Evaluation of an ensemble of regional climate model simulations over South America driven by the ERA-Interim reanalysis: model performance and uncertainties

Abstract

The capability of a set of 7 coordinated regional climate model simulations performed in the framework of the CLARIS-LPB Project in reproducing the mean climate conditions over the South American continent has been evaluated. The model simulations were forced by the ERA-Interim reanalysis dataset for the period 1990–2008 on a grid resolution of 50 km, following the CORDEX protocol. The analysis was focused on evaluating the reliability of simulating mean precipitation and surface air temperature, which are the variables most commonly used for impact studies. Both the common features and the differences among individual models have been evaluated and compared against several observational datasets. In this study the ensemble bias and the degree of agreement among individual models have been quantified. The evaluation was focused on the seasonal means, the area-averaged annual cycles and the frequency distributions of monthly means over target sub-regions. Results show that the Regional Climate Model ensemble reproduces adequately well these features, with biases mostly within ±2 °C and ±20 % for temperature and precipitation, respectively. However, the multi-model ensemble depicts larger biases and larger uncertainty (as defined by the standard deviation of the models) over tropical regions compared with subtropical regions. Though some systematic biases were detected particularly over the La Plata Basin region, such as underestimation of rainfall during winter months and overestimation of temperature during summer months, every model shares a similar behavior and, consequently, the uncertainty in simulating current climate conditions is low. Every model is able to capture the variety in the shape of the frequency distribution for both temperature and precipitation along the South American continent. Differences among individual models and observations revealed the nature of individual model biases, showing either a shift in the distribution or an overestimation or underestimation of the range of variability.