Google maps hedgerows for climate work

Google Research published an open resource on June 16 from Google Earth AI that turns agricultural landscape maps into a vector inventory of hedgerows, stone walls, copses and small woodlands across the United Kingdom. The central fact is specific: the team is no longer only detecting pixels of vegetation. It is releasing a vectorized dataset, meaning usable geometric shapes, so landowners, researchers and public bodies can measure fine ecological features that often disappear from conventional forest inventories.

The practical problem is larger than a mapping exercise. Protecting or planting trees does not always mean creating large forests. On working farmland, hedgerows and wooded corridors can store carbon, reduce erosion, host biodiversity and connect habitats while leaving fields in production. These structures are small, elongated and often mixed with walls or field boundaries. Standard satellite detection can miss them, and a pixel map is still hard to use when someone needs to decide where to restore, measure or finance local action.

To move from pixels to planning, Google describes a processing pipeline that combines deep learning, submeter imagery, 1-meter LiDAR data and Google Earth Engine. The model was fine-tuned from the Remote Sensing Foundations Vision Transformer backbone, which was pretrained on more than 300 million global satellite images. Because the local annotations covered only about 247 km², that pretraining gives the system a broad visual base before it learns the specific shapes of the British countryside. The method then adds a dual-layer labeling system to separate ground-level boundaries from objects above them, and merges geometries that were split by map tiles.

The useful part is the semantics attached to the shapes. A tree cluster, a hedgerow and a woodland patch do not serve the same ecological function. Google says it uses tools including the Polsby-Popper compactness score, a geometric measure, to classify detected outlines into continuous woodland, small copses or linear woody features such as hedgerows. This is not an automatic restoration decision. It is a sharper way to read the landscape. The data can help prioritize wildlife corridors, document carbon storage or check whether local ecological gains are offset by losses just outside a project boundary. The Mindshot signal is clear: climate AI is becoming less theatrical and more cadastral, producing data granular enough to enter real land-use decisions.