Large-Scale nighttime illuminated road extraction in the BTH region of China using SDGSAT-1 nighttime light data

Nighttime illuminated roads (NTIRs) are closely related to residents' nighttime travel and activities. However, traditional nighttime light (NTL) data are not highly applicable for large-scale NTIR extraction tasks because of challenges such as low resolution and data acquisition. In contrast, the open-sourced SDGSAT-1 NTL data released in 2021 which contains a panchromatic band (PAN) of 10-m resolution and red, green, blue bands (RGB) of 40-m resolution, are more suitable for this task. To date, research on NTIR extraction methods is still insufficient, and the application potential of NTIR products has not been fully explored. In this work, a deep learning model called DSC-UNet was proposed for NTIR extraction via SDGSAT-1 NTL imagery of 10-m spatial resolution. The proposed model uses UNet as the basic architecture and incorporates a dynamic snake convolution module to increase the sensitivity of the NTIR pixels. Our experimental results showed that DSC-UNet outperformed seven baseline models. Using the proposed model and vectorization tools, a high-accuracy NTIR centerline product of the BTH region was generated. By applying spatial statistics, the road illumination rate of the BTH region was assessed. The assessment results revealed an imbalance in road lighting levels among different cities in the BTH region.