Enhancing Seismic Data Augmentation with Geostatistical Features and Internal Multiples for Improved Deep Learning Applications

Abstract Seismic Data Augmentation (SDA) is a comprehensive framework designed to generate realistic synthetic seismic datasets for deep learning applications in seismic interpretation. It addresses the limitations of conventional data acquisition by producing structurally and geologically meaningful data that reflect the complexity and variability of real subsurface conditions. We present an advanced SDA workflow that synthesizes realistic seismic datasets through the integration of three key components: structural interpolation, stochastic simulation, and wavefield modeling. First, Image-Guided Interpolation (IGI) is applied to construct continuous velocity models from sparse well log inputs while preserving large-scale structural trends, with the option to incorporate faults and horizons if provided as input. Second, Sequential Gaussian Simulation (SGS) introduces geostatistical perturbations at smaller scales, enriching the models with spatial variability consistent with subsurface heterogeneity. Third, reflectivity-based modeling is used to generate synthetic seismic data that can include primary reflections, internal multiples, or their combination, offering flexibility in simulating realistic wavefield responses. The SDA framework is demonstrated on a public field dataset to illustrate its ability to generate structurally coherent, statistically diverse, and physically plausible seismic volumes. the resulting augmented data provide a strong foundation for both qualitative interpretation and future ML-based workflows. This work highlights the potential of SDA as a scalable and practical solution for generating synthetic seismic datasets that closely resemble real-world observations.