From pixels to planning: Google AI for nature restoration

Google researchers have introduced a deep learning framework that can identify small landscape elements (hedgerows, shelterbelts, copses) that are invisible on standard satellite imagery. These features can increase carbon uptake and biodiversity without removing land from agricultural use. Previously, the Farmscapes 2020 raster dataset (England) was released, and now a vector dataset of hedgerows, stone walls, and copses across the whole United Kingdom has been published. To create it, they used a pretrained Vision-Transformer (ViT) model from Remote Sensing Foundations (Google Earth AI), fine-tuned on 247 km² of annotated data, along with LiDAR. For object classification, they applied the Polsby-Popper compactness score (elongated objects with a coefficient < 0.5). Processing was done in Google Earth Engine. The data is open and available for farmers, scientists, and policymakers.