From Pixels to Planning: Earth AI for Nature Restoration

Google Research developed a high-resolution deep learning framework to map fine-scale woody features such as hedgerows, shelterbelts, and copses that are typically invisible to standard satellite detection. These features can enhance carbon storage and biodiversity without displacing crops. The team released a vectorized dataset for the UK, converting pixel-based maps into actionable inventories of hedgerows, stone walls, and copses. To overcome data scarcity, they used Remote Sensing Foundations Vision-Transformer backbone pre-trained on 300 million global satellite images, fine-tuned for British landscapes. A dual-layer labeling system using submeter imagery and LiDAR data handled overlapping features, while the Polsby–Popper compactness score classified shapes by ecological function. Google Earth Engine enabled parallel processing of millions of features across England. The dataset aims to empower farmers, scientists, and policymakers to protect small-scale features for climate and biodiversity goals.