Road Extraction Based on Deep Learning Using Sdgsat-1 Nighttime Light Data

Previous road extraction research based on remote sensing data mainly used traditional optical remote sensing imagery acquired during the daytime, but some roads on the images were affected by problems such as building shadows and similar spectral features of road with other materials, which reduced the accuracy of road extraction. The Glimmer Imager (GI) equipped on the Sustainable Development Science Satellite 1 (SDGSAT-1) provides nighttime light (NTL) images with relatively high spatial resolution (RGB: 40m; PAN: 10m), which show detail of road networks. In order to explore the potential of the NTL data obtained from SDGSAT-1 on road extraction, a deep learning framework combining ResUnet and Dynamic Snake Convolution (DSC) module was proposed in this study. The experimental results demonstrated the effectiveness of SDGSAT-1 NTL on road extraction tasks and the extracted roads can be used for road dataset updating.