Adaptive semi-supervised learning framework for synthetic aperture radar image classification

Synthetic aperture radar (SAR) images can capture data under all weather conditions and at any time, making them particularly effective for detecting sea surface targets, which are often influenced by the complex and dynamic marine environment. To address the challenges of dataset scarcity and class imbalance in ship target classification using SAR imagery, we introduce a semi-supervised learning framework. We propose an adaptive teacher–student with reliable bank network (adaptive-TSRBNet), designed to leverage a substantial amount of unlabeled data to enhance the model’s robustness in underrepresented classes. This framework consists of three main steps: (1) using a semi-supervised teacher–student network to generate pseudo-labels from the unlabeled dataset; (2) incorporating a dynamic reliable bank to curate a trustworthy repository of pseudo-labels; and (3) introducing a novel loss function that integrates both labeled and unlabeled data to facilitate feature learning within the student network. The FUSAR-Ship dataset used in this study contains SAR images of various vessel types and undergoes processing and augmentation to mitigate class imbalance. Data augmentation, achieved through random transformations such as rotation, flipping, and noise injection, improves model robustness and generalization while preserving class distribution. Experimental results indicate that in highly imbalanced classification scenarios, adaptive-TSRBNet achieves a minimum average accuracy improvement of 2.2% over state-of-the-art methods, with certain categories exhibiting performance gains exceeding 20%, demonstrating its superior capability in learning from minority classes. Furthermore, in small-sample scenarios with three and six categories, classification accuracy improves by 6% and 8%, respectively, highlighting the model’s superiority in long-tail categories.