Global wind turbine detection via installation suitability-constrained deep learning with multi-source geospatial data

The rapid expansion of global wind power infrastructure has urgently necessitated accurate identification of wind turbines (WTs). In this study, we propose a deep learning detection framework constrained by multi-source geospatial data for automated global WT identification using remote sensing imagery. By integrating land-use, topography, anthropogenic buffers, and wind power density constraints, we construct a global WT installation suitability map that serves as a spatial constraint, which effectively reduces the search space by 38.99%. An iterative sample refinement strategy is then employed, where dataset expansion and model training are jointly performed through a feedback loop. The final YOLOv8 model achieves precision values of 96.3% and 90.1% on the training and validation sets, respectively. Within the global latitude range of ±70°, 404,392 WTs are detected, including 48,688 previously unrecorded in OpenStreetMap. The detection results are consistent with known wind resource distributions, predominantly concentrated in China, the United States, and India. Detection performance varies across terrain types, with the highest accuracy observed in inland plains (95.61%). The proposed framework provides a robust and scalable approach for global wind power infrastructure mapping and monitoring, thereby offering valuable insights for energy planning.