Predicting seasonal rainfall in East Africa through a teleconnection-based subsampling informed by AI models

The East African region has increasingly experienced periods of extreme precipitation and drought, indicating changes in its climatological seasonal precipitation cycle. Enhancing the seasonal prediction of precipitation is therefore of critical importance for climate change adaptation. Currently, state-of-the-art global climate models demonstrate good skill in forecasting precipitation during the October-December (OND) season, commonly referred to as the short rains. This study aims to further improve OND precipitation forecasts through a hybrid statistical-dynamical approach that integrates artificial intelligence with teleconnection-based subsampling. The focus is placed on the El Niño Southern Oscillation (ENSO), a primary driver of East Africa rainfall variability.

Specifically, we start from the synthetic ENSO forecast developed by Patil K. R., (2023), which is based on a convolutional neural network trained on observational data. This forecast is used to guide the subselection of ensemble members from five global seasonal forecast models provided by the Copernicus Climate Change Service (C3S) -namely ECMWF, CMCC, UKMO, DWD and Meteo France- for the 1993-2016 period. The subselection process involves comparing the Nino 3.4 index time series of each member from each C3S model's ensemble with the synthetic ENSO forecast. The members whose ENSO evolution most closely matches the synthetic forecast, measured using the Euclidean distance metric, are retained. These selected members form a new ensemble, intended as a new representation of the climate system, which leads to improved prediction skill for OND precipitation when compared to observational data.

Future developments of this work could involve expanding the spatial domain to the entire Sub-Saharan Africa, applying the method to the less predictable March-May (MAM) season and evaluating additional teleconnections such as the Indian Ocean Dipole.