Mapping stormwater inundation potential in data-scarce urban areas: A machine learning approach for Kolkata, India

Kolkata is a major fastest-growing city in India facing severe stormwater inundation yearly. However, it is difficult to identify stormwater inundation areas due to the unavailability of past inundation data on a large scale. This study focused on developing a methodology using Sentinel image and Machine Learning (ML) techniques for identifying inundation potential areas to minimize past inundation data availability problems. Stormwater inundation potential areas of the Kolkata Municipal Corporation (KMC) area are identified from Sentinel data derived inundation area locations and eight influencing factors using three ensemble Machine Learning (ML) models i.e. Random Forest (RF), Bagged CART and Extreme Gradient Boosting (EGB). The validation of these identified inundation potential areas is performed based on a separate stormwater event and field verification of inundated areas. The results show that the EGB (87.34 %) outperforms RF (87.09 %) and Bagged CART (84.46 %) due to its advantage of adding a large number of weak models to improve the result. EGB indicates 30.1 % of areas with high to very high inundation potential, especially in the western part adjacent to Hooghly River, the Eastern part adjacent to East Kolkata Wetland (EKW) and some areas of northern Kolkata. This higher potentiality is mainly due to topographic conditions, built-up density and age-old drainage network. This identified inundation potential areas will help the decision-makers to identify high-priority zones for developing plans to minimize this issue. This developed methodology will also help to identify the inundation potentiality of other inundation data-scare cities of the World.