Optimizing partial receiver operating characteristic curve via curriculum learning and Neyman–Pearson criterion for robust underwater acoustic target detection

This paper addresses the challenge of underwater acoustic target detection, a critical task in marine monitoring and passive sonar systems, which is often hindered by complex noise environments and imbalanced labeled data where the targets appear very sparse in the long collected data. Traditional models take the minimization of the binary cross-entropy (BCE) as the optimization criterion. However, underwater target detection is fundamentally a class-imbalanced classification problem that uses the receiver operating characteristic curve as the evaluation metric instead of the classification accuracy, while BCE maximizes the classification accuracy on training data. To address this, three optimization methods are proposed to directly maximize the area under the receiver operating characteristic curve (AUC). Additionally, the Neyman-Pearson criterion from classical detection theory is incorporated into the AUC optimization framework, forming a curriculum learning strategy that progressively optimizes the partial area under the curve (pAUC). To overcome the scarcity of underwater data, a cross-domain knowledge transfer method is implemented from the airborne to underwater acoustic domains, which accelerates model convergence and improves generalization. Experimental results demonstrate that the proposed AUC- and pAUC-based loss functions outperform BCE and achieve state-of-the-art performance under low signal-to-noise ratio and mismatched conditions.