Predicting Urban Coastal Street Flooding at Hourly Time Scale Using Random Forest And Google WAZE

Coastal cities have been experiencing more frequent flooding due to extreme rainfall and high tide events and this flooding is expected to increase from climate change and sea level rise. Frequent floods in coastal cities disrupt the transportation system causing inconvenience and economic loss. An accurate real time model for predicting street flooding would help to minimize the negative impacts of urban coastal flooding. In this study, we used Random Forest to predict the number of streets flooded in Norfolk, Virginia, USA at an hourly timescale which is a step towards near real time flood prediction. The model was trained and evaluated using environmental data such as rainfall, tide level, wind and groundwater level as input variables and crowdsourced street flood reports as output variable. The street flood reports were made by Google WAZE users from August, 2017 to May, 2018 which provided a uniquely detailed record both temporally and spatially. Within the ten-month record, 82 days had at least one flooded street report. The number of daily flooded streets ranged from 1 to 21 with a maximum of 13 streets flooded in an hour and some of the streets were reported flooded frequently due only to high tide level. We compared our preliminary results to those of a previous study which was done at a daily time step. Both models' normalized RMSE were very similar, however the importance of input variables differed significantly. Although cumulative rainfall was the most important feature in both cases, the average tide level was substantially more important for predicting hourly flooded locations. This suggests that at an hourly time scale shorter-term floods caused by water backing up through the stormwater infrastructures are more accurately predicted than when predictions are made at a daily time scale. However, the model was still unable to predict some of the flood events that occurred during high tide and smaller rainfall events. A longer-term street flooding dataset is expected to improve prediction for flooding caused by less significant events.