Predicting IRI Using Machine Learning Techniques

• Published in: International journal of pavement research & technology 2023-01, Vol.16 (1), p.128-137
• Main Authors: Sharma, Ankit, Sachdeva, S. N., Aggarwal, Praveen
• Format: Article
• Language: English
• Subjects: Building Construction and Design; Civil Engineering; Deep learning; Engineering; Generalized linear models; Machine learning; Mean square errors; Mean square values; Original Research Paper; Pavements; Performance evaluation; Regression models; Rehabilitation; Road maintenance; Roads & highways; Root-mean-square errors; Statistical analysis; Statistical models; Structural Materials; Transportation networks
• ISSN: 1996-6814; 1997-1400
• DOI: 10.1007/s42947-021-00119-w

The behaviour of pavement structure to varying degrees of loads, climate conditions, traffic, drainage conditions and dimensions of road cause difficulty in deciding the maintenance/rehabilitation task on the pavement. International Roughness Index (IRI) is the most commonly used criteria for evaluating pavement performance and determining maintenance/rehabilitation requirements of the pavements. In a road network comprising hundreds of km of the road, it becomes difficult to accurately predict the road’s IRI. The data have been taken from a public database of roads, i.e. long-term pavement performance. In this study, machine learning models have been studied to understand/analyze the IRI of roads. The evaluation/performance of regression models has been done on the basis of commonly used statistical measures. Gradient boosting machine (GBM) model performed best on the test as well as train data set out of five used models, namely GBM, deep learning, extremely random forest, distributed random forest, and generalized linear model. Performance of GBM in the testing dataset had root mean square error (RMSE = 0.176003), root mean square log error (RMSLE = 0.074924), mean average error (MAE = 0.126345), mean square error (MSE = 0.030977), which was minimum of five models, and R 2 (0.86572) which was maximum.