Urban Tree Crown Instance Segmentation Using Cross-Domain Foundation Models

Kurzfassung

Accurate urban individual tree crown segmentation remains challenging due to the complexity of urban environments as well as the unclear boundaries between adjacent trees and fused canopies. In this study, we propose a cross-domain two-step framework to accurately segment tree crowns in urban regions. 1) We apply a VLM-based approach for tree segmentation using GeoVLM. 2) We present cross-domain fine-tuning to Cellpose-SAM to adapt the segmentation model to tree structures. Experiments on the BAMFORESTS dataset outperform existing methods while requiring only limited fine-tuning. Experiments conducted in a selected urban environment further show that domain adaptation under complex condition works.