Skill assessment and hybrid statistical-dynamical approach through teleconnection-based subsampling to improve seasonal rainfall forecasts over Sub-Saharan Africa

Teleconnections are a key source of seasonal predictability, particularly at tropical latitudes where the El Niño Southern Oscillation (ENSO), the Indian Ocean Dipole (IOD) and the Atlantic Niño (ATL) drive a large portion of rainfall variability. Global Circulation Models (GCMs) still struggle to reproduce these modes of variability correctly, limiting the forecast skill over vulnerable regions such as Sub-Saharan Africa. This study presents a two-step framework developed within the ALBATROSS (Advancing knowledge for Long-term Benefits and climate Adaptation ThRough hOlistic climate Services and nature-based Solutions) project. First, we provide a robust assessment of seasonal rainfall skill over Sub-Saharan Africa using a set of multiple models (ECMWF, CMCC, UKMO, DWD, Météo France) and multiple observational datasets (ERA5, GPCP v2.3, CHIRPS v2.0) over the hindcast period 1993-2016. Results identify robust hotspots of predictability across regions and seasons, that are independent of the dataset used. These include East Africa during the October-December (OND) short rains and Southern Africa during January-March (JFM). In contrast, predictability over West Africa during boreal summer (July-September, JAS) is strongly dataset dependent. Second, based on these results, we apply a statistical-dynamical hybrid approach, named the teleconnection-based subsampling, in which AI-based prediction of teleconnection indices is used as a priori information to subsample GCM ensemble members and to generate improved hybrid rainfall forecasts. Convolutional Neural Networks (CNNs), which have been shown to outperform traditional modelling techniques in predicting modes of climate variability, are trained on Sea Surface Temperature anomalies and employed to predict the teleconnection index most relevant to each region and season, selected on the basis of both the skill assessment results and the known physical influence of teleconnection on seasonal rainfall. The CNN architectures are adapted from previous studies. Over East Africa, a CNN trained to predict the IOD index for OND at three months lead time results in hybrid rainfall forecasts that outperform both purely dynamical and purely AI-based approaches, with the largest skill improvements along the coasts of Kenya and Tanzania. Over West Africa, a combination of ENSO and ATL CNN-based predictions highlights the potential of this hybrid methodology during the JAS season, with notable improvements over Ghana. Over Southern Africa, limited improvements suggest that additional drivers, including extratropical modes of variability, may need to be incorporated in future work.  These results demonstrate the value of combining multi-model and multi-observational dataset skill assessment with hybrid methodologies, based on a better knowledge of climate teleconnections, to enhance seasonal rainfall forecast over Sub-Saharan Africa.