Predicting Roadway Flood Severity Based on Waze Traffic Alerts Using Machine Learning

With more frequent intensive precipitation events, aging infrastructure, and increasing imperviousness, urban waterlogging is becoming more frequent due to insufficient drainage capacity. One important adverse impact of intensive rainfalls is traffic disruption due to roadway pluvial flash flooding (PFF) and lane submersion. Predicting PFF severity is crucial to facilitate road safety management, increase travelers' awareness, and enable decision-makers to take proactive actions to address flood conditions. In highly urbanized areas, flood formation is a highly complex and uncertain process due to the lack of information about underground and overland flow interaction. In addition, the impacts of flooding on roadway mobility depend on numerous temporal and spatial variables. Despite the availability of many physics-based hydrodynamic, hydrologic, and empirical models to simulate PFF, their application on the highly localized scales at which traffic disruption occurs remains a challenge due to high computational time and lack of available data. This study detects roadway PFF based on traffic data obtained from Waze, a crowd-sourced navigation app. A hybrid model is developed that combines a hierarchical filling-spilling direct rainfall model with machine learning (ML) algorithms to predict PFF events. The model is applied to a case study of roadways in Dallas-Forth Worth, Texas. High-resolution precipitation data obtained from the Collaborative Adaptive Sensing of the Atmosphere (CASA) radar network and a 1-meter resolution Digital Elevation Model (DEM) are utilized in the rainfall model. The model output and rainfall, topography, and traffic characteristics are then used as inputs to the machine learning algorithms that predict the occurrence of PFF events on an intersection scale. The performance of multiple ML algorithms, including Logistic Regression, Naïve Bayes classifier, Optimal Bayes Classifier, Support Vector Machine, and Multilayer Perceptron Neural Networks, are compared. Results of the study will be presented at the conference.