Graph Neural Networks for Spatial-Temporal Modeling of Urban Flooding using IoT-Generated Topology Data

This problem of urban flood is one of the severe issues that ideally must be predicted in real time for efficient management. In this work, GNN-LSTM model with IoT generated topology data is considered for spatial-temporal modeling of urban floods. The suggested model does better than standard methods, it yields a 92.8 % hit rate, LSTM(87.3 %), CNN -based models (84.5%), and even the Decision Tree Regression of 78.9 %. The RMSE is decreased to 0.18 m while the MAE minimum is 0.12 m to show better accuracy in the evaluation of flood levels. The model is able to incorporate spatial relationship from IoT sensors and temporal behavior of the flood pattern. Evaluation of the time complexity of this model shows that GNN-LSTM has moderate training time of 35 minutes and testing time of 22 ms indicating the model's suitability for real-time use. A validation real world achieved 96.4 % of flood alert correct in 0.5 m, early warning and low false alarms (3.6 %) and false neglects (1.8 %). Therefore, the use of GNNs in combination with the actual IoT data in flood prediction improves its effectiveness and reliability as a solution for managing flood risk in urban areas.