Heavy rainstorms are becoming more common in many cities. When rain falls faster than storm drains can carry it away, water builds up on streets and neighborhoods - causing flooding. These floods can disrupt traffic, damage property, and pose safety risks. To prevent or reduce flooding, city engineers need accurate, real-time information about how much water is flowing through stormwater pipes and where bottlenecks are likely to occur. However, installing and maintaining a dense network of flow sensors across an entire city is very expensive and time-consuming. Cities often have to work with limited budgets, limited staff, and aging infrastructure. This raises a key challenge: how can we effectively monitor and predict flooding using only a few strategically placed sensors?
Our group at UMD addresses this challenge by developing a data-driven sparse sensing (DSS) framework that identifies the best places to install a small number of sensors in a stormwater network. We tested this approach using the Woodland Avenue catchment in Duluth as a case study. First, we obtained an urban hydrologic & hydraulic model for the catchment from Bolten & Menk, and we used it to simulate how stormwater flows through the drainage network under many different rainfall conditions. These simulations produced a large training dataset showing how water flow rates change throughout the system during storms. Next, our DSS framework analyzed this dataset to determine where sensors would provide the most useful information. It uses advanced mathematical tools, specifically singular value decomposition (SVD), to simplify complex flow patterns and QR factorization to select the most informative sensor locations.
When we tested the method, we found that just three sensors - carefully placed among 77 possible locations - could accurately reconstruct flow patterns across the entire network. The system also proved to be robust, meaning it still worked well even when there were small errors or uncertainties in the sensor readings. Its ability to tolerate sensor failures depended on where the sensors were located, but overall performance improved as more sensors were added.
In short, this framework helps cities get the most value out of limited monitoring resources. By combining engineering knowledge with modern data science, it allows for high-accuracy flow estimation with only a few well-placed sensors. This project will continue. We will integrate DSS with predictive models and real-time control systems to support flood early warnings and smarter stormwater management, helping cities like Duluth become more resilient to extreme weather.