Integrating deep learning and ensemble modelling for urban flood susceptibility mapping in eThekwini municipality, South Africa

Flood susceptibility mapping is vital for proactive disaster management. This study proposes a novel hybrid framework combining a Convolutional Neural Network (CNN) with an ensemble of boosting algorithms, Extreme Gradient Boosting (XGBoost), Light Gradient Boosting (LightGBM) and Gradient Boosting (GB) to assess flood susceptibility in eThekwini municipality, South Africa. Model training and validation utilised 19 topographical, environmental and hydrological conditioning factors alongside a spatial inventory of 135 historical flood events. The prop[osed hybrid model demonstrated exceptional predictive capability outperforming individual benchmark algorithms with an Area Under the Receiver Operating Characteristic curve (ROC-AUC) of 0.991. Factor importance analysis identified elevation, historical rainfall, and the terrain ruggedness index as the primary drivers of flood susceptibility. Spatial susceptibility classification revealed that critical risk is concentrated within the municipality; 11.6 % of the total area falls under extreme susceptibility, 4.0 % under very high and 24.5 % under high susceptibility. Regionally, the central zone exhibits he highest vulnerability, with 26% of its area classified as extreme risk. By pinpointinghyper-local exposure patterns, this high precision hybrid approach provides actionable intelligence for spatial planning, enabling targetted infrastructure interventions and climate resilience strategies in high-risk zones.