Self-Supervised Continual Learning for SAR-ATR: A Local Feature Adaptation Framework

Label issue in synthetic aperture radar automatic target recognition (SAR ATR) stems from two sources: 1) radar imagery requires domain experts for slow annotation; and 2) the privacy-sensitive nature of target types exists. When most continually acquired data remain unlabeled and labels are delayed/scarce, the update and improvement of SAR ATR is hindered, and the system's timeliness in dynamic task environments decreases. In this article, a self-supervised continual feature adaptation (SscFA) framework is proposed to separate continual representation learning from supervised classifier updating. A self-supervised channel continually refines base representations from the unlabeled stream using Barlow-Twins contrastive loss regularized by memory-aware synapses. When only a handful of labels become available, a supervised adapter rapidly specializes base representations through elementwise multiplication while probabilistically freezing base blocks, which could mitigate negative feature transfer and catastrophic forgetting. Experiments on three popular datasets under label-scarce class-incremental learning settings show that SscFA outperforms baselines in feature transferability and stability in SAR ATR.