| description abstract | Current roughness prediction models require extensive input data including pavement distress, climatic, and traffic data, which may be difficult to collect. In addition, these models have geographical limitations; therefore, a significant bias is expected if these models are used without recalibration. Pavement digital images can reflect both surface distresses and other surface irregularities that may affect pavement roughness conditions. In this study, convolutional neural networks (CNNs) were used to classify pavement sections into different roughness categories and to estimate International Roughness Index (IRI) values using pavement surface images. Furthermore, the effectiveness of artificial neural network (ANN) and multinomial logistic (MNL) regression models to categorize pavement sections into different roughness conditions was investigated. A pretrained CNN model was trained and validated using 850 three-dimensional (3D) pavement surface images, which were extracted from the Louisiana DOT and Development (LaDOTD) pavement management system (PMS) inventory. In addition, 1,142 test observations including IRI measurements and distress data were used to develop the ANN-based pattern recognition and MNL models. The developed CNN model outperformed the ANN and MNL models with an accuracy of 93.4% in the training stage. In addition, the CNN model predicted IRI values with a coefficient of determination (R2) of 0.985 and an average error of 5.9%. | |
