Comparison of statistical methods for determining risk of leakage from soil gas monitoring

The objective of this paper is to test different methods for detecting anomalies in environmental monitoring of CO2 geological storage sites. This type of monitoring relies on a baseline that is performed before injection. Data collected during injection are then compared to this baseline in order to detect deviations from a normal behavior. A robust method for detecting anomalous measurements would have few false positives (i.e. when an anomaly is wrongly detected) and false negatives (when a real anomaly is not detected). We use a dataset of soil gas measurements that were collected in 36 different locations in the surrounding of a CO2 storage pilot in the South-West of France. We consider CO2 and O2 concentrations in soil. The dataset is modified in order to simulate a leak. We compare the performance of four different methods. The first method uses a threshold in CO2 concentrations. The second method compares measurements with a threshold in the CO2 and O2 relationship (respiration line). The third method fits a Bayesian hierarchical model to the data, accounting for variability between measurements locations. The fourth method fits a Bayesian model to the CO2 / O2 relationship. The Bayesian models both use the same method: a “normal” model is fitted to the baseline data. An additional “anomalous” model (same model but with higher deviation) is tested with the measurements collected during injection. Bayesian inference is used in order to check which of these two models fits best to each measurement. With the data used in this paper, the Bayesian version of the CO2 / O2 relationship has the best results in terms of false positives and false negatives. One advantage of using a Bayesian approach is that the result is a probability that a data point is an anomaly, allowing for a natural interpretation. However, Bayesian approaches may be harder to communicate to lay people.