|The monitoring of winter road surface conditions (RSCs) is essential to transportation agencies and the traveling public, since the former needs to be aware of the location and severity of existing RSCs in order to effectively maintain safe roadways with minimal environmental impact, while the latter uses RSC information to make informed travel decisions. However, current RSC monitoring practice still relies on methods that are time-consuming, labour-intensive and lacking in objectivity, therefore limiting their ability to provide sufficient spatial and temporal coverage across a road network. This research was motivated by the need for accurate, timely and reliable RSC monitoring for winter maintenance personnel and the travelling public. To achieve this objective, the field performance of a smartphone-based RSC monitoring system was evaluated on a section of Highway 6 in Ontario, Canada during the winter of 2014. A comparison between this system and current monitoring methods indicated that the former was capable of providing reliable results particularly at the maintenance route level; however, classification accuracy was found to vary according to RSC type.
To improve the results produced by the smartphone-based system, this thesis proposes a connected- vehicle (CV) based RSC monitoring system that utilizes Road Weather Information System (RWIS) data in addition to the smartphone-based system’s data. Three techniques in artificial neural networks (ANNs), random trees (RTs), and random forests (RFs) were tested as the underlying models of the CV system, and the results indicated that all three models successfully increased the classification accuracy of the smartphone-based system. RFs were found to provide the most accurate RSC classifications for the standard (three-class) classification scheme while RTs were found to be most accurate when using a more detailed (five-class) classification scheme. Model transferability was also tested using data captured from a different test site during the winter of 2015; and it was found that although the proposed CV system significantly increased the reliability of RSC classifications, the underlying models were non-transferable and would therefore require local calibration before being used at different sites across a road network.