A noise-masking marine mammal sound classification method via multi-task learning

Underwater noise pollution hinders passive acoustic monitoring (PAM) of marine mammals, as ambient noise masks target signals and degrades classification performance. To address this, we propose MT-MaskNet, an end-to-end multi-task learning (MTL) framework based on ResNet18. This framework jointly optimizes marine mammal sound classification and noise-type classification as the primary and auxiliary tasks, respectively. Guided by the auxiliary branch, a mid-level gating mechanism dynamically suppresses noise-related activations while preserving salient acoustic patterns. We adopt a two-stage training strategy to decouple sound and noise representations before fine-tuning the gating module. This approach mitigates negative transfer and promotes positive transfer to the primary task. On a synthesized dataset featuring three dolphin species with realistic noise superposition, MT-MaskNet achieved a mean accuracy of 95.00% ± 0.8%. This performance significantly outperformed single-task baselines (p = 0.0018). Evaluations on real-world PAM recordings further demonstrated a marked accuracy improvement from 44.00% to 73.75%. Overall, the gating mechanism enables robust feature enhancement through coarse-grained noise indication with minimal computational overhead. • Proposes MT-MaskNet, an end-to-end multi-task learning framework for noise-robust marine mammal sound classification. • Introduces a mid-level gating mechanism, guided by an auxiliary noise-type classification branch, for adaptive feature-level noise suppression. • Employs a two-stage training strategy to decouple sound and noise representations, mitigating negative transfer while promoting positive transfer to the primary task. • Achieves superior performance, significantly outperforming single-task baselines on both synthetic and real-world Passive Acoustic Monitoring (PAM) datasets. • Demonstrates that the gating mechanism primarily provides coarse-grained noise indication, enabling robust feature enhancement with only a modest parameter overhead.