Contrastive Learning With Intra-Recording Pairs for Small-Sample Underwater Acoustic Target Recognition

Abstract

Passive acoustic monitoring (PAM) is a pivotal technology in ocean observation and surveillance. Although deep-learning-based passive underwater acoustic target recognition (UATR) has advanced rapidly, supervised approaches still face two challenges. First, labeled passive sonar data required for model training are scarce and expensive to obtain, while large volumes of historical recordings are often underexploited. Second, underwater acoustic signals, such as ship-radiated noise, exhibit substantial variability due to shifts in marine environments and operating conditions. This makes it difficult for models trained with limited labeled samples to remain robust when test recordings are collected under different conditions. To address these issues, we propose an intra-recording contrastive learning (IRCL) framework for small-sample UATR. It samples two distinct clips from the same recording as a positive pair. A branch-adversarial strategy is further introduced to improve representation robustness under recording-condition variability. Experiments under same-recording and different-recording evaluations show that IRCL outperforms supervised training and conventional self-supervised contrastive pretraining, and remains competitive with identity-supervised pretraining. These results demonstrate the effectiveness of intra-recording contrastive pretraining for improving small-sample UATR under recording-condition variability.