source: google research: from pixels to planning: earth ai for nature restoration
level: technical
google research released a vectorized version of its farmscapes 2020 map, transforming pixel-based detections into precise outlines of hedgerows, stone walls, and small woodlands across england. the original raster map identified fine-scale woody features often missed by standard satellite monitoring, but the new dataset converts these into a practical inventory for conservation and carbon accounting. the work was done with the leverhulme centre for nature recovery at the university of oxford.
the team built a deep learning framework using a vision transformer pretrained on over 300 million satellite images, then fine-tuned it on a small annotated dataset of british landscapes. they used submeter imagery and lidar to separate ground boundaries from above-ground features, solving the problem of overlapping elements like a hedgerow over a stone wall. a polsby–popper compactness score programmatically classified detections into woodlands, woody patches, and linear features based on shape, isolating the narrow corridors vital for wildlife movement.
processing was scaled across england using google earth engine, which handled millions of features in parallel by splitting the area into s2 grid cells and merging geometries at tile borders. the resulting open dataset aims to help landowners and policymakers measure and expand these small features without displacing farmland. future work will explore using the technology for silvopasture monitoring and detecting conservation leakage, where local gains are offset by losses elsewhere.
why it matters: this dataset makes previously invisible ecological features measurable, enabling data-driven restoration planning that balances carbon capture and biodiversity with food production.
source: google research: from pixels to planning: earth ai for nature restoration