We demonstrate that integrating 3D reflection seismics with machine learning (ML) can bring many benefits for the future development of the mining industry. We use a serial integration of reflection seismics, which identifies economic horizon-depression structures known as potholes within the western Bushveld Complex. Thereafter, agglomerative clustering is applied to the resulting data, using features engineered from the physical characteristics of the potholes. Our results indicate that potholes can be divided into several classes based on characteristic features; e.g., large potholes are substantially less steep than small potholes. Furthermore, we model this empirical relationship and show that it can be used to predict average sizes of potholes given their typical in-mine exposures. We also demonstrate that pothole formation is likely to have been initiated depth-wise, followed by lateral increases in size. Lastly, we demonstrate that our serial application of seismically based data generation and ML-based data analytics is a viable alternative to conventional geostastistical analysis, especially for the classification, prediction, and modeling of geologic structures such as potholes.