Road surface condition (RSC) is an important performance indicator for winter road maintenance personnel to maintain safe driving conditions. This becomes more apparent in inclement weather events where timely clearing of snow is highly prioritized. Considering the vast road networks that need to be covered, many transportation agencies have been using camera images to view real-time RSC directly; however, monitoring conditions via these cameras is still being done manually, thereby hindering its full utilization for optimizing maintenance services. Many studies have attempted to develop a deep-learning-based approach known as convolution neural network (CNN) to automate the process of RSC image classification. When implemented, RSC can be extracted from road images without human involvement. However, efforts made thus far have been focused on rural highways, with performance in the urban context being the least explored. Furthermore, CNN models developed in previous studies have been trained either from scratch or via transfer learning, but only a few studies have investigated transfer learning using a pre-trained RSC model. To address these gaps, an urban RSC model was developed in this study via transfer learning using a pre-trained RSC CNN model. The image dataset used contains 3914 urban images collected in a residential area south of Edmonton, Alberta. With these images, the pre-trained RSC model was fine-tuned via transfer learning and underwent hyperparameter optimization to boost performance further, yielding a high classification accuracy of 98.21% and an F1-score of 98.4%, which surpassed the accuracy of the model trained from scratch.

This study investigates the utility of machine learning (ML) in understanding and mitigating winter road risks. Despite their capability in managing complex data structures, ML models often lack interpretability. We address this issue by integrating Shapley Additive exPlanations (SHAP) with a support vector machine (SVM) model. Utilizing a comprehensive dataset of 231 snowstorm events collected in the city of Edmonton across two winter seasons, the SVM model achieved an accuracy rate of 87.2%. Following model development, a SHAP summary plot was employed to identify the contribution of individual features to collision predictions—an insight not achievable through ML alone. Next, SHAP waterfall plots were used to assess the reliability of individual predictions. The findings enhanced our understanding of the complex SVM model and provided greater insights into the diverse factors affecting winter road safety.

Although road surface friction is considered the most effective performance measure for maintenance operations, it is not commonly used due to the high cost of collection. As a result, most jurisdictions use subjective visual indicators that qualitatively describe the state of the road surface, even though they create measurement inconsistencies and offer less detailed maintenance tracking. For maintenance personnel to transition into using friction, the collection cost must be reduced. This paper attempts to do so by proposing a low-cost, machine learning-based method for predicting road surface friction using dash camera imagery and demonstrates its feasibility through a case study. The dataset used for this project was collected in the City of Edmonton, AB, Canada during its 2021/2022 winter season. Three models were developed using tree-based algorithms, where all three displayed high performance with an average root mean squared error of 0.0796 or 79.3% accuracy based on RMSPE.

Jurisdictions currently provide information on winter road conditions through qualitative descriptors like bare and fully snow-covered. Ideally, these descriptors are meant to warn drivers beforehand about hazardous roads. In practice, however, discerning between safe and unsafe roads is sometimes unclear due to intermediate RSC classes covering too wide a range of conditions. This study aims at solving this safety ambiguity issue by proposing a framework for predicting collision likelihood within a road segment. The proposed framework converts road surface images into friction coefficients, which are then converted into continuous measurements through an interpolator. To find the best-performing interpolator, we evaluated geostatistical, machine learning, and hybrid interpolators. It was found that ordinary kriging had the lowest estimation error and was the least sensitive to changes in distance between measurements. After developing an interpolator, collision likelihood models were developed for segment lengths ranging from 0.5 km to 20 km. We chose the 6.5 km model based on its accuracy and intuitiveness. This model had 76.9% accuracy and included friction and AADT as predictors. It was also estimated that if the proposed framework were implemented in an environment with connected vehicles and intelligent transportation systems, it would offer significant safety improvements.