Mapping individual tree crowns from aerial imagery helps us understand forest structure and monitor ecosystem health. But a major bottleneck remains: AI models often need new, manually labeled examples when moving to a different landscape or image resolution. Our lab’s new study introduces UniTree, a framework that transfers existing knowledge rather than requiring new crown annotations for each target scene.
Led by Jiaqi Yang, the study tackles a challenging transfer: learning from 20-centimeter aerial imagery in Denmark to map trees in 60-centimeter imagery in Yosemite National Park. UniTree uses existing Danish crown labels alongside unlabeled Yosemite imagery during training; manually delineated Yosemite crowns are reserved for evaluation.
How does it work? UniTree combines three complementary strategies: adapting to differences between regions, transferring knowledge from a “teacher” network working with sharper imagery to a “student” working with coarser imagery, and progressively shifting training toward crown boundaries and geometric structure. Together, these strategies help the model recognize trees despite changes in their appearance and the detail available in the images.
Across three Yosemite test areas, UniTree outperformed the comparison segmentation models, achieving mean intersection-over-union scores—a measure of overlap between predicted and reference maps—of 0.78–0.88. Its tree counts also showed broad agreement with independent lidar-based estimates.
The broader promise is more scalable forest monitoring with less repetitive labeling. Shadows and visually ambiguous crowns remain challenges, but UniTree offers a promising route toward automated digital forest inventories and tree-level ecological assessments.
Publication: Yang et al. (2026), “Unified knowledge transfer boosts individual tree crown segmentation without scene-specific labels.” Remote Sensing of Environment, 347, 115646. DOI: 10.1016/j.rse.2026.115646.
