Urban flood prediction from limited data: a data-model integrated approach

Flooded parking lot of shopping strip mall

Project overview

Increasing resilience to floods under climate change has become one of the major challenges for urban area. Annual losses due to major flood events in the US are projected to increase from $7-9B for the period 1903-2014 to $19B by 2100. Minnesota is expected to receive more precipitation in response to increasing global temperatures. Great amounts of funding as part of climate change resilience are expected to be invested in the improvements of stormwater systems across the state. To obtain a complete understanding of the flooding events and achieve accurate, timely prediction, hydrologists and water managers are challenged by the desire to measure and model at high spatio-temporal resolution. However, time, budget, and technology constraints limit the resolution that can be achieved in practice. Process-based models are limited by high computational and calibration requirements; Machine learning models also suffer from challenges in physical interpretability and transparency and high requirement of training data. Therefore, broad questions facing data-driven flood prediction include: Can we predict urban floods with temporally and spatially sparse data? What times and locations are measurements most informative of urban flood occurrence? And how can we strategically deploy measurements to maximize the effectiveness of flood prediction models?

The overarching goal of this project is to apply data-driven sparse sensing (DSS) methods to develop optimal monitoring strategies for urban flood prediction. This goal will be achieved by

  1. Developing an urban hydrologic model that can simulate flood depths in urban catchments

  2. Identifying optimal monitoring locations and minimum data requirements for accurate flooding prediction. 

This project will leverage computationally-efficient and interpretable linear algebra tools. Specifically, the singular value decomposition (SVD) and matrix QR factorization will be used for dimension reduction and sensor allocation; compressive sensing will be used for flood depth prediction.