Loss-Aware Data Augmentation With Dynamic Trimming and Weighting for Underwater Acoustic Target Classification

Abstract

Limited availability of labeled data presents a significant challenge for underwater acoustic target recognition (UATR), often resulting in model overfitting and poor generalization. Data augmentation (DA) has been a major strategy to increase effective data diversity, yet prevailing methods often lack explicit mechanisms to discriminate the informational value of augmented samples. This letter presents two DA approaches, Loss-Aware Trimming Augmentation (LATA) and Learnable Weight-Based Augmentation (LWBA), to enhance the UATR task under restricted annotated data scenarios. LATA adaptively prunes both excessively difficult and trivial augmented samples based on real-time loss evaluation, while LWBA introduces sample-wise learnable weights to balance the influence of each augmentation during model training. Experiments conducted on the public DeepShip dataset validate the superiority of the proposed framework, with an average improvement of 3.44% in accuracy and 3.49% in F1-score compared to the baselines.