Key points are not available for this paper at this time.
Urban green space (UGS) maps are essential for identifying and assessing the multifunctional benefits of nature in cities. However, obtaining reasonable-quality UGS data across the Global North and South remains challenging due to methodological inconsistencies and the high costs of field-based data collection. We developed a scalable and replicable framework that leverages freely available, moderate-resolution satellite images, combined with artificial intelligence-based (AI) image segmentation, to detect and map UGSes. Sentinel-2 images were retrieved across 16 cities in North America, Europe, the Middle East, and South Asia. Using raw and processed Sentinel-2 spectral information, we trained and validated an AI hybrid model combining U-Net and ResNet-50 on varying combinations of data layers (i.e., normalized difference vegetation index NDVI, normalized difference water index NDWI, and normalized difference built index NDBI). The trained models achieved approximately 90% accuracy in identifying UGS and demonstrated a substantial overlap with the ground-truth data across diverse urban settings. However, consistent with known limitations of moderate-resolution imagery, the models underperformed in detecting relatively small UGS patches. To test the geographic transferability of the model, we applied the trained model to detect UGS in an African city (Kampala, Uganda), where ground-truth data were unavailable. We found that the UGS identified from the model partially overlapped with the UGS in Kampala, as derived from OpenStreetMap data, suggesting that combining AI-derived and volunteered geographic information can produce more comprehensive UGS inventories. Overall, this scalable framework for identifying UGS in places with limited existing data could enable cities to inventory their UGS and target the Sustainable Development Goals.
Zhao et al. (Tue,) studied this question.