Velocity model building of VSP data through Semblance analysis

Vertical seismic profiling (VSP) plays an important role during the exploration/production phase in both mining and oil and gas industries. The main application of VSP is to establish a time-depth-tie that is used for converting surface reflection seismic images into depth domain and to correlate it with other downhole geophysics logs including gamma, neutron, sonic and televiewer images. The quality of VSP velocities is highly dependent on the accuracy of the first arrivals picked within VSP records. A VSP velocity model can be built via manual and automatic picking of the first arrivals. Manual picking involves visual observation of the first break inflection point by a specialist, which is a labour-intensive procedure. Reliability of such picks can be subjective, specifically in the presence of significant noise. Hence, it is important to use more objective velocity analysis methods that combine manual picking with automated workflows. Within the past decades, a variety of automated first break picking solutions have been developed including cross-correlation of adjacent traces, short-term average to long-term average ratio, and machine learning. Limitations of these techniques include sensitivity to noise, sparsity of traces, and requirement for large training datasets. Moreover, these methods are typically locked into commercial software that are not available in public domain. We propose to use an automatic VSP velocity analysis workflow using semblance function, which is a powerful tool to quantify coherency of seismic wavefronts. During the past decades, seismic velocity model building using semblance, has become a routine stage in most surface seismic data processing workflows. Moreover, it is employed in first break picking of sonic log transient times. In this study, we describe the automatic VSP velocity analysis workflow and apply it on a synthetic seismic dataset generated through finite-difference modelling in noise-free, noisy, and sparsely recorded data conditions.