Using Statistical Analysis and Computer Vision to Understand Uncertainty in Subsurface Exploration

Abstract

As part of the 4th annual Smoky Mountains Data Challenge hosted by Oak Ridge National Laboratory, we sought to quantify uncertainty in subsurface exploration of the underground to facilitate decision-making. To provide some context, in the collection of seismic data, sounds waves are transmitted into the ground and their reflections recorded by a receiver. However, due to inconsistencies of the subsurface medium, accurate localization of underground layers is difficult without directly digging down to confirm. To combat this issue, we used several statistical and computer vision to quantify uncertainty of seismic data images by labelling each pixel of a seismic survey (realistic models of subsurface density) to indicate its volatility. After thorough analysis, we could conclude that not one “good” metric exists to accomplish our de- fined goal; uncertainty is defined differently depending on the specific methods one employs. Every uncertainty map that was generated using a unique technique highlighted distinct areas of the seismic surveys. More experimentation and feedback from experts are needed to identify what optimal combination of these (or other) techniques would be best to arrive at the best measurement by which to measure subsurface uncertainty.