Towards a comprehensive approach to parameter estimation in land surface parameterization schemes

In climate models, the land-atmosphere interactions are described numerically by land surface parameterization (LSP) schemes. The continuing improvement in realism in these schemes comes at the expense of the need to specify a large number of parameters that are either directly measured or estimated. Also, an emerging problem is whether the relationships used in LSPs are universal and globally applicable. One plausible approach to evaluate this is to first minimize uncertainty in model parameters by calibration. In this paper, we conduct a comprehensive analysis of some model diagnostics using a slightly modified version of the Simple Biosphere 3 model for a variety of biomes located mainly in the Amazon. First, the degree of influence of each individual parameter in simulating surface fluxes is identified. Next, we estimate parameters using a multi-operator genetic algorithm applied in a multi-objective context and evaluate simulations of energy and carbon fluxes against observations. Compared with the default parameter sets, these parameter estimates improve the partitioning of energy fluxes in forest and cropland sites and provide better simulations of daytime increases in assimilation of net carbon during the dry season at forest sites. Finally, a detailed assessment of the parameter estimation problem was performed by accounting for the decomposition of the mean squared error to the total model uncertainty. Analysis of the total prediction uncertainty reveals that the parameter adjustments significantly improve reproduction of the mean and variability of the flux time series at all sites and generally remove seasonality of the errors but do not improve dynamical properties. Our results demonstrate that error decomposition provides a meaningful and intuitive way to understand differences in model performance. To make further advancements in the knowledge of these models, we encourage the LSP community to adopt similar approaches in the future.