Advances in Surface-Wave Analysis Using Single Sensor Seismic Data and Deep Neural Network Algorithm for Near Surface Characterization

Abstract

AbstractIn hydrocarbon exploration, seismic surface waves are used to characterize the near-surface by imaging the subsurface shear wave velocity for geo-hazard investigation and near surface seismic corrections to avoid false structures in the final seismic image. Surface waves, identified in a conventional surface acquisition experiment, can be analyzed in the frequency wavenumber (FK) domain to generate dispersion curves at each shot location. The subsurface shear wave velocity is represented as a 1D profile with lateral variations can be handled using laterally constrained inversion or by applying spatial interpolation of 1D results. We identify two fundamental challenges to perform surface wave analysis. First, inadequate sampling of the surface wave in conventional sensor arrays may create artifacts in the frequency-wavenumber domain, which introduces further distortion in the signal. The use of broadband single-sensor single source land 3D seismic data provides adequate sampling of surface wave energy that is captured with negligible aliasing and high signal power. This makes it possible to record fundamental and higher surface wave modes at large frequencies. Second, it is common in seismic exploration to deal with large amounts of seismic data on several tens of thousands shot gathers of the single sensor survey making manual picking of dispersion curves a tedious and time-consuming job. We developed a deep belief network (DBN) with multiple hidden layers to pick fundamental modes in the phase velocity spectrum. The neural network workflow was trained on 1500 gathers and validated on several 100 gathers. Finally, the automated picking was applied to roughly 50,000 gathers using frequency range (3-30 Hz). The resulting dispersion curves show high spatial correlation and are geologically consistent. The fundamental mode pseudo-section shows smooth changes with significant lateral variations of Rayleigh-wave phase velocities. The second and third mode of dispersion curves are observed in some shots in the F-K spectrum but usually they have weaker energy than fundamental mode. The recent advances in surface wave analysis is presented over a complex structure where the raw data are characterized by strong Rayleigh waves dominated by a fundamental mode. Dispersion curves were inverted using nonlinear conjugate gradients to generate a shear wave velocity model with high vertical resolution for the first 50 m depth. The recent development in seismic data acquisition using single sensor broad band data, and advances in seismic processing using deep neural network lead to a novel technology that enable automatic picking of dispersion curves.