Underwater Acoustic Target Classification Using Hybrid Temporal–Spectral Feature Learning Network

Abstract

Underwater acoustic target classification (UATC) seeks to identify unknown acoustic sources through passive sonar in oceanic remote sensing applications. However, the highly dynamic marine environment and various background noise pose significant challenges to improving UATC performance. To address these challenges, we develop a hybrid temporal–spectral feature learning neural network that integrates three core components: a Fourier analysis network (FAN)-based temporal feature learner, a mixture of expert networks (MoEN)-based spectral feature learner, and an adaptive feature fusion (AFF) module. Specifically, the FAN-based learner encodes temporal feature by capturing multiple periodic patterns associated with underwater acoustic targets. In parallel, the MoEN-based learner models spectral dependencies and emphasizes spectral-selective patterns. The AFF dynamically then balance the contributions of the dual-branch learners through an adaptive weighting mechanism. The proposed algorithm achieves 98.58% accuracy on ShipsEar dataset, outperforming six state-of-the-art methods with moderate model parameters and computational cost.