Abstract
Solar photovoltaic (PV) power generation, known for its affordability and environmental benefits, is a key component of the global energy supply. However, the lack of comprehensive, timely, and precise global PV datasets has limited spatial analysis of PV potential. We developed a new method to identify PV panels globally, producing an annual 20-meter resolution dataset for 2019-2022. This dataset offers unprecedented detail and accuracy for future research and policy-making. A two-stage PV classification framework was built using U-Net and positive unlabelled learning with random forest (PUL-RF). U-Net first recognizes PVs from sub-meter Google Earth images, expanding positive PV samples for the second stage, where PUL-RF classifies Sentinel-2 images on a large scale. The dataset was evaluated with IoU and F1-Score metrics, achieving over 90% accuracy. Compared to existing datasets, it provides better precision and spatial detail, showing global PV growth of over 60% between 2019 and 2022, with developing countries leading the increase.