Cross-Domain PolSAR Image Classification Using Complex-Valued Few-Shot Learning Network

With the impetus of supervised deep learning-based classifiers requiring extensive human-annotated training sets, polarimetric synthetic aperture (PolSAR) image classification has advanced remarkably. However, obtaining dependable labeled samples is labor-intensive and time-consuming. Furthermore, due to actual domain disparities, noticeable performance degradation often occurs when applying a trained classifier to unseen domains. Therefore, this article proposes a complex-valued cross-domain few-shot learning classification (CCFSLC) method for PolSAR images. The goal is to integrate domain adaptation with few-shot learning under the complex-valued network framework to remedy the aforementioned problems. For the proposed CCFSLC, based on the source domain with adequate label information, the created transferable knowledge learning module is first trained to learn an effective complex-valued feature encoder (CVFE) for extracting discriminative transferable knowledge. Subsequently, the deep few-shot learning module, constructed using the pretrained CVFE, undergoes training episodes in both source and target domains to learn reliable domain-invariant features, utilizing just minimal target labeled samples. Meanwhile, the adversarial domain adaptation module is employed to eliminate domain shift, thereby further enhancing cross-domain classification accuracy. The proposed CCFSLC mainly focuses on reducing the domain gap to recognize new categories in unseen domains with only a few annotated samples, while exploring more comprehensive and abundant discriminative information without compromising the integrity of the raw PolSAR data. Comprehensive experimental results on typical datasets demonstrate the superiority of CCFSLC over state-of-the-art methods for cross-domain PolSAR image classification.