Urban flood susceptibility mapping using supervised regression and machine learning models in Toronto, Canada

Floods are one of the most devastating and damaging natural hazards, with adverse effects on human life and well-being, property and the environment. A severe urban flash flood event in Toronto, Canada in 2013 caused insured damages of $943 million. Considering this, the development of flood prediction models to identify flood-prone areas is important for decision-makers. Hydrologic and hydraulic models can be used to guide in the implementation of flood prevention and mitigation measures but their use is considered problematic due to the various types of required hydro-geomorphological data-sets, intensive computation time, and in-depth knowledge required regarding various hydrological parameters. One of the key alternatives is to use data-driven, statistical machine learning models to identify flood-prone areas.

The aim of the current research is to utilize and compare the performance of an array of simplistic to advanced machine learning approaches for urban flood susceptibility mapping. Seven different machine learning techniques ranging from the simple ones like logistic regression to advanced methods such as Artificial Neural Network (ANN) were used to produce flood susceptibility maps for a case study in Toronto. For the analysis, 15 conditioning factors were considered, including elevation, slope, plan curvature, topographic wetness index, stream power index, permeable and impermeable surfaces, geology, land-use, proximity to natural drainage, sewers and stormwater ponds, population density, precipitation, normalized vegetation index, drainage density and the type of sewer system. The models were evaluated and validated using statistical measures such as Receiver's Operating Curve, accuracy, root-mean-squared error and Kappa index. The accuracy assessment of all seven models showed fair prediction accuracy ranging from 70 to 75%. The results of the current research are being used to identify flood susceptible areas, where targeted measures can be implemented to minimize the adverse effects of flooding and provide fast emergency response services.