Institute for Transportation

About the research

Road weather information systems (RWIS), in both stationary and mobile forms, have become increasingly popular in recent decades for their ability to collect and disseminate road weather and surface data. In addition to meteorological measurements, highway agencies rely heavily on RWIS imagery data to guide winter road maintenance (WRM) operations. However, the analysis of imagery data is still performed manually by trained personnel. Moreover, the limited number of stationary RWIS stations and the infrequent deployment of mobile RWIS units result in significant spatial gaps along the highway network. In our previous project, we developed methodologies based on convolutional neural networks (CNNs) to automatically recognize road surface conditions (RSC) from dash camera imagery and employed regression kriging (RK) to estimate RSC in unmonitored areas using limited point measurements. These methods demonstrated feasibility and robustness in real-world case studies. Building on these efforts, this project aimed to further advance CNN development specifically for stationary RWIS imagery and assess its reliability using explainable artificial intelligence (XAI) techniques, including SHapley Additive exPlanations (SHAP) and class activation map (CAM)-based methods. Additionally, to automatically estimate snow coverage ratios from stationary RWIS imagery, two distinct deep learning-based computer vision techniques, pix-to-pix generative adversarial network and semantic segmentation, were employed. Furthermore, the RK method was revisited to better accommodate a wider range of weather events while considering their variability, and the potential monetary benefits of this approach were also explored. To address the limitations of RK in handling categorical variables, a novel geostatistical method, namely nested indicator kriging (NIK), was developed to interpolate RSC in unmonitored areas directly using CNN classification results. These methods were evaluated using data from two major highways, Interstate 35 and Interstate 80 in Iowa, US, spanning a five-year period and encompassing over 20,000 images. The results demonstrated high accuracy and reliability. Additionally, a web application was developed to integrate these methods, offering real-time monitoring, estimation, and historical data archiving. This project equips decision-makers with a powerful tool to implement WRM activities more swiftly, efficiently, and cost-effectively, ultimately promoting a safer, more mobile, and sustainable winter transportation system.