Urban Flood Susceptibility Mapping for Toronto, Canada, Using Supervised Regression and Machine Learning Models

Abstract

Floods are one of the most devastating natural hazards, causing adverse effects on human life, well-being, property, and the environment. The application of five machine-learning techniques in pluvial flood susceptibility mapping was investigated using the case study of two severe storms (2005 and 2013) in Toronto, Canada. Sixteen flood conditioning factors, including elevation, slope, topographic wetness index, stream power index, amount of permeable and impermeable surfaces, and more, were used to evaluate their importance in terms of flooding impacts for the 2005 and 2013 severe storms. Extreme gradient boosting (XGBoost) and an ensemble method are identified as the best models for the tracks of severe storms in 2005 and 2013. The AUROC (Area under the Receiver's Operating Characteristic Curve) analysis shows that precipitation was the most critical variable, followed by groundwater level and distance from sewers, during the two major storm events investigated. However, the flood susceptibility maps are specific and depend on the storm track and intensity-duration characteristics for each significant storm event. Depending on the seasonal groundwater levels and the storm sewer drainage capacity of an area, the system may be overwhelmed, and houses may be flooded if the rainfall intensity and duration exceeds the urban stormwater drainage system capacity. This research provides a foundational understanding of the factors influencing urban flood risk and the statistical models that result from pluvial rainfall events. However, there is a need for more research on rainfall events with different tracks, intensities, and durations to provide reliable ensemble flood susceptibility mapping that could be used to calculate the flood risk for a given area.