Enhancing pavement temperature prediction using eXtreme gradient boosting transfer learning for data-scarce scenarios

Accurate prediction of pavement temperature profiles is crucial for proactive maintenance and effective infrastructure management. However, traditional machine learning models often require large, location-specific datasets that are costly and time-consuming to collect, limiting their practical application in real-world scenarios. To address this challenge, we propose a novel transfer learning framework based on eXtreme Gradient Boosting (XGB-T), designed to enhance prediction accuracy under data-scarce conditions. The approach involves pre-training the model on a comprehensive dataset from central Taiwan to capture pavement temperature profile, followed by fine-tuning using a minimal dataset from a geographically distinct target site in southern Taiwan. This two-stage process allows the model to transfer learned knowledge to a new domain with minimal additional data. The performance of XGB-T was rigorously evaluated against a standard eXtreme Gradient Boosting (XGB) model trained from scratch and several alternative transfers learning models, including Light Gradient Boosting Machine, Random Forest, Adaptive Boosting, Long Short-Term Memory, and Gated Recurrent Unit. Results show that XGB-T significantly outperforms the baseline XGB model, achieving a 58.96% reduction in Mean Absolute Error, a 56.77% decrease in Mean Absolute Percentage Error, and a 4.08% improvement in the coefficient of determination (from 0.942 to 0.987), averaged across all depths in the testing phase. These findings validate XGB-T as a highly effective and resource-efficient solution for pavement temperature forecasting, significantly reducing data collection burden and cost while maintaining high predictive accuracy in diverse, data-limited environments. This robust framework supports timely decision-making, improves infrastructure maintenance strategies and ultimately enhances road safety.