A Semisupervised Learning-Based International Roughness Index Estimation Framework Using Smartphone-Collected Public Vehicle Response

Abstract Smartphone-collected vehicle response data have been used to estimate the International Roughness Index (IRI). While data-driven methods were approved to be effective, a model must be trained by substantial labeled data (ground truth IRI), which are scarce due to data collection costs. Yet, a wealth of unlabeled response data could be crowdsourced from road users, presenting an opportunity to employ semisupervised learning (SSL) algorithms to leverage these data for IRI estimation. This paper introduces an SSL framework to enhance a deep learning model for IRI estimation. The results demonstrate that the SSL-trained model reduces the estimation root mean square error (RMSE) compared to a fully supervised model trained solely on labeled data. These findings suggest the potential for applying SSL as a robust and cost-effective approach for continuous pavement condition monitoring.