A machine learning approach to evaluate the spatial variability of New York City's 311 street flooding complaints

Urbanization, accompanied by the creation of roads, pavements, and sidewalks creates an environment where there is limited infiltration capacity, leaving metropolitan areas especially vulnerable during intense rain events. Furthermore, within an urban setting, there is spatial variability, as certain areas, owing to location, topography, land feature conditions, population and physical attributes or precipitation patterns, are more prone to flood damages. To detect neighborhoods with increased flood risk, crowdsourced data, which is the consolidation of eyewitness accounts, affords particular value. With an intent to understand how factors affect the spatial variability of street flooding, the Random Forest regression machine learning algorithm is employed, where the 311 street flooding reports of New York City (NYC) serve as the response, while the explanatory variables include topographic and land feature, physical and population dynamics, locational, infrastructural, and climatic influences. This study also analyzes socio-economic variables as predictors, as to allow for better insight into potential biases within the NYC 311 crowdsourced platform. It is found that catch basin complaints have overwhelmingly the greatest predictor importance, at 41%, almost sixfold higher than that of the second highest ranked predictor, slope, at 6.7%. Thus, NYC has an apparent issue with debris blocking the basins, and this may be remediated by increased cleaning efforts or public awareness to maintain clear streets, particularly during forecasted rain events. Furthermore, more than a third of the top predictors are land feature and topographical conditions, with building characteristics dominating the category. Often excluded in urban flood models, building effects, with a combined total importance of 11.7%, have greater significance than commonly considered flooding factors, such as percent impervious cover or elevation. Another major finding is the significance of the ‘commuters who drive alone’ variable, which alerts to the prospect of more reports being filed by those more affected by street flooding, as opposed to reflecting the actual occurrence of flooding (more reports being filed by those who drive on flooded roads versus those who do not). Overall, the leading contribution of this study is the identification of the top flooding factors in NYC, along with the presentation of their specific impacts towards street flooding variability among zip codes.