Assessment of model uncertainty for soil moisture through ensemble verification

The Community Land Model (CLM2.0) has been used to simulate land surface processes in a small corn field. The subdivision of grid cells into patches in the CLM2.0 was explored for the generation of Monte Carlo simulations for use in calibration and ensemble generation. A distributed multiobjective calibration was developed for the optimal estimation of parameters and initial state variables for 36 soil moisture profiles. Since the resulting parameter and initial state values did not lead to perfect simulations for soil moisture, and in order to better understand the forecast uncertainty, ensemble runs were generated. The ensembles generated by CLM2.0 have been verified by several methods that are commonly used in meteorology. It was shown that the perfect model approach cannot be applied for bounded hydrological applications and that perturbation of parameters is a necessity to obtain a realistic assessment of the forecast error. Perturbation of forcings only captures more of the model uncertainty than perturbation of initial conditions only, but also causes a too limited spread in the ensembles. The generation of ensemble members through perturbation of the parameter set, found through calibration, does not necessarily result in ensembles that surround the calibrated deterministic control run for soil moisture. This is partially due the nonlinearity of the model in the parameters. It may also indicate that some parameter sets are not robust and not appropriate to perturb for ensemble generation. Consequently, the resulting ensemble mean may not represent the best forecast or a priori state estimation. During periods of extreme drought or precipitation, the ensemble probability density function (pdf) deviates far from normality and the model behaves very nonlinearly. For state estimation, methods like the ensemble Kalman filter are best suited for the propagation of the first moments to account for the nonlinear dynamics during crucial events for hydrological simulations. However, the a posteriori estimate for this technique will only be optimal in the limited class of linear filters, since the underlying pdfs cannot be assumed to be Gaussian.