Urban Flood Hazard Digital Twin Using a 2-D Flood Model and Machine-Learning Surrogate

Where does rain water collect in a city, and which planning change helps? This project builds a flood-hazard digital twin of south Chennai (24 x 24 km at 60 m) from open satellite data. The physics is the project's own 2-D rain-on- grid flood model (local-inertial scheme, as in LISFLOOD-FP), run for five design storms of 40-240 mm in 6 hours. A LightGBM surrogate learns that model and draws a new hazard map in about a second; it is tested on a storm and on 3 x 3 km tiles it never saw, always next to the classical HAND threshold and a depression (blue-spot) threshold. On the unseen 120 mm storm the surrogate reaches a CSI of 0.611 against 0.292 for HAND. A web application lets planners try pervious surfaces and storm depths and see the residents and facilities in deep water. The maps are relative hazard for planning, not a flood forecast: the surrogate imitates the own model, the elevation model still contains buildings and trees, and there are no drains and no river inflow.

Category: School of Architecture and Planning

Architecture & Planning Project AP06 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.