Google researchers have released a vectorized dataset that transforms high-resolution satellite imagery into actionable ecological inventory, revealing fine-scale features like hedgerows and copses across England. The Earth AI framework uses deep learning to detect these woody elements that standard satellite detection misses, offering new tools for nature restoration on working agricultural lands without sacrificing food production.

The technical breakthrough required solving three major challenges: complex spatial topologies where features overlap, semantic classification to distinguish between forests, corridors, and patches, and computational scale for processing over 130,000 km². The team built a dual-layer labeling system combining submeter imagery with 1-meter LiDAR data, and developed algorithms to merge artificially sliced features across map tiles.

Starting with a Vision-Transformer pre-trained on 300 million global satellite images, the researchers fine-tuned the model using limited annotated data (~247 km²). They applied the Polsby-Popper compactness score to automatically classify features by shape, defining linear woody corridors as those with scores below 0.5. Google Earth Engine enabled parallel processing of thousands of S2 cells to generate vector geometries for millions of features simultaneously.

This vectorized inventory empowers landowners and conservationists to measure and expand small-scale woody features throughout the UK, supporting carbon accounting and biodiversity goals. The open data release provides a pathway to scale restoration efforts while monitoring for 'leakage' events that could offset local environmental gains.