DeepForest's Mask2Former model segments 18,507 tree crowns from UAV imagery

Researchers from DeepForest Technologies Co., Ltd., Kyoto University, and collaborators have developed a highly detailed instance segmentation model for delineating individual tree crowns in broadleaf forests using UAV imagery. The model, based on the Mask2Former architecture, was trained on 18,507 manually labeled crown polygons from orthomosaic images collected across seven forests in Japan. Broadleaf forests pose unique challenges due to diverse crown shapes and unclear treetops, but the model achieves high segmentation performance using only RGB imagery—without needing multispectral or LiDAR data. The best configuration generalizes well to geographically distinct forests within Japan and, notably, to biologically different tropical rainforests in Borneo, demonstrating robustness across diverse ecosystems.

The team emphasizes that the large, high-quality annotated dataset was critical for achieving detailed and generalizable results. The model has been integrated into DF Scanner Pro, a commercial software that enables wide users—from researchers to forest managers—to automatically extract tree-level information from UAV flights. This tool could significantly streamline forest monitoring, carbon stock estimation, and biodiversity assessments by replacing manual crown delineation with automated, scalable AI. The paper is available on arXiv and the model weights are expected to be released, though no open-source code is mentioned.