End-to-end underwater acoustic target classification with frequency separation strategy

Abstract

Most underwater target classification methods require data preprocessing to convert waveforms into spectrograms. Recent advancements have led to the emergence of waveform-based approaches that directly extract discriminative features from raw acoustic signals, eliminating the need for preprocessing. However, waveform-based architectures often exhibit inferior classification performance compared to spectrogram-based methods. This study identifies two key limitations of existing waveform-based approaches: (1) the lack of explicit frequency feature extraction mechanisms, and (2) information degradation due to downsampling. To address these challenges, we propose an end-to-end frequency separation network (FSNet), which includes two innovative components: a frequency separation (FS) module that explicitly captures discriminative frequency characteristics, and a scale-adaptive max pooling (SAMP) layer that preserves critical information during dimensionality reduction. Comprehensive evaluations on ShipsEar and DeepShip datasets demonstrate that our framework achieves competitive accuracy performance (82.91% on ShipsEar and 78.39% on DeepShip) while maintaining exceptional computational efficiency, requiring only 0.49M parameters—over three times fewer than the second-smallest model.