Pothole detection and segmentation in the Bushveld Complex using physics-based data augmentation and deep learning

Abstract

Potholes are local depression structures that disrupt stratigraphic continuity, such as in layered igneous intrusions. In the Bushveld Complex (South Africa), potholes range from a few to hundreds of meters in width, and may disrupt orebodies, cause ore loss and pose geotechnical challenges. However, potholes are of scientific value as they are proxies of magma chamber processes that are not directly observable. Unfortunately, it is seldom possible to map the full 3D geometry of potholes directly. Reflection seismics has the potential to map many potholes indirectly. However, the accurate segmentation of potholes in seismic data remains unresolved, particularly using geodata science-based methods. Here, we present a prototype segmentation framework that: (1) uses a physics-based, forward modelling method to synthesize 3D reflection seismic data and augments the training data; and (2) implements a standard deep learning, voxel classification-based pothole detection workflow using the data generated in step (1). Both components of the framework are general enough to permit further development, for example, as deep-learning architectures evolve or as the knowledge of potholes improve. We demonstrate that a self-reinforcing feedback loop of knowledge-driven data engineering and deep learning has the potential to overcome data quality issues in supervised tasks of seismic data analysis. We apply the trained model on augmented data to 3D seismic data acquired from a platinum group element Bushveld Complex orebody and demonstrate that automated pothole prediction is practical. Furthermore, physics-based data augmentation, as opposed to inferential types, provides a realistic path to recursive data augmentation that does not incur problems caused by the use of inferential data synthesis, such as model collapse.