Underwater Acoustic Target Recognition Based on Multi-Scale Features Fusion

Underwater acoustic target recognition is critical for ocean exploration, marine monitoring, and defense applications. However, its performance is severely constrained by complex noise, channel fading, and, most notably, the scarcity of labeled data. Traditional methods rely on handcrafted features with strong physical interpretability but often fail in complex environments. In contrast, deep learning approaches, such as convolutional neural networks (CNNs), can effectively extract discriminative representations from time-frequency spectrograms, yet they typically require large-scale labeled datasets and lack physical interpretability. To address these challenges, this paper proposes an Underwater Acoustic Multi-scale Feature Fusion Network (UMFNet). First, a multi-scale feature set is constructed by integrating eight types of physically interpretable one-dimensional (1D) acoustic features-such as pitch, loudness, periodicity, and timbre-with two-dimensional (2D) time-frequency representations, embedding domain knowledge into the model. Second, instead of direct feature concatenation, the intrinsic relationships among 1D acoustic features are modeled via a semantic graph topology to form structured representations, while 2D features are extracted independently to preserve their spatial characteristics. These heterogeneous features are then effectively fused for robust representation learning. Finally, experiments conducted on the limitedsample Watkins Marine Mammal Sound Database demonstrate that UMFNet outperforms existing methods, highlighting the effectiveness of the proposed multi-scale heterogeneous fusion strategy.