Forecasting agricultural drought using a multi-modal ensemble approach

ABSTRACT Frequent droughts cause severe socio-economic and ecological impacts in Southern Africa, where agriculture and water security are crucial. This study aimed at developing and evaluating a model employing a multi-modal ensemble approach that integrates random forest, feedforward neural networks and transformer models for forecasting drought conditions in Southern Africa, and to assess model effectiveness in predicting the standardised precipitation-evapotranspiration index (SPEI). The methodology combines innovative machine learning techniques using Bayesian model averaging to enhance predictive reliability. This study utilises SPEI indices (SPEI1, SPEI3, and SPEI6) across eight climatic zones between 1972 and 2022. SPEI6 models consistently outperformed SPEI3 and SPEI1 models. Stratified K-fold cross-validation is used to address class imbalances. The ensemble model demonstrates superior performance across all evaluation metrics when compared to the individual algorithms. The addition of Platt scaling to the ensemble model marks a key advancement. It enables more accurate probability outputs and enhances confidence clarity. Particularly strong performance is observed in one zone with the SPEI6 model. It demonstrates an accuracy of 0.99, a Nash–Sutcliffe Efficiency (NSE) of 0.9, and a correlation coefficient (R) of 0.98. The model reports mean absolute error (MAE) 0.02, mean squared error (MSE) 0.02, and mean bias error (MBE) 0.0. It also delivers strong discrimination, with a precision–recall area under the curve (PR AUC) of 0.9462 and a receiver operating characteristic area under the curve (ROC AUC) of 0.9754. Findings emphasise the potential of machine learning to strengthen climate resilience in rainfed agriculture throughout Southern Africa.