A GNN-Based Urban Flood Situational Awareness Framework Integrating Radar Rainfall Forecasting and Real-Time Sensor Observations

The increasing frequency of localized heavy rainfall driven by climate change has intensified urban flood damage, demanding situational awareness systems capable of capturing the complete flood lifecycle—from pre-event onset through spatial inundation propagation—in real time. This study presents a GNN-based urban flood situational awareness framework that integrates radar rainfall forecasting and real-time operational sensor observations to simultaneously predict flood onset timing, flow pathways, and inundated areas across a complex urban drainage network.

In the rainfall forecasting component, hazard-triggering rainfall is classified using the morphological characteristics of radar-observed precipitation fields and storm motion vectors, after which a deep learning-based short-range prediction model (KICT-RAIN-AI) generates high-resolution gridded rainfall fields as forecast inputs. Within the GNN model, the urban drainage system is reconstructed as a topological network in which manholes and sewer pipes are represented as nodes and links, respectively. The forecast rainfall is incorporated as dynamic nodes at each time step, enabling the model to learn spatiotemporal correlations between predicted rainfall and real-time sensor measurements, including sewer water levels and road surface inundation depths. Training data are derived from physics-based inundation scenarios generated by a coupled one-dimensional SWMM and two-dimensional GIAM framework, comprising 76 rainfall events and 1,296 time-series inundation maps, followed by domain-adaptive fine-tuning using field sensor records.

A key distinction of the proposed framework lies in its explicit encoding of drainage network topology within the learning architecture, which enables the model to trace inundation propagation pathways along hydraulically connected structures—a capability that conventional grid-based deep learning approaches cannot resolve. The study area is the Gwanak-gu district of Seoul, a topographically vulnerable catchment characterized by a valley-type terrain, high impervious surface ratio, and dense pipe network, where multiple historical flood events including the August 2022 extreme rainfall episode will be used for validation. The proposed system is expected to provide flood onset predictions with a lead time exceeding 30 minutes, supporting more accurate and timely decision-making during extreme weather events. By unifying radar rainfall forecasting, drainage network topology learning, and real-time sensor fusion into a single operational framework, this study provides a critical technological foundation for proactive urban flood management and early warning.

 Acknowledgments: The research for this paper was carried out under the KICT Research Program (Project no. 20260161–001, Development of Digital Urban Flood Control Technology for the Realization of Flood Safety City) funded by the Ministry of Science and ICT.