Forest Research UK implements Meta DINOv2 AI for accurate tree mapping

The British forestry research organization Forest Research is integrating Meta's DINOv2 artificial intelligence model to significantly enhance tree mapping accuracy. This initiative is part of a broad program to monitor the country's forest stands, aimed at supporting the ambitious government's Environmental Improvement Plan. The model is being used to update and refine data, including information for the Trees Outside Woodlands (ToW) map, expected to be published in April 2025.

DINOv2 is an open computer vision model, developed by Meta and trained on an extensive dataset of 18 million satellite images. The result of its training is a global 1-meter resolution tree canopy height map, launched in April 2024, capable of identifying individual trees worldwide. Forest Research is applying this advanced technology, including through funding from the UK's Natural Capital and Ecosystem Assessment (NCEA) program.

Previously, Forest Research relied on costly and labor-intensive ground surveys and LiDAR data to assess forest cover. These traditional methods proved inefficient for continuous monitoring, especially when tracking individual trees, small groups, and forest plots less than 0.5 hectares in area, which, however, constitute about 30% of England's total tree cover. The implementation of DINOv2 offers a much more efficient and accurate solution to these longstanding problems.

Enhanced mapping accuracy is crucial for tracking the UK's progress in achieving its ambitious environmental protection goals. The government has set targets: by 2050, to ensure every resident in England has access to a green space within a 15-minute walk, visibility of at least three trees from home, and a 2% increase in forest cover. DINOv2 provides reliable data needed by the Department for Environment, Food and Rural Affairs (Defra) for informed policy decisions.

Freddie Hunter, Head of Remote Sensing at Forest Research, called the DINOv2 model a "game-changing" solution for detecting and monitoring individual trees at a national scale. His team plans to use tree canopy height maps in conjunction with aerial photographs to create more up-to-date assessments of the area and height of both individual trees and small forest stands, as well as to calculate timber volume losses from felling. Applying the model to aerial photographs can also provide synthetic tree canopy height models that surpass existing national LiDAR surveys in identifying individual trees. Forest Research also intends to use the structural information obtained to more accurately map tree types across the country and assess forest cover in complex urban environments.

Interest in DINOv2-based models and maps has already been shown by many governments worldwide, seeking to improve their own forest restoration programs. This underscores the global potential of Meta's open technology in natural resource monitoring.