Line Spectrum Feature-Guided Deep Convolutional Network for Low-SNR Underwater DOA Estimation

Direction of Arrival (DOA) estimation is a crucial task in underwater target detection. However, both data-driven and model-driven algorithms can be regarded as threshold detection based on the signal energy in the direction of the wave vector. In complex underwater environments characterized by low Signal-to-Noise Ratio (SNR) and limited snapshots, performance significantly degrades. To address this challenge, this paper proposes a line spectrum feature-guided deep convolutional network (LSF-CNN). Unlike traditional deep learning methods that directly input the covariance matrix, our approach utilizes the radiated noise generated by naval ships during forward movement for target recognition, which exhibits certain frequency domain characteristics. Specifically, we design a feature extraction module to identify the main line spectrum components from the Power Spectral Density (PSD) and construct a Line Spectrum Enhanced Covariance Matrix (LSE-CM). This matrix serves as a robust input feature map for a multi-layer Convolutional Neural Network (CNN), which learns spatial features while suppressing broadband noise. Simulation results demonstrate that the proposed LSF-CNN achieves superior estimation accuracy compared to the classic MUSIC algorithm and standard CNNs, particularly under severe noise conditions (e.g., SNR below -10 dB), showing great potential for robust underwater target localization.