Advanced Hybrid CNN-GRU Model for IRI Prediction in Flexible Asphalt Pavements

Abstract

The international roughness index (IRI) is an important parameter for road surface roughness assessment, which is crucial for road safety. This study developed a road roughness prediction model for predicting the IRI of asphalt pavement with deep learning technology. A gate recurrent unit (GRU) and convolutional neural network (CNN) model are integrated, forming a novel hybrid CNN-GRU model for the IRI prediction. The Long-Term Pavement Performance (LTPP) database is adopted to acquire pavement data considering various variables, such as road age, traffic load, pavement performance, and climate condition. The correlation coefficients between these variables were analyzed and generated. The database consisted of 4,782 observations from 975 road sections in the LTPP program. From the data set, 80% of the data were randomly sampled for training the hybrid CNN-GRU model and the other 20% were used for testing. The coefficient of determination R2 for training and testing can reach 0.903 and 0.893, respectively. Compared with the random forest model, the R2 of the CNN-GRU model are 0.108 and 0.130 larger for the training set and the testing set, an increase of 13.58% and 17.04%, respectively. Moreover, the Shapley additive explanations (SHAP) method was employed to assess the significance of variables, revealing that the initial IRI holds the most substantial influence on prediction outcomes. Rutting and transverse cracking also exhibit a considerable impact. The effect of climate parameters on IRI prediction was also investigated. When not considering climate parameters, the R2 for the training set and testing set are 0.893 and 0.886, respectively, which is lower than the results when considering climate parameters. The results indicate that the climatic parameters indeed play a vital role in the prediction model.