Machine Learning Architectures for Modelling International Roughness in Cold Region Pavements

One of the most commonly used pavement performance indicators is the International Roughness Index (IRI). Currently used IRI models are often developed using regression analysis with little emphasis on climate. Recent studies have started using Machine Learning (ML) for IRI model development; however, the studies' scope is limited and often restricted to algorithms such as neural networks. Additionally, a systematic comparison between different ML algorithms in modelling IRI cannot be found in the literature. This study develops and systematically compares IRI models using regression analysis and ML methods. The economic and environmental implications of using site-specific models over general models are also examined in this study. This study also analyzes the impacts of climate change on pavement roughness for pavements with different subgrade soil types. This study's results support the use of ML, especially gradient-boosted ensemble algorithms, in developing IRI models as they have superior predicting capabilities and can provide much more value than traditional regression methods, such as regression analysis. The results also found that ML was able to produce meaningful results when regression analysis failed to do so.