Machine Learning Analysis of Satellite Imagery for Detection of Permian Basin Water Impoundments

Abstract

Abstract A private company based in Houston, Texas, collects and analyzes water source, logistics, recycling and disposal data for the upstream oil and gas industry. The company recently started applying machine learning methods to satellite imagery of the Permian Basin to identify frac water impoundments and facilitate water trading between operators. Freshwater frac pond locations in Texas are not available from any public record data source, so identification of their locations and other characteristics through satellite imagery analysis may reveal new insights about energy industry operations. The company partnered with the University of Texas Bureau of Economic Geology to analyze eleven quarters of historical central Midland Basin satellite imagery, a time series beginning in January 2016, to test correlations between frac ponds and hydraulic fracturing activity. The study identified a significant positive correlation between total frac pond surface area in the satellite imagery and IHS data on aggregate water use for hydraulic fracturing in the Midland Basin area studied. This study also found a striking 76% increase in average frac pond volume over the 27-month study period, and found that the percentage of available frac pond water injected for hydraulic fracturing grew from 11% to 20% from January 2016 to July 2017. The most notable observation from this study, however, may be that the permit-based (IHS) control data appeared to lose integrity between six and ten months prior to the study date, while the satellite-derived data appeared to maintain integrity right up to the study date. As of May 2018, the IHS data shows water injected for hydraulic fracturing in the Midland Basin dropping 90% beginning in July 2017 because of absence of complete data. The satellite data, however, shows steady growth in total frac water supply throughout that period. This timeliness-of-data comparison was not the purpose of the study, but the study conclusions were limited because reliable control data could not be obtained for the time period of about ten months prior to the study date. This suggests that oilfield market research could be improved through the use of more satellite imagery analytics versus public permit and industy reported data.