Quantification of the confidence that can be placed in land-surface model predictions : applications to vegetation and hydrologic processes

The research presented here informs the confidence that can be placed in the simulations of land-surface models (LSMs). After introducing a method for simplifying a complex, heterogeneous land-cover dataset for use in LSMs, I show that LSMs can realistically represent the spatial distribution of heterogeneous land-cover processes (e.g., biogenic emission of volatile organic compounds) in Texas. LSM-derived estimates of biogenic emissions are sensitive (varying up to a factor of 3) to land-cover data, which is not well constrained by observations. Simulated emissions are most sensitive to land-cover data in eastern and central Texas, where tropospheric ozone pollution is a concern. I further demonstrate that interannual variation in leaf mass is at least as important to variation in biogenic emissions as is interannual variation in shortwave radiation and temperature. Model estimates show that more-humid regions with less year-to-year variation in precipitation have lower year-to-year variation in biogenic emissions: as modeled mean emissions increase, their mean-normalized standard deviation decreases. I evaluate three parameterizations of subsurface hydrology in LSMs (with (1) a shallow, 10-layer soil; (2) a deeper, many-layered soil; and (3) a lumped aquifer model) under increasing parameter uncertainty. When given their optimal parameter sets, all three versions perform equivalently well when simulating monthly change in terrestrial water storage. The most conceptually realistic model is least sensitive to errant parameter values. However, even when using the most conceptually realistic model, parameter interaction ensures that knowing ranges for individual parameters is insufficient to guarantee realistic simulation. LSMs are often developed and evaluated at data-rich sites but are then applied in regions where data are sparse or unavailable. I present a framework for model evaluation that explicitly acknowledges perennial sources of uncertainty in LSM simulations (e.g., parameter uncertainty, meteorological forcing-data uncertainty, evaluation-data uncertainty) and that evaluates LSMs in a way that is consistent with models’ typical application. The model performance score quantifies the likelihood that a representative ensemble of model performance will bracket observations with high skill and low spread. The robustness score quantifies the sensitivity of model performance to parameter error or data error. The fitness score ranks models’ suitability for broad application.