A Multi-Time-Frequency Feature Fusion Approach for Marine Mammal Sound Recognition

Accurate acoustic identification of marine mammals is vital for monitoring ocean health and human impacts. Existing methods often struggle with limited single-feature representations or suboptimal fusion of multiple features. This paper proposes an Evaluation-Adaptive Weighted Multi-Head Fusion Network that integrates CQT and STFT features via a dual-branch ResNet architecture. The model enhances intra-branch features using channel attention and adaptive weighting of each branch based on its validation accuracy during training. Experiments on the Watkins Marine Mammal Sound Database show that the proposed method achieves superior performance, reaching 96.05% accuracy and outperforming baseline and attention-based fusion models. This approach offers an effective solution for multi-feature acoustic recognition in complex underwater environments.