Low-Shot Learning for Synthetic Aperture Sonar Image Classification Using Hierarchical Pretraining & AirSAS

Deep learning for undersea remote sensing typically demands extensive labeled data to obtain good performance, a significant challenge in synthetic aperture sonar (SAS) image classification due to the difficulty of obtaining diverse, welllabeled datasets. To mitigate this problem, we propose a low-shot learning approach for SAS (defined as fewer than 100 target exemplars per class), leveraging Hierarchical Pretraining (HPT), a recent self-supervised learning (SSL) method, and unlabeled imagery from AirSAS, a benchtop system that rapidly generates cost-effective, SAS-like data. This work makes two primary contributions: first, we demonstrate that applying HPT improves low-shot learning performance for SAS classification by leveraging unlabeled, out-of-domain imagery during pretraining; second, we show that HPT effectively utilizes unlabeled data from a terrestrial benchtop system (AirSAS) to enhance low-shot learning performance on underwater SAS imagery, establishing a new paradigm for rapid automatic target recognition (ATR) development. Our method achieves competitive low-shot performance on a challenging real-world SAS dataset, demonstrating superior performance over ImageNet-only pretraining across various undersea environments and performance metrics.