The development of a road network flood risk detection model using optimised ensemble learning

Floods are natural phenomena that invade different parts of the globe annually, leaving severe adverse effects on the natural landscape, particularly in areas where humans practice large-scale urban development activities. Inadequate corporate strategic stormwater management or maintenance plans make the issue even worse. Therefore, efforts have been consolidated to confront flood hazards by incorporating precautious and proactive strategies toward better forecasting and preventing flood disasters or even mitigating the damage if it occurs. Further, there is inadequacy in incorporating computational intelligence techniques to detect flood susceptibility in road networks. This paper proposes a data-driven flood risk areas detection model incorporating various ensemble and generic machine learning techniques as well as autoML tools. A systematic and iterative approach is followed to examine different solutions to tackle the data balancing problem, including oversampling and undersampling techniques. Also, hyperparameter optimisation is undertaken using both grid search and genetic algorithms, and the models' performance using each designated strategy is reported and assessed. The empirical results from the experiments indicate that the optimised ensembled Extremely Randomized Trees classifier proves an ability to perform well on the given dataset with an averaged ROC AUC value of 90% and outperforms several state-of-the-art generic, ensembled, and autoML classifiers in all evaluation measures. The proposed framework could be possibly generalised to other road networks in similar topographical and geological regional areas.