Automated identification of changes in electrode contact properties for long-term permanent ERT monitoring experiments

Electrical resistivity tomography (ERT) monitoring experiments are being conducted more often to image spatiotemporal changes in soil properties. When conducting long-term ERT monitoring, the identification of suspicious electrodes in a permanent spread is of major importance because changes in electrode contact properties of a single electrode may affect the quality of many measurements on each time-slice. An automated methodology was developed to detect these temporal changes in electrode contact properties, based on a Bayesian approach called "weights of evidence." Contrasts (C-W) and studentized contrasts (C-S) are estimators of the influence of each electrode in the global data quality. A consolidated studentized contrast (CCS) is introduced to consider the proportion of rejected quadripoles which contain a single electrode. These estimators are computed for each time-slice using Q-factor (coefficient of variation of repeated measurements) threshold values, from 0 to 10%, to discriminate between selected and rejected quadripoles. An automated detection strategy is proposed to identify suspicious electrodes by comparing the CCS to the MECS (maximum expected C-S values when every electrode is good for the given data set). These MECS are computed using Monte-Carlo simulations of a hundred random draws where the distribution of Q-factor values follows a Weibull cumulative distribution, with k = 0.421 and lambda = 1.303, fitted on a background data set filtered using a 5% threshold on absolute reciprocal errors. The efficiency of the methodology and its sensitivity to the selected reciprocal error threshold are assessed on synthetic and field data. Our approach is suitable to detect suspicious electrodes and slowly changing conditions affecting the galvanic contact resistances where classical approaches are shown to be inadequate except when the faulty electrode is disconnected. A data-weighting method is finally proposed to ensure that only good data will be used in the inversion of ERT monitoring data sets.