Bayesian optimization based random forest and extreme gradient boosting for the pavement density prediction in GPR detection

Abstract

The Ground penetrating radar (GPR) adopts non-destructive and continuous electromagnetic (EM) monitoring technology that is influenced by the dielectric properties of transmission mediums and is of interest in the density detection of asphalt pavement. Prediction of in-place density based on the dielectric properties of pavement materials is a crucial aspect of GPR density detection. This study introduced the random forest (RF) and extreme gradient boosting (XGBoost) approaches that combine the state-of-the-art Bayesian hyper-parameter optimization (BHPO) and 5-fold cross-validation for density prediction. The pavement density data collected from the literature, laboratory, and field tests in the G42S expressway are used for training, testing and validating the models. The input to the RF and XGBoost models was the data for seven material parameters, and the output was the corresponding asphalt mixture or pavement density. The results indicated that BHPO significantly increased the accuracy of density prediction and captured the optimal combination of hyper-parameters for RF and XGBoost models. The XGBoost and RF models were well defined after BHPO and the XGBoost outperformed the RF and other theoretical models. In the validation set, machine learning (ML) algorithm-based density models reduced pavement density prediction error from 3% to about 0.3%. Features importance analysis illustrated that the ratio and density of asphalt to aggregate is the decisive factor for pavement density. The critical factor affecting the density monitoring is the accuracy of the dielectric constant detection using the GPR. The present study implied that the ML algorithms could significantly improve the accuracy of the GPR system. The coming decade promises considerable advances in combining the GPR system and ML.