Generalized Zero-Shot Domain Adaptation for Unsupervised Cross-Domain PolSAR Image Classification

Abstract

Cross-domain polarimetric synthetic aperture radar interpretation is urgently needed, due to the rapid data growth and label scarcity. However, the class distribution shift problems hinder the reuse of labeled samples among cross-domain images. Most of the existing domain adaptations can only handle the cross-domain case of same categories between source and target domains, while the categories of the target domain are usually more abundant than those of the source domain. To improve the usability of labeled samples among cross-domain images, an unsupervised generalized zero-shot domain adaptation (uGZSDA) based on scattering component semantics (SCSs) is proposed. By using SCSs and limited labeled samples (seen categories) in the source domain, more land cover types (seen and unseen categories) in the unlabeled target domain can be inferred. First, a stacked autoencoder (SAE) extracts source/target-domain features, and SCSs of typical land covers are constructed by cross-domain databases and statistical scattering components. Second, combining SAE features and source-domain samples, the most likely seen class samples in the target domain are selected by probability sorting, and the SAE is retrained by obtained selected seen samples. Third, the unseen class samples in the target domain are inferred by the retrained SAE, classification probability, and semantic similarity. Finally, the selected seen and inferred unseen class samples in the target domain are used to further retrain the SAE, and the target domain is classified by the retrained SAE and the classifier. The proposed uGZSDA is verified among 16 cross-domain PolSAR datasets. Using SCS and two to three types of seen samples from the source domain, the accuracies of seven types of land covers in the unlabeled target domain can reach 76–83.96%.