Proceedings of the 2019 Mississippi Water Resources Conference

Year: 2019 Authors: Asquith W.H.

Information gaps can be detected by quantifying statistical efficacy in estimation of phenomena such as groundwater levels at unmonitored locations for the Mississippi River Valley alluvial (MRVA) aquifer located within the Mississippi Alluvial Plain (MAP), south-central United States. Multi-agency water-level networks containing wells screened in the MRVA aquifer collect data in space (horizontal and vertical dimension) and time. Groundwater levels are also influenced by a given hydrogeologic framework (aquifer geometry and properties), well construction, local and regional pumping histories, and contexts of seasonal recharge and discharge. One common stakeholder inquiry concerns identification of information gaps. To quantify information gaps, a two-method approach for water-level prediction is proposed. Two statistics of interest were spring 2018 maxima (March-May) and fall 2018 minima (September-November) based on use of water-level data collected during these same months from 2014-2018 with prediction made for 2018. Spring maxima represent maximum seasonal aquifer recovery, whereas fall minima represent maximum aquifer drawdown attributable in part to irrigation demands. Data included for this study were computed from 1,411 unique wells for which 6,304 measurements (discrete or daily mean) were available. Our focus is not on the estimation of water levels per se but on the divergence between estimates using two methods (generalized additive models [GAMs] and support vector machines [SVMs]). Spatial coordinates, land-surface altitude, MRVA aquifer bottom altitude, and year were used as predictor variables. GAMs and SVMs are powerful estimation methods in their own right, but by their radically different mathematics, perform differently as extrapolation increases when predictions are increasingly made away from hyperspace of predictor variables and not necessarily away from spatial coordinates. GAMs can have curvatures away from the global mean, but SVMs must curve back to the global mean. Throughout the MAP and aligned to the 1-kilometer National Hydrogeologic Grid (NHG), absolute differences between GAM and SVM predictions were computed. Spatial depiction of the results on the NHG are shown for the entire MAP as well as for subdivision-specific GAM and SVM computations for the Boeuf, Cache, Delta, Grande Prairie, and St. Francis subdivisions. Various local areas in the MAP can be seen with large GAM-SVM divergence, and hence these areas have potential information gaps, indicating the need for additional water-level monitoring. Stakeholders are thus provided information on which to judge allocation of future resources in monitoring of the MRVA aquifer.