Inverse conditioning of a high resolution integrated terrestrial model at the hillslope scale: the role of input data quality and model structural errors

Understanding the soil-vegetation-atmosphere continuum is essential to improve hydrological model predictions. Particularly the characterization and prediction of the spatio-temporal variability of soil water content (SWC) and its controlling factors are of high interest for many geoscientific fields, since these patterns influence for example the rainfall-runoff response and the partitioning of the net radiation into latent and sensible heat fluxes while interacting with the vegetation cover. Within this context, this PhD thesis explores the degree of model complexity that is necessary to adequately represent heterogeneous subsurface processes, and the benefit of merging soil moisture data with an integrated terrestrial model. This includes an uncertainty analysis of model forcing (i.e. precipitation) and evaluation data (actual evapotranspiration). On this account, the fully coupled land surface-subsurface model ParFlow-CLM, which is part of the terrestrial system modeling platform (TerrSysMP), was applied to the 38 ha Rollesbroich headwater catchment located in the Eifel (Germany). Detailed long-term data for model setup, calibration, and evaluation were provided by the TERENO infrastructure initiative, the North Rhine-Westphalian State Environment Agency, and the Transregional Collaborative Research Center 32. It was expected that this combination of process orientated model and extensive observation data contributes to the understanding of the complex processes of the energy and water cycle at the hillslope, the elementary unit for the runoff generation process. The first part compared different measurements of actual evapotranspiration (ETa) obtained by a set of six weighable lysimeters, ETa estimates acquired with the eddy covariance (EC), and evapotranspiration calculated with the full-form Penman Monteith equation (ETPM) for the Rollesbroich site. This comparison of ETa included a correction of the energy balance deficit for the EC method, which is not often reported in literature and allows a deeper insight into the performance of both methods. The evaluation was conducted for the year 2012 and showed that both measurement methods are in good agreement with a total difference of 3.8 % (19 mm). The smallest relative difference (< 8 %) between the monthly ETa of lysimeters and EC was found in summer. Both ETa estimates were close to the ETPM which was used as reference for the observations. This indicates that ETa was energy limited but not limited by water availability during the entire evaluation period. The ETa difference between lysimeter and EC were found to be mainly related to differences in grass height due to harvest and EC footprint. The lysimeter data further were used to estimate precipitation in combination with a filter and smoothing algorithm for high-precision lysimeters by Peters et al. (2014). The comparison with the on-site standard tipping bucket rain gauge revealed that precipitation amounts from the lysimeter were 16 % higher for 2012. Applying the correction method of Richter (1995) reduced this amount to 3 %. With the help of an on-site camera system, it was shown that dew and rime contributed 17 % to the total yearly difference between both measurement methods, while fog and drizzling rain explained another 5.5 %. During snowfall, the tipping bucket underestimated precipitation (7.9 % of the total difference). The largest proportion of the total yearly difference (36 %) was associated with snow cover without apparent snowfall. The strong overestimation of precipitation under these conditions most likely can be explained with snow drift and snow bridges. The remaining 33 % of the total yearly precipitation difference could not be explained. In the second part, the results of TerrSysMP simulations with the 3D variably saturated groundwater flow model (ParFlow) and the Community Land Model (CLM) are compared to soil water content measurements from a wireless sensor network, actual evapotranspiration recorded by lysimeters and EC, and discharge. Therefore, the ParFlow-CLM model with a lateral resolution of 10 × 10 m and a variable vertical resolution (0.025-0.575 m) was tested with different subsurface parametrization strategies: (i) completely homogeneous, (ii) homogeneous parameters for different soil horizons, (iii) different parameters for each soil unit and soil horizon and (iv) heterogeneous stochastic realizations. The soil hydraulic parameters for these simulations were either (i) sampled from measured soil texture and the Rosetta pedotransfer functions (ROS), or (ii) estimated using a 1D subsurface model in combination with shuffle complex evolution (SCE). The results showed that spatially heterogeneous soil hydraulic properties in combination with topography dominate the spatial variability of SWC at the hillslope scale. Whereas topography in combination with homogeneous soil hydraulic properties underestimated the observed spatial SWC variability, the full SWC variability could be explained with a heterogeneous distribution of soil hydraulic parameters. All model simulations reproduced the seasonal evapotranspiration, while the discharge showed a high bias compared to the observations. This poor performance of the uncalibrated model could be explained with a lack of information about the bedrock characteristics and the on-site drainage system. In general, a better performance was found for the simulation of the SCE setup with interpolated 1D inverse parameter estimates. Showing a higher inverse air entry parameter within the van Genuchten subsurface parameterization, the SCE-setup was in better correspondence with the observed SWC during dry periods than the simulation of the ROS-setup. This demonstrates the difficulty to transfer small scale local measurements into larger scale. In the third part, SWC data was assimilated with Ensemble Kalman Filter (EnKF) into a Parflow-CLM Model of the Rollesbroich study site. Dense information from a soil moisture sensor network was available at the site and assimilated with the EnKF in a coupled land surface-subsurface model with 10 m lateral resolution. Different data assimilation scenarios with the EnKF were applied including the update of model states with or without updating of model parameters as well as different time update intervals for parameters and states. The ensemble size was 128 (or 256) with 3D- heterogeneous fields of saturated hydraulic conductivity and Mualem-van Genuchten parameters α and n. Furthermore, simulations were also carried out with a synthetic test case mimicking the Rollesbroich site, with the aim to get more insight in the role of model structural errors. The results of the real-world scenarios showed that the exclusive update of soil hydraulic properties was not enough to improve the SWC characterization. On average the RMSE at the sensor locations was reduced by 14 % if states and parameters were updated jointly and discharge estimation was not improved significantly. Simulations for the synthetic case showed much better results with an overall RMSE reduction by 55 % at independent verification locations in case of daily SWC-data assimilation including parameter estimation. Individual data assimilation scenarios with parameter estimation also showed NSE-increase for discharge from -0.04 for the open loop run to ca. 0.61. This shows that data assimilation in combination with high-resolution physically based models can potentially strongly improve soil moisture and discharge estimation at the hillslope scale. The large differences in the performance for the synthetic case and the real-world case can be explained with model structural errors like the representation of the on-site drainage system. Further studies should show whether the role of model structural errors was site specific, or that in general small scale processes play a more dominant role which are not captured by the modelling approach. In summary, these studies show that integrated terrestrial high-resolution models at the hillslope scale are a valuable tool to better understand the basic hydrological processes. However, on the way to a predictive application of high resolution models many challenges remain. For more complex models a larger parameter space has to be explored (e.g. by additional soil hydraulic properties) in combination with multiple observation types. Further, an improved representation of significant processes (e.g. preferential flow) will contribute to the benefit of data-fusion techniques. On this account, high-resolution integrated terrestrial models can be used to identify significant information gaps in the existing measurement infrastructure