This paper examines the predictive capability of 14 different Machine Learning (ML) classifiers and a Deep Learning algorithm (VGG16) for the classification of unpaved rural roads surface condition using high-resolution satellite imagery. An extraction and segmentation method has been developed for extracting the surface of road networks from satellite imagery and segmenting the surface image into 6x6 pixel tiles. ML classifiers are trained and tested with a road surface condition characteristics dataset, which produces a 24-dimensional feature vector (colour eigenvectors and texture eigenvectors) calculated from each tile. While the VGG16 is trained and tested directly using satellite images. The prediction of all models regarding the categorisation of roads into four categories, good, fair, poor, and bad, is compared with a ground truth condition survey. Based on the results of this study, a blended model achieves the highest accuracy of 88.8%, with high results also for precision, recall and F1. This indicates that satellite imagery combined with machine learning can be used to identify the surface condition of unpaved rural roads as a feasible alternative to conventional driven surveys.
This paper was awarded First Prize by ITS Asia.
This paper is only available on request. Please use the contact form to ask for a copy.