A Dual-Branch Deep Learning Approach for Passive Sonar Underwater Target Classification

Underwater target recognition faces significant challenges due to complex noise interference, which results in signal attenuation and distortion during propagation. These factors lead to blurred features and increased intraclass variance, ultimately degrading recognition performance. To overcome these challenges, the proposed model incorporates the GAM attention mechanism, the Kolmogorov–Arnold network (KAN), and a dual-channel structure, where each component is designed to mitigate noise interference, capture nonlinear feature relationships, and exploit multifeature complementarity, respectively. The model processes various types of underwater acoustic features through two independent network channels and fuses the outputs from both channels for final classification. This architecture fully exploits the complementarity of different features, thereby enhancing the model’s target recognition capability in low signal-to-noise ratio (SNR) environments. To validate the model’s effectiveness, multiple experiments were conducted, and its performance was compared with the existing methods under low SNR conditions. The experimental results demonstrate that the proposed dual-channel model achieves superior recognition accuracy.