Combining statistical methods for detecting potential outliers in groundwater quality time series

Quality control of large-scale monitoring networks requires the use of automatic procedures to detect potential outliers in an unambiguous and reproducible manner. This paper describes a methodology that combines existing statistical methods to accommodate for the specific characteristics of measurement data obtained from groundwater quality monitoring networks: the measurement series show a large variety of dynamics and often comprise few (< 25) measurements, the measurement data are not normally distributed, measurement series may contain several outliers, there may be trends in the series, and/or some measurements may be below detection limits. Furthermore, the detection limits may vary in time. The methodology for outlier detection described in this paper uses robust regression on order statistics (ROS) to deal with measured values below the detection limit. In addition, a biweight location estimator is applied to filter out any temporal trends from the series. The subsequent outlier detection is done in z-score space. Tuning parameters are used to attune the robustness and accuracy to the given dataset and the user requirements. The method has been applied to data from the Dutch national groundwater quality monitoring network, which consists of approximately 350 monitoring wells. It proved to work well in general, detecting outliers at the top and bottom of the regular measurement range and around the detection limit. Given the diversity exhibited by measurement series, it is to be expected that the method does not give 100% satisfactory results. Measured values identified by the method as potential outliers will therefore always need to be further assessed on the basis of expert knowledge, consistency with other measurement data and/or additional research.