Automated Extraction of Rayleigh Wave Dispersion Curves from DAS Railway Monitoring

Distributed Acoustic Sensing (DAS) enables continuous assessment of transport infrastructure conditions through passive surface wave analysis. However, operational deployment may be problematic as the dispersion curves from thousands of frequency-velocity dispersion images are generated during routine monitoring. Passive DAS recordings from train-induced vibrations exhibit severe fragmentation, strong higher-mode interference, and temporal variability driven by seasonal moisture changes that consistently challenge manual or semi-automated picking methods. We present a training-free hybrid algorithm that combines marker-controlled watershed segmentation with physics-informed trajectory optimization to extract fundamental mode of dispersion curves from complex operational DAS data. Applied to a 15-month monitoring campaign on a 350 m railway embankment in the UK, the methodology operates directly on dispersion images exported from standard MASW processing software. The algorithm proceeds through five stages: (1) binary masking with morphological noise suppression establishes candidate energy regions; (2) watershed transformation with internal markers separates touching fragments that should constitute independent segments; (3) bidirectional amplitude-maximum propagation with adaptive vertical search radii extracts local trajectory estimates within each isolated fragment; (4) velocity band filtering combined with forward monotonic chaining reassembles disconnected segments by enforcing kinematic consistency and rejecting physically-implausible connections; and (5) global sigmoid fitting with constrained horizontal extension produces smooth, inversion-ready dispersion curves validated against aliasing boundaries. Validation against manual picking demonstrates that the algorithm bridges spectral gaps exceeding 5 Hz, correctly isolates fundamental from higher modes even when energy amplitudes are comparable, and maintains trajectory continuity through severe fragmentation where conventional peak-following methods fail. Beyond immediate operational utility, automated extraction from real-world DAS railway data enables generating computationally labeled training datasets that preserve physical consistency and interpretability. We demonstrate how this computer vision approach produces high-quality dispersion curve labels across diverse geological settings and complexity levels.