Urban pluvial flooding is one of the most costly natural hazards worldwide. Risks of flooding are expected to increase in the future due to global warming and urbanization. The complexity of the involved processes and the lack of long-term field observations means that many crucial aspects related to urban flood risks still remain poorly understood. In this paper, the possibility to gain new insight into urban pluvial flooding using citizen flood observations is explored. Using a ten-year dataset of radar rainfall maps and 70,000 citizen flood reports for the city of Rotterdam, we derive critical thresholds beyond which urban pluvial flooding is likely to occur. Three binary decision trees are trained for predicting flood occurrences based on peak rainfall intensities across different temporal scales. Results show that the decision trees correctly predict 37%-52% of all flood occurrences and 95%-97% of all non-flood occurrences, which is a fair performance given the uncertainties associated with citizen data. More importantly, all models agree on which rainfall features are the most important for predicting flooding, reaching optimal performance whenever short- and long-duration rainfall peak intensities are combined together to make a prediction. Additional feature selection using principal component analysis shows that further improvement is possible when critical rainfall thresholds are calculated using a linear combination of peak rainfall intensities across multiple temporal scales. The encouraging results suggest that citizen observatories, although prone to larger errors and uncertainties, constitute a valuable alternative source of information for gaining insight into urban pluvial flooding.
Keywords: Citizen flood observations; Critical rainfall thresholds; Decision tree learning; Urban pluvial flooding; Weather radar.
Copyright © 2019 Elsevier B.V. All rights reserved.