From pixels to planning: Earth AI for nature restoration

Details

Google Research and Oxford's Leverhulme Centre for Nature Recovery have released Vectorized Farmscapes 2020, converting their earlier raster Farmscapes map into a vector dataset identifying discrete hedgerows, copses, and stone walls across England's agricultural landscape. These fine-scale woody features fall below the detection threshold of standard satellite imagery and are excluded from national forest inventories — making them invisible to conservation programs despite meaningful ecological value. The underlying deep learning framework had to solve a non-trivial spatial topology problem: agricultural features frequently overlap or adjoin — a hedgerow running beside a stone wall, for instance — and single-layer models cannot represent these compound structures. The vectorized output lets landowners and conservationists generate inventory counts, measure coverage, and identify expansion opportunities at precision sufficient for carbon accounting and biodiversity planning. The ecological argument is compelling: hedgerows and shelterbelts can sequester carbon and support biodiversity without displacing cropland, addressing the core tension between food production expansion and climate commitments. This is a planning tool for working agricultural lands, not a substitute for large-scale afforestation. The methodology transfers beyond England: similar high-resolution deep learning frameworks could produce national-scale inventories of fine-scale vegetation in other countries. No commercial product was announced — this is a research dataset release with practical utility for the UK conservation and land management sector.