFDA achieves superior classification accuracy and generalization ability compared to existing domain adaptation methods.