Model Quality Impacts Projections of Summer Rainfall
Lawrence Livermore National Laboratory and University of Hawai’i scientists have developed a Gaussian approach for evaluating climate change perturbations to precipitation that equally weights models irrespective of the different number of realizations available from each model. The Gaussian approach is compared to non-parametric Monte Carlo sampling which does not assume an underlying statistical model for the probability distributions (PDF’s) of the CMIP5 Historical and RCP8.5 simulation anomalies. We evaluate PDFs of summer precipitation amount, onset date, withdrawal date, and length of season. We test the sensitivity of the climate change projections to using all models, one model per modeling group, and sub-selecting models based on their fidelity in simulating the annual cycle of precipitation.
The 1% significant lower- and upper-bounds of changes in characteristics of summer precipitation are presented using CMIP5 RCP8.5 and Historical experiments. The results indicate that sub-sampling models on annual cycle quality leads to demonstrable differences in the climate change projection compared to using all models or one model per modeling group. The fidelity evaluation indicates that climate models have the most difficulty in representing the annual cycle of precipitation over the tropics and continental interiors.
We compare late 21st Century and late 20th Century simulations of 5-day averaged precipitation from the CMIP5 RCP8.5 and Historical experiments. We evaluate probability density functions (PDFs) of summer precipitation amount, onset date, withdrawal date, and length of season. Climate change projections were generated using all models, one model per modeling group to account for overconfidence, and sub-selecting models on annual cycle fidelity. The results indicate that sub-sampling models on annual cycle quality lead to demonstrable differences in the climate change projection compared to using all models or one model per modeling group, with differences of up to ±50%, especially in the tropics and subtropics. Sensitivity testing indicates that the Gaussian t-test and the non-parametric Mann-Whitney U-test (the latter using Monte Carlo sampling) yield consistent results for assessing where the climate change perturbation is significant at the 1% level, even in cases where skewness and excess kurtosis indicate non-Gaussian behavior. Similarly, in terms of climate change-induced perturbations to below-normal, normal, and above-normal categorical probabilities, the Gaussian results are typically consistent with the non-parametric estimates. These sensitivity results promote the use of Gaussian statistics to present global maps of the lower-bound and upper-bound of the climate change response, given that the non-parametric calculation of confidence intervals would otherwise not be tractable in a desktop computing environment due to its CPU intensive requirement.