MSSInvNet: a multi-component surface-wave dispersion spectrogram inversion network

The estimation of near-surface shear-wave ( V s ) velocity models plays a critical role in environmental and engineering geophysics. Traditional surface-wave methods usually rely on the manual extraction of dispersion curves from spectrograms, which is time-consuming and prone to error in the manual recognition and picking of multi-modal surface waves. An alternative way to overcome these difficulties is to invert the dispersion spectra instead of the dispersion curve. Conventional dispersion-spectral inversion approaches utilize vertical-component spectra only, thereby neglecting valuable information contained in the other seismic components. In this study, we propose a multi-component surface-wave spectrogram inversion network (MSSInvNet), a deep-learning approach that simultaneously inverts both horizontal- and vertical-component dispersion spectra to estimate V s models. Synthetic datasets are generated through finite-difference simulations of the wave equation and are used for training, testing, and validation of MSSInvNet. Experimental results demonstrate that MSSInvNet outperforms single-component inversion networks, yielding relatively higher accuracy. In addition, we analyze how dataset size affects model performance and show that selecting an optimal dataset size helps balance accuracy and training efficiency of the deep-learning network. Field data example shows that MSSInvNet exhibits strong robustness against noise and is capable of the rapid and accurate estimation of reliable V s profiles. These results show that MSSInvNet offers a promising and efficient approach for characterizing near-surface V s structures.