Towards a Road Weather Impacts Prediction System: Data-Driven Modeling and Prediction of Flash Flood Impacts

This work presents a prototype of a rapidly updating, decision-support system based on potentially impactful weather information, coupled with a routing decision support tool to aid the Missouri Department of Transportation (MODOT) in effectively responding to road-weather hazards. The presented work will lay the foundation for developing a road weather impacts prediction system (RWIPS). While the main focus of this project's initial stage is on road impacts from flash flooding, the potential for implementing similar models for additional and distinct road hazards (such as winter weather) is the ultimate goal.

Based on an extensive record of road closures due to flooding in the state of Missouri (~13,000 events between 2017 and 2023), an event-based machine learning-enabling data set has been constructed by incorporating diverse data sources such as MRMS precipitation estimates and accumulations, FLASH rainfall average recurrence intervals and streamflow estimates, as well as geomorphological and vulnerability information. A systematic, data-driven approach was used to define event scales (spatial and temporal) for matching road closure reports to model and product outputs.

A pilot study demonstrating the feasibility of utilizing compound transformer-based, computer vision-oriented architectures was performed (e.g. VideoMAE, ViT), which is being designed to provide state-wide probabilistic estimates for flash flood-impacted roads in real time. Detailed results on both the construction of our data set, as well as model implementation will be presented and discussed in depth. Prospective insights towards real-time implementation and the incorporation of additional hazards will be discussed.