Geostatistical interpolation of positively skewed and censored data in a dioxin-contaminated site

A correct delineation of hazardous areas in a contaminated site relies first on accurate predictions of pollutant concentrations, a task usually complicated by the presence of censored data (observations below the detection limit and highly positively skewed histograms. This paper compares the prediction performances of four geostatistical algorithms (ordinary kriging, log-normal kriging, multi-Gaussian kriging, and indicator kriging) through the cross validation of a set of 600 dioxin concentrations. Despite its theoretical limitations, log-normal kriging consistently yields the best results (smallest prediction errors, least false positives, and lowest total costs). The cross validation has been repeated 100 times for a series of sampling intensities, which reduces the risk that these results simply reflect sampling fluctuations, indicator kriging (IK), in the simplified implementation of median IK, produces good predictions except for a moderate bias caused by the underestimation of high dioxin concentrations. Ordinary kriging is the most affected by data sparsity, leading to a large proportion of remediation units wrongly declared contaminated when less than 100 observations were used. Last, decisions based on multi-Gaussian kriging estimates are the most costly and create a large proportion of false positives that cannot be reduced by the collection of additional samples.