资源环境科技发展态势分析平台

This study provides the first assessment of CMIP5 model performances in simulating southern Africa (SA) rainfall variability in austral summer (Nov-Feb), and its teleconnections with large-scale climate variability at different timescales. Observed SA rainfall varies at three major timescales: interannual (2-8 years), quasi-decadal (8-13 years; QDV) and interdecadal (15-28 years; IDV). These rainfall fluctuations are, respectively, associated with El Nino Southern Oscillation (ENSO), the Interdecadal Pacific Oscillation (IPO) and the Pacific Decadal Oscillation (PDO), interacting with climate anomalies in the South Atlantic and South Indian Ocean. CMIP5 models produce their own variability, but perform better in simulating interannual rainfall variability, while QDV and IDV are largely underestimated. These limitations can be partly explained by spatial shifts in core regions of SA rainfall variability in the models. Most models reproduce the impact of La Nina on rainfall at the interannual scale in SA, in spite of limitations in the representation of ENSO. Realistic links between negative IPO are found in some models at the QDV scale, but very poor performances are found at the IDV scale. Strong limitations, i.e. loss or reversal of these teleconnections, are also noted in some simulations. Such model errors, however, do not systematically impact the skill of simulated rainfall variability. This is because biased SST variability in the South Atlantic and South Indian Oceans strongly impact model skills by modulating the impact of Pacific modes of variability. Using probabilistic multi-scale clustering, model uncertainties in SST variability are primarily driven by differences from one model to another, or comparable models (sharing similar physics), at the global scale. At the regional scale, i.e. SA rainfall variability and associated teleconnections, while differences in model physics remain a large source of uncertainty, the contribution of internal climate variability is increasing. This is particularly true at the QDV and IDV scales, where the individual simulations from the same model tend to differentiate, and the sampling error increase.