Accessible road detection from remote sensing image based on the deep learning method

Numerous earthquake rescue experiences indicate that obtaining the accessible roads in time after the destructive earthquakes is the cornerstone of the emergency rescue. Traditional methods, such as on-site surveys and manual interpretation of optical data, are time-consuming and don’t meet the requirement of emergency rescue. The rapid development of computer technology has promoted the application of the deep learning in the road extraction. In this study, we first used the high-spatial resolution optical image of Beijing to produce the road labels containing the road surface, road edges and road centerline. Then a network model connecting two CNNs in series is constructed. This network model is designed to extract road surface and road edges at same time, then the road centerline is calculated using the skeletonization method. This model achieved a Precision of 0.86 and a Recall of 0.91 on road surface segmentation; a ODS of 0.82 and an OIS of 0.84 on road edges and centerline extraction. Therefore, it means that building a pre-trained road extraction model that learning the local road characteristics can quickly and accurately extract the spatial distribution of accessible roads from optical images of areas attacked by the earthquake, and provide data support and technical reserves for the emergency rescue planning and rescue forces deployment.