Modelling hydrocarbon concentrations in groundwater monitoring networks

Abstract

Groundwater networks provide a critical resource across the world by supplying fresh water for a wide array of scenarios including extraction for drinking water and irrigation. Protection of these naturally occurring geological features is an important component of the wider climate problem. Pollution to groundwater networks can occur in many forms, including nitrates, radioactive material and the focus of this thesis, hydrocarbons. Due to their carcinogenic nature, data is collected at groundwater monitoring sites for regulatory compliance and to ensure safe concentration levels are not exceeded. Data collection involves extraction of a water sample, in situ, to be later analysed in a laboratory capable of measuring hydrocarbon concentrations above a certain “non-detection” limit. This process is less than desirable as our data is left-censored at laboratory-dependent thresholds and it requires the construction of several groundwater monitoring wells. Furthermore, observations may be missed due to faulty wells, unsafe working conditions and other potential obstructions. Hence, the aim of this thesis is to investigate whether statistical modelling of hydrocarbon concentrations based on measurements of predictors that are easier to obtain can provide more insight with less information. Models proposed in this thesis take the form of a regression where the dependent variable is a left-censored analyte of interest and the regressors are indicators of water quality such as temperature, pH and dissolved oxygen that could be more feasibly obtained using sensors and telemetry in the future. An application with such complexity requires an inter-disciplinary approach and this thesis presents an exploratory data analysis, machine learning methods and mechanistic transport models based on physical laws. Following these results, we propose models that avoid replacing censored data with half the detection limit; leverage the high correlation between analytes; apply mixture models to deal with non-linearity and a varying intercept model that makes use of the spatial aspect of the wells from which the data are sampled.