Stratification of climate projections for efficient estimation of uncertainty and variation using weather-driven models

The range of uncertainties inherent in climate models can only be portrayed by provision of multiple climate projections. Unfortunately, such provision poses a challenge to model-based impact studies, since driving the relevant impact models using weather data from large numbers of climate projections may not be computationally feasible. Hence, it is important to investigate how to draw sub-samples of climate projections in a manner that reduces the subsequent computational burden. We describe a stratification-based protocol for sub-sampling climate projections to drive crop models with strata based on changes in mean temperature and changes in relative mean rainfall. As an example of the protocol's utility, simulated weather for each selected climate projection was used to drive 3 contrasting process-based models of plant–environment interactions to predict yields of spring barley, managed grassland, and short-rotation coppice. Many of the questions about potential impact that we wish to answer are related to variation in predicted yields. Variance components analyses of predicted yields for each of 2 time periods (2040s and 2080s) indicated that, after allowing for variability between grid squares, between 16 and 61% of the remaining variance in annual yields was uncertainty due to climate projections, the corresponding range for mean yields over 9 yr being from 63 to 93%. We found that our stratification procedure enhanced the precision in the estimate of the variance component due to climate projection, enabling reductions of up to 20% in the number of climate projections required to achieve equivalent precision compared to simple random sampling.