Full Abstract

Urban flooding due to extreme precipitation poses an increasing threat to cities, necessitating improved design guidelines and best practices for urban stormwater-drainage infrastructure and accurate, timely predictions of flood characteristics. In practice, however, these systems do not always meet their assumed performance expectations. Current guidelines focus on understanding the performance of specific infrastructure elements — such as storm drains, culverts and pipelines — in isolation, rather than as part of a broader urban landscape. This fails to account for the complexity of flood ‘connectivity’: how flood flows can interact in subterranean pipelines, open channels, and over natural and human-made surfaces. Starting with analysis of unexpected flooding in southeast Michigan, USA, in 2014, we find that human-engineered infrastructure can introduce additional stormwater flow connectivity in urban environments that may increase flood risk (Fig. 1).          

Paradoxically, design errors can reverse stormwater network functionality in the real world and lead, for instance, to flooding even in the absence of local rainfall. To advance predictive capabilities, we combine high-fidelity hydrodynamic urban flood model (tRIBS-Urban), stochastic field decomposition methods, and machine learning tools to develop a novel approach optimizing flood sensor placement and reconstructing inundation maps from (synthetic) observations at a few locations. The approach rapidly produces real-time forecasts for 1-12 hour lead times. We further present a general framework in development for real-time flood-forecasting in urban environments with lead time of several days.