Classification machine learning models for urban flood hazard mapping: case study of Zaio, NE Morocco

Floods have become increasingly frequent and devastating in recent decades, posing unignorable risks as highly destructive natural hazards. To effectively manage and mitigate these risks, accurate flood hazard mapping is crucial. Machine learning models have emerged as valuable approaches for flood hazard assessment. In this study, six machine learning (ML) models, including Maximum Entropy, Support Vector Machine, Extreme Gradient Boosting (XGB), Random Forest (RF), multi-layer perceptron, and Naive Bayes, were utilized to evaluate urban flood hazard in Zaio, NE Morocco, and estimate the flood presence extent. Nine flood conditioning factors were used as input variables. Historical flood presence and absence data were employed for models training and testing, incorporating 663 flood presence and absence locations dating back to past flood events. Performance evaluation metrics such as Kappa statistic, accuracy, sensitivity, specificity, and area under the curve (AUC) were calculated for each model. RF (AUC = 0.92) and XGB (AUC = 0.9) models showed excellent classification capabilities, surpassing the performance of the other models, while the other models exhibited lower but recognizable performances. Additionally, the hazard presence extent maps generated by the ML models exhibited a decent alignment with a historical flood event maps created by the hydrodynamic and the cellular automata models. The results imply that ML models offer effective solutions for mapping urban flood hazards. The innovative integration of various ensemble and single ML models demonstrates their potential in urban flood hazard susceptibility and extent mapping, effectively surpassing the limitations associated with limited availability of hydrologic/hydraulic data and computational burden. These mapped results can be instrumental for local authorities in shaping mitigation strategies in the city of Zaio.