The good, the bad and the outliers: automated detection of errors and outliers from groundwater hydrographs

Peterson, Tim J.; Western, Andrew W.; Cheng, Xiang

The good, the bad and the outliers: automated detection of errors and outliers from groundwater hydrographs | Les bonnes, les mauvaises et les aberrantes: détection automatisée des erreurs et des données aberrantes des hydrogrammes des eaux souterraines Lo bueno, lo malo y lo extraño: detección automatizada de errores y valores atípicos de los hidrogramas de agua subterránea 理想值、不理想值及异常值:地下水水位图中误差和异常值的自动检测 O bem, o mau e os dados discrepantes: detecção automática de erros e dados discrepantes a partir de hidrogramas das águas subterrâneas

2018

Peterson, Tim J. | Western, Andrew W. | Cheng, Xiang

Suspicious groundwater-level observations are common and can arise for many reasons ranging from an unforeseen biophysical process to bore failure and data management errors. Unforeseen observations may provide valuable insights that challenge existing expectations and can be deemed outliers, while monitoring and data handling failures can be deemed errors, and, if ignored, may compromise trend analysis and groundwater model calibration. Ideally, outliers and errors should be identified but to date this has been a subjective process that is not reproducible and is inefficient. This paper presents an approach to objectively and efficiently identify multiple types of errors and outliers. The approach requires only the observed groundwater hydrograph, requires no particular consideration of the hydrogeology, the drivers (e.g. pumping) or the monitoring frequency, and is freely available in the HydroSight toolbox. Herein, the algorithms and time-series model are detailed and applied to four observation bores with varying dynamics. The detection of outliers was most reliable when the observation data were acquired quarterly or more frequently. Outlier detection where the groundwater-level variance is nonstationary or the absolute trend increases rapidly was more challenging, with the former likely to result in an under-estimation of the number of outliers and the latter an overestimation in the number of outliers.

Show more [+]