Automatic classification of municipal call data for quantitative urban drainage system analysis

Flooding in urban areas can be caused by heavy rainfall, improper planning or component failures. Quantification of these various causes to urban flood probability supports prioritisation of flood risk reduction measures. In many cases, a lack of data on flooding incidents impedes quantification of probability and risk. In the proposed paper we use municipal call data that describe citizens’ observations of urban flooding incidents to quantify flood probabilities in a quantitative fault tree analysis. Given the unstructured nature of call information, calls are first assigned to classes that correspond with causes of flooding as represented in the fault tree. Since manual classification of calls is very timeconsuming, pattern recognition routines are used to automatically classify the call data. The aim of this study is to assess whether by automatic classification of citizen’s calls, sufficient accuracy can be obtained to allow for use of the results in quantitative risk analysis. This is illustrated by application of automatic classification results in a quantitative fault tree analysis for urban flooding, for two cases with datasets of approximately 6000 calls. The results show that straightforward automatic classification routines can reach error rates below 20%. Largest classification errors occur for small classes, where few data are available to train the classifiers. Automatic classification errors lead to small deviations in the outcomes of quantitative fault tree analysis. Still, conclusions about the ranking of contributions to urban flooding that are to be drawn from fault tree analysis, remain intact.Water ManagementCivil Engineering and Geoscience