Speaker
Description
In urban environments, floods result from complex interactions between intense rainfall, soil properties, topography and urbanization patterns. This study investigates the influence of these factors on the modelling and mapping of flood susceptibility using machine learning in the Bobo-Dioulasso agglomeration, with the aim of quantifying the relative contribution of each variable to flood-risk prediction. We adopt an approach based on several Machine Learning algorithms, in particular XGBoost, Random Forest (RF) and a Multi-Layer Perceptron (MLP). Model interpretability is ensured through SHAP (SHapley Additive exPlanations) values, which quantify the contribution of each predictor to the model outputs. The results show that flood-controlling variables such as the Topographic Position Index (TPI), elevation, slope, aspect, NDVI, soil type and drainage density emerge as the most influential factors in Bobo-Dioulasso. Areas combining low-permeability soils, concave topographic forms, high drainage density and strongly impervious surfaces exhibit the highest flood susceptibility. These findings enable the production of physically interpretable flood-susceptibility maps, which are useful for effective flood-risk management and urban planning. This study highlights the central role of soils in the mechanisms governing urban flooding and underlines their importance in susceptibility modelling. Future work will focus on hydrological modelling of surface runoff in other cities of Burkina Faso.
Keywords: urban flooding, susceptibility, soil properties, Machine Learning, SHAP values, Bobo-Dioulasso.