<p>The success of data assimilation systems strongly depends on the suitability of the generated ensembles. While in theory data assimilation should correct the states of an ensemble of models, especially if model parameters are included in the update, its effectiveness will depend on many factors, such as ensemble size, ensemble spread, and the proximity of the prior ensemble simulations to the data. In a previous study, we generated an ensemble-based data-assimilation framework to update model states and parameters of a coupled land surface-subsurface model. As simulation system we used the Terrestrial Systems Modeling Platform TerrSysMP, with the community land-surface model (CLM) coupled to the subsurface model Parflow. In this work, we used the previously generated ensemble to assess the effect of uncertain input forcings (i.e. precipitation), unknown subsurface parameterization, and/or plant physiology in data assimilation. The model domain covers a rectangular area of 1&#215;5km<sup>2</sup>, with a uniform depth of 50m. The subsurface material is divided into four units, and the top soil layers consist of three different soil types with different vegetation. Streams are defined along three of the four boundaries of the domain. For data assimilation, we used the TerrsysMP PDAF framework. We defined a series of data assimilation experiments in which sources of uncertainty were considered individually, and all additional settings of the ensemble members matched those of the reference. To evaluate the effect of all sources of uncertainty combined, we designed an additional test in which the input forcings, subsurface parameters, and the leaf area index of the ensemble were all perturbed. In all these tests, the reference model had homogenous subsurface units and the same grid resolution as all models of the ensemble. We used point measurements of soil moisture in all data assimilation experiments. We concluded that precipitation dominates the dynamics of the simulations, and perturbing the precipitation fields for the ensemble have a major impact in the performance of the assimilation. Still, considerable improvements are observed compared to open-loop simulations. In contrast, the effect of variable plant physiology was minimal, with no visible improvement in relevant fluxes such as evapotranspiration. As expected, improved ensemble predictions are propagated longer in time when parameters are included in the update.</p>
The Land Variational Ensemble Data Assimilation fRamework: LaVEnDAR
