Mapping the first dataset of global urban land uses with Sentinel-2 imagery and POI prompt

Abstract

An up-to-date, detailed global urban land use map is essential for disclosing urban structures and dynamics as well as their differences across different regions. However, generating an accurate global urban land use map remains challenging due to the complex diversity of land use types and the uneven availability of data. Existing methods, which either rely solely on remote sensing imagery or treat supplementary data like point-of-interest (POI) as mandatory inputs, fail to account for regional data disparities and the complex relationships between different data modalities. In this study, we propose an urban land use mapping framework optimized with POI prompts. First, we acquire global urban Sentinel-2 imagery, POI data, and labeled LU samples from Google Earth Engine (GEE) and OpenStreetMap (OSM), and then propose a POI-Prompt Urban Land-use Mapping Network (PPUL-Net), which utilizes POI prompts to enhance classification accuracy and produce reliable predictions in those regions lacking POI data. Consequently, a global land use dataset (GULU) with a 10 m resolution in 2020 has been produced at the first time. Experimental results show that GULU has an overall accuracy of 84.59 %, demonstrating robust performance across continents. Incorporating POI data improved the accuracies for some challenging categories, such as commercial and institutional lands, by 7.32 % and 17.66 %, respectively. Additionally, using POI as prompts instead of direct pixel-level fusion with imagery increased accuracy by 2.92 %. Finally, analysis of the GULU dataset reveals that Europe and North America exhibit high land use diversity, implying mature urban structures, whereas sub-Saharan Africa and South America are predominantly characterized by residential and undeveloped areas. This dataset provides invaluable insights for urban planning, development monitoring, and sustainable development assessments.