Extended Data Fig. 1: Deep learning framework diagram.
From: Advancing solar and wind penetration in China through energy complementarity

The workflow consists of four steps: (1) manual sample annotation; (2) deep learning model training; (3) nationwide image inference; and (4) manual verification. During the manual verification step, for regions in which certain installations were missed during the initial Google imagery inference, Jilin-1 satellite imagery was used as an alternative data source to re-infer these areas, thereby recovering undetected installations and improving overall recall. Base map derived from Natural Earth shaded relief data (public domain).