Peatland foundation model
Self-supervised geospatial backbone for UK peatland monitoring.
Peatlands are the largest terrestrial carbon store in the UK, and their erosion features are difficult to delineate from imagery alone — labels are scarce, sensors are heterogeneous, and the visual signal is subtle.
Under an Innovate UK grant, we developed a self-supervised geospatial backbone (BEiT / MAE variants) on ~20,000+ sq. km. of multiresolution UK peatland imagery, combining RGB orthophotos with digital surface models. A downstream UPerNet segmentation head targets erosion features on the ground. To handle domain shift across sensors and regions, active-learning sample-selection loop picks the most informative tiles from heterogeneous data sources for human labelling.
The deployed system delivers a ~10× mIoU improvement over the incumbent DEFRA-backed baseline and is in production use at Calterra for ecosystem monitoring.