Multi-model forecast skill for mid-summer rainfall over southern Africa

Abstract

Southern African December-January-February (DJF) probabilistic rainfall forecast skill is assessed over a 22-year retroactive test period (1980/1981 to 2001/2002) by considering multi-model ensembles consisting of downscaled forecasts from three of the DEMETER models, the ECMWF, Meteo-France and UKMO coupled ocean-atmosphere general circulation models. These models are initialized in such a way that DJF forecasts are produced at an approximate 1-month lead time, i.e. forecasts made in early November. Multi-model forecasts are obtained by: i) downscaling each model’s 850 hPa geopotential height field forecast using canonical correlation analysis (CCA) and then simply averaging the rainfall forecasts; and ii) by combining the three models’ 850 hPa forecasts, and then downscaling them using CCA. Downscaling is performed onto the 0.5° × 0.5° resolution of the CRU rainfall data set south of 10° south over Africa. Forecast verification is performed using the relative operating characteristic (ROC) and the reliability diagram. The performance of the two multi-model combinations approaches are compared with the single-model downscaled forecasts and also with each other. It is shown that the multi-model forecasts outperform the single model forecasts, that the two multi-model schemes produce about equally skilful forecasts, and that the forecasts perform better during El Nino and La Nina seasons than during neutral years.