Comparison of Machine Learning, Deep Learning and Transfer Learning-Based Approaches for Crack Detection on Flexible Airport Runway Pavements Using Unmanned Aerial Vehicle Images

In this study, machine learning, deep learning and transfer learning-based approaches have been compared to determine the presence of pavement cracks of the airport runway using a total of 4032 images captured by an unmanned aerial vehicle. In the machine learning approach, Local Binary Pattern, Grey-Level Co-occurrence Matrix, and Histogram of Oriented Gradient are used for feature extraction. For classification, k-Nearest Neighbours and Support Vector Machine are used. Among them, the best performance (82.96%) is achieved whenever the Grey-Level Co-occurrence Matrix and k-Nearest Neighbours combination is used. The convolutional neural network approach gives 83.89% accuracy, which is higher than the machine learning approach. Maximum accuracy is achieved when the convolutional neural network trained with 6-layer blocks, which corresponds to 24 layers in total. In the transfer learning approach, among the nine pretrained models evaluated, VGG-19 achieved the highest validation accuracy under the evaluated runway-specific validation setting, with an accuracy of 90.09%, a precision of 96.63%, a recall of 90.52%, and an F1-score of 93.47%. This study indicates that it is possible to detect the cracks on flexible airport pavements using images captured from an unmanned aerial vehicle, with an F1-score of approximately 93%.