Hybrid seasonal rainfall predictions in sub-Saharan Africa through a teleconnection-based subsampling, informed by AI model

Seasonal forecasts generated by General Circulation Models (GCMs) provide essential information for early warning and climate services, particularly in regions highly vulnerable to rainfall variability, such as sub-Saharan Africa. To enhance the predictive skill of GCMs, hybrid forecasting systems have been developed that combine physically based dynamical models with data-driven models.  

In this study, we apply a statistical-dynamical hybrid method, referred to as teleconnection subsampling, in which AI-based prediction of large-scale teleconnection indices influencing sub-Saharan Africa rainfall -such as the El Niño Southern Oscillation (ENSO), the Indian Ocean Dipole (IOD) and the Atlantic Niño (ATL)- is used as a priori information to select a subsample of GCM ensemble members and generate hybrid rainfall forecast. Previous studies have demonstrated that Convolutional Neural Networks (CNNs) outperform traditional modelling techniques in predicting modes of climate variability; therefore, a CNN-based prediction of a teleconnection index is adopted in this work. 

The analysis is based on the ECMWF seasonal forecasting system, provided by the Copernicus Climate Change Service, with ERA5 reanalysis used as the observational reference for skill assessment. The study focuses on three regions -East, West and Southern Africa- characterized by distinct rainfall regimes and lying within the framework of the ALBATROSS (Advancing knowledge for Long-term Benefits and climate Adaptation ThRough hOlistic climate Services and nature-based Solutions) project. The period considered ranges from 1993 to 2016, corresponding to the ECMWF hindcast period. Total precipitation rate and sea surface temperature (SST) fields from ECMWF and ERA5 are used.  

Focusing on East Africa, we develop a CNN-based prediction of the IOD index for the October-December season at a lead time of three months. The CNN, trained on Indian Ocean SST anomalies using even years and validated on odd years, is based on the architecture proposed by Tao (2024). The resulting hybrid prediction outperforms both the purely AI-based and purely dynamical predictions and leads to improved rainfall skill, particularly over coastal regions of Kenya and Tanzania. 

In addition, experiments assuming perfect knowledge of ENSO and ATL teleconnection indices highlight the potential for further development of CNN-based prediction for June–September rainfall over West Africa, with particularly promising results for the Ghana region. Conversely, limited skill improvements over Southern Africa suggest the need to investigate the role of additional drivers, such as extratropical modes of variability, in future work.