Machine Learning Approaches for Near-Surface Velocity Model Building

Abstract Near-surface complexities pose major challenges for land seismic data processing and interpretation. Accurate modeling of the near-surface velocity is key to compensate the effects of these shallow heterogeneities, which can distort seismic signals and reduce image quality. Uphole data are of great importance in improving the estimation of the near-surface velocity model as they provide direct vertical travel time measurements of the seismic wave traveling through the weathered layers. Near-surface attributes, namely velocity, depths and thicknesses, can be characterized by the interpretation of the uphole travel time data. In this study, we develop an automated velocity model-building method of uphole data based on neural networks to address the need for consistency and efficiency. The goal is standardizing and automating what is otherwise a laborious and error-prone interpretation workflow that often relies on subjective analyst input. The neural network model, trained with synthetically generated uphole data, is able to construct velocity profiles from uphole travel time surveys by learning from representative patterns in the input data. We demonstrate the effectiveness of the trained model in producing robust velocity predictions using synthetic as well as real uphole datasets, highlighting its potential for improving near-surface velocity characterization and supporting seismic processing in complex environments.