Improving impervious surface mapping accuracy in arid regions using Sentinel-1 and Sentinel-2 data combined with OSM data

Impervious surface mapping plays a crucial role in urban planning, environmental management, and sustainable development. However, existing land cover datasets often exhibit limited accuracy in mapping impervious surfaces in arid regions. This is primarily attributable to the presence of extensive barelands that are easily misclassified as impervious surfaces and the challenges in obtaining representative samples due to the dispersed villages in arid regions. This study combines three land cover products (FROM-GLC, GLC_FCS30D, and ESRI-LULC) to generate uniformly distributed regional training samples based on geographic grids. Considering that OSM data contain local knowledge and experience, this study introduces OSM data, generating local impervious surface training samples based on OSM data combined with raster kernel density. By combining regional and local training samples without any labeled samples, the impervious surface mapping accuracy in arid regions improved significantly compared with the other three datasets, with OA improving by 11.10%-24.73% and kappa by 0.1408-0.3161, demonstrating the effectiveness of the proposed method.