A Method for Road Surface Anomaly Detection from Remote Sensing Data by Deep Convolutional Networks

In the area of transportation construction and road safety, road structures need to be frequently inspected in order to discover damage threatens and potential risks. Therefore, the automatic technique of detecting road surface anomaly is very important. The paper aims to extract road surface anomaly from remote sensing data. The anomalies can be regarded as possible road damage candidates. This is a very challenging task, and we focus on the targets whose sizes are not too tiny to be seen in remote sensing image, like landslide, pothole, ponding, and so on. The basic idea of the proposed approach is to firstly identify the road pavement materials, and then analyze their characteristics to find abnormal conditions. We treat the task of road material identification as an image object segmentation/classification process, for which the methods based on deep learning framework can be applied. Then, an approach is developed to detect if there exits any abnormal subjects in the road region, and these subjects are the candidates of desired anomalies. These candidates are further analyzed based on the shape and spectral features, to output the final anomaly results.