Utilization of Convolutional Neural Networks and Satellite Images for the Prediction of Road Quality in the Philippines

Road transportation is the largest transport sub-sector in the Philippines – accounting for 98% and 58% of all passenger and cargo traffic respectively. Despite the heavy reliance on road infrastructure, the maintenance process of roads in the Philippines still involves an approach that requires plenty of human intervention at each step which is tedious, costly, and prone to bias. In this paper, a more efficient automated process of collecting road quality data through the utilization of satellite imagery is studied. The paper contributes to existing literature by applying and optimizing a proven technique to a different context. Through machine learning using Convolutional Neural Networks (CNNs), a model which is capable of predicting road quality given just coordinates and satellite images is developed. The dataset for this study consists of an extensive number of roads in the Philippines amounting to around 33,000 km and 54,000 images classified into "good", "fair", "poor", and "bad". The best performing models achieved approximate accuracies of 40% for the four-class classification task and 74% on the binary classification task. A relative accuracy of 83% for the best performing model was also recorded and shows promise for this preliminary study. These results show that a machine learning model is a hopeful and viable replacement for the laborious process that is currently used in the country.