Eco-friendly pavement raveling detection based on in-situ data and transfer learning

Abstract

Pavement raveling is characterized by the progressive loss of surface aggregates and poses significant challenges to road durability and safety. The early detection and monitoring of this disease are crucial for implementing timely maintenance interventions. This study adopts a transfer learning approach for automated raveling detection in asphalt pavements. A comprehensive in situ raveling dataset comprising 530 images and 215 normal pavement images was collected using the LS-40 portable three-dimensional surface analyzer. Following histogram equalization for noise reduction, the dataset was augmented to 1600 images through mirroring and rotation techniques. The transfer learning was fine-tuned on four base convolutional neural network structures: the VGG16, the EfficientNet-B0, the InceptionV3, and the RegNet. The models were trained using standardized parameters, including an input size of 224 × 224, a batch size of 32, and 200 epochs, with regularization and dropout techniques applied to mitigate overfitting. A comparative analysis of optimization algorithms showed that RMSprop outperformed both SGD and Adam for this specific task. Transfer learning significantly enhanced the performance of all models, with the EfficientNet-B0 achieving outstanding results—achieving both high accuracy (99.7%) and low energy consumption in raveling detection. The selected models achieved AUC values exceeding 0.95, while the transfer learning significantly reduced training time and enhanced feature extraction capabilities, as confirmed through convolutional layer visualization. These findings establish the EfficientNet-B0 as the optimal structure for practical deployment in automated pavement inspection systems, providing a robust foundation for intelligent infrastructure maintenance strategies.