Abstract

Floods and droughts are among the most devastating and costly disasters in Africa, damaging food, infrastructure, and livelihoods. Current forecasting and early warning systems for floods and droughts in Africa are limited in terms of accuracy and precision, fail to assess uncertainty, and offer minimal guidance for decision-making. Exacerbated by the effects of climate variability and climate change, floods and droughts are occurring more frequently and more severely, leaving current systems unable to keep up with the impacts of these disasters. This work proposes and explores the capability of machine learning approaches to improve the accuracy, flexibility, and interpretability of predictions of floods and droughts in Africa.
The approach leverages climate and hydrological data from Climate Research Unit (CRU), NASA - Earth Observing System Data and Information System (EOSDIS) and African Centre of Meteorological Applications for Development (ACMAD) datasets. The dataset covers 1980-2025 for five countries, that is, Nigeria, Kenya, Ethiopia, South Africa, and Ghana. The data includes Rainfall, Temperature, Soil moisture, River Discharge, and vegetation indices. A combination of Bayesian inference and Machine learning algorithms (Random Forests, Gradient Boosting, Neural Networks) is used to account for uncertainties, and non-linear relationships. The models are evaluated using cross validation, RMSE, MAE and AUC.
Results show that the hybrid Bayesian-machine learning models outperform regression benchmarks by statistically significant margins. The performance gain in the case of flood forecasting was 19%, 21% and 18.7% in Nigeria, Ghana, and Kenya respectively. For drought prediction, the Bayesian models led to gains of 22% in MAE and to increases in AUC metrics in Ethiopia, South Africa of 23% and 15% respectively. The models captured 46–49% of flood occurrence with rainfall variability; 37–41% of drought severity with soil moisture; 12–15% of drought detection with Normalised Difference Vegetation Index (NDVI).
In conclusion, the fusion of Bayesian and machine learning techniques advances disaster prediction, surveillance and early warning systems in Africa. Areas for policy action comprise including hybrid predictive models into national disaster and early warning systems, prioritising climate data infrastructures, and capacity building. Our work holds the potential of mitigating vulnerability and thus avoiding livelihood losses in climate sensitive areas.