A Machine Learning Approach to Flood Prediction

Abstract

Climate change, driven by both natural processes and human activities, has significantly disrupted living
conditions across many countries. Among its most devastating effects is flooding, which impacts millions of
people globally. Predicting the timing and severity of future floods remains a major challenge. This study adopts
a data-driven methodology, employing machine learning techniques to forecast both the location and magnitude
of floods based on historical flood data from Africa. We also investigate the most appropriate probability
distribution models for recorded precipitation levels. Our findings indicate that, although Africa is a
geographically distinct region that has received limited attention in the literature, its rainfall patterns can be
effectively modeled using well-established probability distributions. Additionally, we identify the weeks with
the highest and lowest rainfall as significant risk factors among various predictors of flooding. Our analysis
further demonstrates that the accuracy of flood predictions is highly dependent on the choice of machine
learning algorithm; with the optimal model, we achieve a prediction accuracy of approximately 85% for flood
occurrence in targeted areas. These findings suggest that while certain flood predictors in Africa align with those
commonly observed in other regions, region-specific factors must still be considered