Parameter Estimation of a Regional Seawater Intrusion Model Using Well and Geophysical Data: Insights from Data Assimilation and Data Worth

Abstract:

Seawater intrusion is a main concern for groundwater management in coastal and island aquifers, and numerical models are a key tool for managing these issues. Seawater intrusion models generally use variable density codes, which solve the groundwater flow and solute transport equations. These models are computationally expensive, which limits the possibility of parameter estimation, and require concentration data to constrain the solute transport equation. When building regional models, sharp-interface codes are a computationally efficient alternative in which mixing processes are not simulated. While these offer reduced run times and facilitate parameter estimation, few data assimilation methods have been described and it is unclear which observations should be used. We developed a data assimilation method for a regional, sharp-interface model designed for management purposes. We built a sharp-interface model for an island aquifer using the SWI2 package for MODFLOW and implemented a correction factor to rectify interface elevations. We then assembled an observation dataset including ‘interface’ observations alongside the more traditional head observations. The former comprised both direct observations from open wells and geophysically-derived, indirect observations from time-domain electromagnetics (TDEM) and electrical resistivity tomography (ERT) surveys. All data types and their associated uncertainties were assimilated to constrain parameter estimation, which was conducted using the model-independent code PEST. A data worth analysis was finally carried out using linear analysis. A satisfying fit was obtained between simulated and observed data given the effective implementation of the correction factor. Model residuals provided insight on the potential of different observation groups to constrain parameter estimation. The data worth analysis provided insight on the importance of different observations types in reducing the uncertainty of model forecasts (head vs interface observations, direct vs indirect interface observations). Ultimately, the study provides recommendations for data assimilation strategies for future regional sharp-interface models. We demonstrated this approach on a real-world example in the Magdalen Islands (Quebec, Canada).