MSSInvNet: A Multi-component Surface-wave Spectrogram Inversion Network

Summary In this study, we present MSSInvNet, a convolutional neural network (CNN)-based method for multi-component surface-wave inversion, designed to directly predict S-wave velocity (Vs) models from dispersion spectrograms. Unlike traditional approaches that require the extraction of dispersion curves for inversion, MSSInvNet employs an end-to-end framework to map multi-component spectrograms directly to Vs profiles. A dataset comprising 3,000 synthetic models, including horizontal (X) and vertical (Z) components, was generated for training and testing, with random noise added to enhance robustness. The model was optimized using the root mean squared error (RMSE) loss function in conjunction with data augmentation techniques. The results demonstrate that MSSInvNet achieves high accuracy, with RMSE values below 10 m/s, and surpasses single-component methods in both precision and noise resilience. An error distribution analysis indicates the absence of systematic bias. By effectively leveraging multi-component data, MSSInvNet significantly improves the efficiency and accuracy of subsurface structure reconstruction, providing a novel and practical solution for surface-wave inversion applications.