Near Surface Characterization in Southern Oman : Multi-Wave Inversion by Machine Learning

Abstract

Shallow stratigraphy in Southern Oman is characterized by the presence of an anhydrite layer (RUS formation) causing a strong velocity inversion which makes seismic imaging particularly difficult. This known shallow sharp velocity inversion cannot be easily captured with methods relying on reflection or diving wave energy. We propose here to use multi-wave inversion using first breaks and dispersion curves of surface waves to provide near-surface high resolution velocity models in the shallow range depths (0-400m). The success of Multi-Wave Inversion strongly depends on the reliability of the surface wave velocity picking, which could be much more challenging compared to the conventional first break picking. Heavy preconditioning is often the solution to increase dispersion curves quality and to obtain a narrower velocity corridor. To improve reliability, we use K-means clustering, an unsupervised machine learning method in order to filter out the outliers as well as to define geologically dependent corridors. The unsupervised machine learning clustering helps to define more stable dispersion curves picking corridors for different areas, in order to extract better quality surface wave dispersion curves's, especially at low frequencies where their quality is low. The multi-wave inversion, fed with the optimized phase velocity picks, captures the shallow velocity inversion, which is impossible to recover with either first break tomography only or diving wave full waveform inversion only. The combination of two recently developed technologies allows us to characterize accurately the near surface for the first time in the South of Oman. The velocity inversion caused by the RUS formation is well captured and the velocity trend of the updated model follows correctly the checkshot trend down to 500m, confirming the reliability of the dispersion curves picks at a very low frequency. By incorporating this shallow inversion layer into the velocity model, the resulting seismic image is significantly improved and more interpretable. Geological features such as faults appear clearly and seismic layering in the tilted blocks is significantly improved with the multi-wave Inversion machine learning-guided workflow.