Data Assimilation in Integrated Subsurface Flow Models: Making Optimal Use of Cross-Compartmental Interactions

The ensemble Kalman filter is often used for data assimilation (DA) to forecast states and fluxes in integrated subsurface flow problems. Integrated subsurface flow models include two compartments: the vadose zone and groundwater aquifers. In DA with such models, observations from both compartments can be used with the goal of improving forecasts in the whole system. To study the potential risks and benefits of cross-compartmental DA in this context, we apply different DA methods, drawing observations from multiple heterogeneous virtual realities. Updating soil moisture with point observations can consistently improve spatially averaged soil moisture predictions at the land surface but frequently deteriorates estimates of the groundwater table height (worse for models with non-resolved layers). Bias correction and vertical localization can mitigate these deteriorations. Groundwater table assimilation that is limited to updating aquifer states can improve the estimation of spatially averaged groundwater tables (RMSEs on average reduced by (Formula presented.) in our test examples), but rigorously taking out the partly saturated parts introduces balancing fluxes that impair the benefits of the groundwater assimilation. Additionally updating pressure heads in an area above the groundwater table can, on average, reduce RMSEs of groundwater table estimates by more than double (Formula presented.). Multivariate assimilation of both soil moisture and groundwater tables leads to the best results for predicting root zone soil moisture. Groundwater recharge predictions could often be improved in a subsequent forecast without DA if groundwater tables had been updated before.