Leveraging optical and SAR data with a UU-Net for large-scale road extraction

Abstract

Road datasets are fundamental and imperative for traffic management and urban planning. Different high-resolution optical remote sensing images are widely used for automatic road extraction but the results are usually limited to local scale and spectral confusions in barren and cropland, while accurate large-scale road extraction remains challenging. In this study, we incorporated medium resolution optical and SAR data, i.e., 10-meter resolution Sentinel-1 and Sentinel-2, for road extraction at a large scale and evaluated the contribution of different data sources. We developed a United U-Net (UU-Net) to fuse optical and SAR data for road extraction, which was trained and evaluated on a large-scale multisource road extraction dataset. The UU-Net achieved better accuracy than traditional deep convolutional networks with optical or SAR data alone, which obtained an average F1 of 0.5502 and an average IoU of 0.4021, outperforming in 160 out of 200 (80%) 0.5-by-0.5 degree evaluation grids. The results indicated that SAR contributes more to road extraction in barren land, while optical data contributes more to large slope areas. The road accuracy is positively related to elevation and urban percentage, which distributes higher in eastern China and lower in western. The road centerline from 10 m road showed comparable results with that from Open Street Map (OSM), indicating its promising applications to support large-scale urban transportation studies.