Model-Guided Deep Learning for Line Segment Detection in Time–Frequency Spectrograms of an Ocean Waveguide

The application of machine learning in underwater acoustics is often limited by the lack of high-quality data. One method to avoid this data issue is to use modeled data to train a machine learning algorithm, called model-guided learning. In this study, a U-Net-based model-guided deep learning approach was developed to identify dispersion curves in an oceanic waveguide. The U-Net is trained using supervised learning with modeled data generated from an ocean propagation model to detect line segments in a time–frequency spectrogram. The evaluation of U-Net with the test data, based on the performance metrics, such as probability of false alarm, probability of detection, and normalized cross-correlation coefficient, reveals that it effectively extracts the dispersion curves. The proposed network was successfully applied to unseen simulated and experimental data. Our results demonstrate that the dispersion curve images generated through model-guided deep learning can serve as concise image features, including information regarding ocean environments.