Structured exploration of machine learning model complexity for spatio-temporal forecasting of urban flooding

Urban flooding disrupts socio-economic systems and endangers lives, necessitating the employment of prediction tools. This study applies spectral clustering to delineate flood-prone zones in New York City (NYC), followed by an evaluation of statistical and neural network models, including feed-forward and graph neural networks. Among these, the graph wavenet (GWN) excels due to its proficiency in capturing dynamic spatio-temporal relationships, improving mean R2 by 0.15 and achieving up to R2 of 0.72 in certain areas. The study emphasizes that augmenting spatio-temporal components and adopting graph-based architectures enhances predictive accuracy, particularly in data-scarce settings. By combining clustering and advanced modeling, this research provides novel insights into urban flood prediction. These findings equip NYC urban planners and emergency responders with a valuable framework to mitigate the socio-economic impacts of flooding and improve response strategies.