A Machine Learning Based Novel Approach of Predicting International Roughness Index(IRI) from Traffic Characteristics using Random Forest Regression

International Roughness Index (IRI) stands as a well-established metric for assessing pavement roughness and overall condition. Predicting IRI is crucial for maintaining pavement infrastructure. In this study, we present a novel approach to predict IRI using Random Forest regression, focusing exclusively on traffic characteristics as predictive variables. Existing studies considered a wide range of factors, including pavement materials, climate, structural attributes, and various pavement distress indicators alongside traffic data where we developed our model using only traffic characteristics. We have used Long-Term Pavement Performance Program (LTPP) dataset for training our models. We have compared our Random forest model with three other models (XGBoost, SVM regression, Gradient Boosting). R squared value and Mean Squared Error (MSE) were taken as performance evaluation metrics. Random forest showed R squared value of 0.70623 and MSE of 8.22 x 10-6 where Gradient Boosting, XGBoost and SVM had R squared value of 0.5737, 0.497, and 0.3455 respectively.We also compared between two hyperparameter tuning methods(Random Search and Grid Search) used in our models and found Random search to perform better. We have also presented a comparative analysis of existing IRI prediction models with our model. Finally we present a SHAP(SHapley Additive exPlanations) analysis to interpret our model and find the contribution of each input feature on our model. We found Annual ESAL (Equivalent Single Axle Load) to be the most dominant factor to predict IRI from traffic characteristics.