Machine Learning Framework for Flood Susceptibility Modeling in a Fast-Growing Urban City of Southern India

Flooding in urban areas often results severe loss of life and property and has many negative socio-economic impacts. Therefore, identifying the flood prone areas is necessary for future flood hazard mitigation, early warning, and land use planning for infrastructure developments in urban areas. In this study, flood susceptibility modeling is carried out for Kozhikode urban and per-urban area, which is severely affected by 2018 Kerala flood. To begin with, a flood inventory map is prepared with 307 flood location points marked immediately after 2018 flood. Thereafter, the inventory is randomly classified into 70% for model training and remaining 30% for model testing. In addition, twelve independent variables such as land use/land cover, soil texture, lithology, elevation, slope angle, slope aspect, valley depth, topographical wetness index, profile curvature, plan curvature, convergence index, and channel network base level were prepared and used. Subsequently, final modeling is carried out using these flood conditioning factors and flood inventory locations using machine learning random forest method. The result shows that ~ 13.78% of the study area is very highly susceptible to the occurrence of flood. The predicted model shows 85.2% accuracy (ROC-AUC) in training phase and 78.5% in testing phase. Therefore, the model is trustworthy and can be used for future hazard mitigation and land use planning in Kozhikode urban and per-urban area.