Computationally Efficient Ranking of Sampling Locations by Redundancy in Groundwater Monitoring Networks

The long‐term and frequent sampling of groundwater monitoring networks is common practice for monitoring the concentrations of constituents of potential concern (CoPC) over space and time. Sampling can be a costly and time‐intensive process that also incurs health and safety risks. A redundancy‐based ranking of candidate sampling locations can provide empirical evidence (after statistical modeling) on sites that are more influential for concentration estimation across the network, thus potentially reducing operational costs. Feedback for the open‐source groundwater monitoring data analysis software GWSDAT highlighted the lack of dedicated statistical tools for such analyses. Using cross‐validation (CV) to estimate the redundancy of candidate locations (by evaluating how the removal of each location affects the concentration estimates after model fitting) is computationally inefficient given large monitoring networks, due to the need for repeated model fitting and the large number of potential well combinations. In this paper, a novel, computationally efficient approach is proposed ‘well redundancy analysis’, which closely approximates CV‐based results and only requires one model fitting to acquire an initial redundancy ranking of candidate sampling locations. The approach is based on Cook's distance (CD), an influence diagnostic for regression analysis, which is commonly used to quantify the influence of individual data points on regression models. The proposed approach uses the median CD value (computed after fitting a spline‐based spatiotemporal model to historic observations at each candidate location) to rank the locations. The median CD values of the candidate locations are inversely proportional to their redundancy for the model. Using synthetic and real groundwater CoPC concentration monitoring data, this paper demonstrates that the proposed approach produces redundancy rankings that closely approximate the CV‐based results, while providing computational advantages. The impact of this study is highlighted by the implementation of the proposed approach in the well redundancy analysis feature of GWSDAT.