Semblance-based velocity picking using unsupervised machine learning

Ninoska Amundaray, Daniel O. Trad

Normal moveout (NMO) correction depends on the identification of optimal velocity-time pairs to flatten the hyperbolic character associated with seismic reflections. Semblance panels are ideal attributes to accomplish this task. However, they are highly affected by the level of noise in the data and poor calculations for short offsets. These are often overcome with additional seismic attributes, which we suggest to replace with the introduction of velocity trends based on semblance. In this study, we demonstrate that our method works as an adequate filtering technique, capable to generate inputs for applications of unsupervised machine learning (ML) in velocity analysis. The performance of three different types of clustering methods known as K-Means, Gaussian Mixture Models and DBSCAN, is investigated by the identification of velocity-time pairs to guide the NMO correction in two datasets simulated for the Marmousi model. In both tests, deep reflectors are corrected independently of the clustering technique used; whereas shallow events are only flattened at near and mid offsets, and under corrected or stretch at far offsets.