Biases and Model Agreement in Projections of Climate Extremes over the Tropical Pacific

Abstract

Abstract Using the Coupled Model Intercomparison Project phase 3 (CMIP3) general circulation models (GCMs), projections of a range of climate extremes are explored for the western Pacific. These projections include the 1-in-20-yr return levels and a selection of climate indices for minimum temperature, maximum temperature, and precipitation, and they are compared to corresponding mean projections for the Special Report on Emission Scenarios (SRES) A2 scenario during 2081–2100. Models are evaluated per variable based on their ability to simulate current extremes, as well as the overall daily distribution. Using the standardized evaluation scores for each variable, models are divided into four subsets where ensemble variability is calculated to measure model uncertainty and biases are calculated in respect to the multimodel ensemble (MME). Results show that higher uncertainty in projections of climate extremes exists when compared to the mean, even in those subsets consisting of higher-skilled models. Higher uncertainty exists for precipitation projections than for temperature, and biases and uncertainties in the 1-in-20-yr precipitation events are an order of magnitude higher than the corresponding mean. Poorer performing models exhibit a cooler bias in the mean and 1-in-20-yr return levels for maximum and minimum temperature, and ensemble variability is low among all subsets of mean minimum temperature, especially the lower-skilled subsets. Higher-skilled models project 1-in-20-yr precipitation return levels that are more intense than in the MME. The frequency of temperature extremes increase dramatically; however, this is explained by the underpinning small temperature range of the region. Although some systematic biases occur in the higher- and lower-skilled models and omitting the poorer performers is recommended, great care should be exercised when interpreting the reduction of uncertainty because the ensemble variability among the remaining models is comparable and in some cases greater than the MME. Such results should be treated on a case-by-case basis.