Automatic LOFAR Line-Spectrum Extraction with Hybrid Dataset Construction and a Continuity-Aware U-Net

Line-spectrum features in ship-radiated noise are essential for the analysis and recognition of passive sonar targets. Robust automatic extraction from low-frequency analysis and recording (LOFAR) spectrograms remains challenging in underwater acoustic environments owing to strong background fluctuations and interference. Supervised learning-based methods are further constrained by the limited availability of manually annotated data. This study proposes an automatic LOFAR line-spectrum extraction method that combines hybrid dataset construction with a continuity-aware U-Net (CAU-Net). Simulated and measured samples are integrated into a hybrid training dataset. A pseudo-label generation strategy combining two-pass split-window (TPSW) responses with inter-frame continuity constraints incorporates unlabeled measured samples into training. In addition, a temporal continuity modeling module combines multi-range inter-frame context with local frequency information, improving the extraction of weak components with pronounced energy variations. On an independent test set with known line-spectrum references, CAU-Net achieved an F1 score of 0.9635±0.0012 and a line-location accuracy (LLA) of 0.9756±0.0023 over five random seeds. It also maintained the highest F1 and LLA across the tested signal-to-noise ratio (SNR) range. Qualitative results on complete ShipsEar recordings illustrate that CAU-Net provides visually clearer weak narrowband responses while suppressing scattered background and transient-interference responses.