Flooding caused by heavy rain is one of the most frequent and damaging natural hazards facing U.S. cities today. Unlike fluvial floods that result from rivers overflowing into floodplains, pluvial floods occur when stormwater systems are overwhelmed by intense rainfall. These events can inundate streets, basements, and critical facilities, often in places far from rivers. As climate change intensifies extreme rainfall and as urban areas expand, pluvial floods are expected to become even more costly and disruptive. This NSF-funded project aims to better understand how pluvial floods form, how urban drainage infrastructure influences their severity, and how modern modeling techniques, including machine learning, can make flood prediction more accurate, computationally efficient, and more useful for decision makers. The research team worked in two highly flood-prone regions, specifically, Houston, Texas and the Red Run watershed near Detroit, Michigan, while also developing tools applicable across the entire United States.

In terms of intellectual merit, the project advanced flood science in several key ways:

1. Understanding rainfall and uncertainty
   The team showed that observed rainfall data from weather radar can be used to generate various scenarios preserving essential storm features to simulate a wide range of realistic rainfall scenarios and capture the uncertainty associated with storm forecasts.
2. Revealing hidden flood pathways
   Using high-resolution LiDAR maps, building footprints, transportation networks, and storm sewer data, the project team built detailed computer models of study urban watersheds. They found that drainage infrastructure such as pipes, culverts, and outfalls does not always reduce flooding. In some cases, the system can actually worsen floods by moving water into areas where sewer backflow occurs. This mechanism, termed a fluvio-pluvial flood, highlights the importance of treating stormwater systems as interconnected networks rather than isolated components.
3. Developing advanced modeling frameworks
   The project combined three approaches: (a) high-fidelity process-based models that capture physical details of water flow; (b) reduced-order models that run quickly by mimicking complex simulations of high-fidelity models; and (c) machine learning models that improve accuracy by correcting biases of flood predictions. This overall hybrid framework enabled both precision and speed, making real-time or large-scale flood forecasting possible.
4. Real-time forecasting and monitoring design
   The team ran hundreds of simulated storms for Detroit’s 2014 flood, generating ensembles that made it possible to design optimized networks of water-level sensor locations. Additionally, machine learning models were trained to provide near-instant estimates of flood inundation extent, paving the way for real-time forecasting tools.

Collectively, these outcomes advanced fundamental understanding of rainfall-runoff processes leading to urban pluvial floods and their interactions with infrastructure, they also improved flood predictive modeling, while creating novel computational tools for the scientific and engineering community.

The broader impacts findings have broad significance for society, policy, and education:

1. Revealing hidden risks in insurance and planning
   By analyzing decades of U.S. flood insurance claims, the team found that 87% of claims outside the official 100-year floodplains were caused by pluvial floods. This highlights a major blind spot in current risk management practices and shows that millions of properties face greater risk than is reflected in existing maps or policies.
2. Informing infrastructure and policy
   The research revealed that outdated stormwater design guidelines and lack of a systems-level perspective are contributing to worsening flood impacts. Publications in Nature Cities and other high-impact outlets shared these insights with engineers, policymakers, and the public. The work underscores the need for integrated stormwater management strategies.
3. Education and workforce development
   The project trained three doctoral students, a postdoctoral scientist, and an undergraduate researcher in advanced hydrologic modeling, geospatial analysis, and machine learning. These individuals presented results at national conferences, published papers in leading journals, and gained professional development experiences in teaching, mentoring, and outreach. The project therefore contributed to the next generation of scientists and engineers trained to address pressing challenges of urban flood resilience.
4. Public outreach and communication3
   The team engaged in a broad dissemination, including interviews with major news outlets, invited lectures internationally, and public-facing articles such as those in The Conversation and the University of Michigan News. These efforts translated complex science into accessible language for the public, raising awareness of the risks and possible solutions to urban flooding.

Last Modified: 08/25/2025
Modified by: Valeriy Y Ivanov