Effects of Spatiotemporal Constraints on Geostatistical Analysis of Groundwater Depth

Geostatistical evaluation of the groundwater depth (GWD) in California's South Coast hydrologic region, and its sensitivity to different spatiotemporal assumptions, is presented in this paper. We obtain a pseudo-stationary representation of the groundwater depth, using the publicly available, online database from the GAMA GeoTracker project, while tracking the associated uncertainty throughout the process. We create nine different sub-datasets, using different temporal constraints, such as seasonal partitioning and different long-term variability filtering criteria. The geostatistical analysis and comparison between the different maps highlight the trade-off between spatial and temporal accuracy. For example, when moving to stricter filtering criteria, despite removing a large number of sites from the interpolation, the root mean squared error (RMSE) calculated in the analysis either decreased or only slightly increased. This suggests that the long-term variability filter is a good representation of the GWD accuracy and that the cross-validation RMSE captures both the stability effect as well as spatial density of the measurement points. We further find that the point-specific standard error is strongly correlated with the associated GWD prediction and that the mean relative error is approximately 60% of the prediction. Hence, it is highly recommended to account for such error in a forward-engineering application, by introducing a GWD distribution rather than a single value into the analysis. Finally, we analyze seasonal fluctuations in the study region and find that they are on average 2.5 m with a standard deviation of 8 m.