On the Verification and Comparison of Extreme Rainfall Indices from Climate Models

Abstract The interpretation of model precipitation output (e.g., as a gridpoint estimate versus as an areal mean) has a large impact on the evaluation and comparison of simulated daily extreme rainfall indices from climate models. It is first argued that interpretation as a gridpoint estimate (i.e., corresponding to station data) is incorrect. The impacts of this interpretation versus the areal mean interpretation in the context of rainfall extremes are then illustrated. A high-resolution (0.25° × 0.25° grid) daily observed precipitation dataset for the United States from Climate Prediction Center (CPC) is used as idealized perfect model gridded data. Both 30-yr return levels of daily precipitation (P30) and a simple daily intensity index are substantially reduced in these data when estimated at coarser resolution compared to the estimation at finer resolution. The reduction of P30 averaged over the conterminous United States is about 9%, 15%, 28%, 33%, and 43% when the data were first interpolated to 0.5° × 0.5°, 1° × 1°, 2° × 2°, 3° × 3°, and 4° × 4° grid boxes, respectively, before the calculation of extremes. The differences resulting from the point estimate versus areal mean interpretation are sensitive to both the data grid size and to the particular extreme rainfall index analyzed. The differences are not as sensitive to the magnitude and regional distribution of the indices. Almost all Intergovernmental Panel on Climate Change (IPCC) Fourth Assessment Report (AR4) models underestimate U.S. mean P30 if it is compared directly with P30 estimated from the high-resolution CPC daily rainfall observation. On the other hand, if CPC daily data are first interpolated to various model resolutions before calculating the P30 (a more correct procedure in our view), about half of the models show good agreement with observations while most of the remaining models tend to overestimate the mean intensity of heavy rainfall events. A further implication of interpreting model precipitation output as an areal mean is that use of either simple multimodel ensemble averages of extreme rainfall or of intermodel variability measures of extreme rainfall to assess the common characteristics and range of uncertainties in current climate models is not appropriate if simulated extreme rainfall is analyzed at a model’s native resolution. Owing to the large sensitivity to the assumption used, the authors recommend that for analysis of precipitation extremes, investigators interpret model precipitation output as an area average as opposed to a point estimate and then ensure that various analysis steps remain consistent with that interpretation.