Benefits and Cautions in Data Assimilation Strategies: An Example of Modeling Groundwater Recharge

Abstract

Assimilating recent observations improves model outcomes for real‐time assessments of groundwater processes. This is demonstrated in estimating time‐varying recharge to a shallow fractured‐rock aquifer in response to precipitation. Results from estimating the time‐varying water‐table altitude (h) and recharge, and their error covariances, are compared for forecasting, filtering, and fixed‐lag smoothing (FLS), which are implemented using the Kalman Filter as applied to a data‐driven, mechanistic model of recharge. Forecasting uses past observations to predict future states and is the current paradigm in most groundwater modeling investigations; filtering assimilates observations up to the current time to estimate current states; and FLS estimates states following a time lag over which additional observations are collected. Results for forecasting yield a large error covariance relative to the magnitude of the expected recharge. With assimilating recent observations of h, filtering and FLS produce estimates of recharge that better represent time‐varying observations of h and reduce uncertainty in comparison to forecasting. Although model outcomes from applying data assimilation through filtering or FLS reduce model uncertainty, they are not necessarily mass conservative, whereas forecasting outcomes are mass conservative. Mass conservative outcomes from forecasting are not necessarily more accurate, because process errors are inherent in any model. Improvements in estimating real‐time groundwater conditions that better represent observations need to be weighed for the model application against outcomes with inherent process deficiencies. Results from data assimilation strategies discussed in this investigation are anticipated to be relevant to other groundwater processes models where system states are sensitive to system inputs.