Acoustic target recognition of large yellow croaker via small-sample deep learning with a proposed dataset partitioning strategy.

Research on acoustic target recognition of the large yellow croaker (Larimichthys crocea) holds significant implications for precise monitoring of their population status and maintaining marine ecological balance. Due to the complex acoustic fields in underwater aquaculture environments and the high time cost of data collection, it is often challenging to gather sufficient effective samples for target recognition tasks. To address the practical challenges in small-sample scenarios-such as limited effective acoustic data, high recognition difficulty, low accuracy, and sensitivity to anomalous samples-this study systematically integrates and adapts a set of established techniques, including data partitioning and ensemble learning, specifically tailored for the acoustic characteristics of the large yellow croaker in aquaculture environments. The proposed approach employs a pre-grouping strategy on the training dataset and incorporates a loss-based weighting mechanism to adjust sub-model contributions, with further optimization focused on efficient data partitioning under small-sample conditions. Experimental results on a dedicated small-sample acoustic dataset of the large yellow croaker demonstrate that the method achieves a recognition F1-score of 87.8% and helps mitigate feature-learning imbalances, indicating its practical effectiveness for this specific application.