Domain-Aware Transfer Learning With SAM-Assisted Mask R-CNN for Urban Tree Crown Delineation From UAV Orthophotos

Accurate tree crown delineation from very high-resolution UAV orthophotos in urban areas remains challenging due to landscape complexity and the scarcity of representative annotated data. This study proposes an enhanced Mask R-CNN approach, supported by semi-automatic annotation using the Segment Anything Model-HQ (ViT-H Quant), to improve tree crown delineation. The approach evaluates the effect of backbone architectures (ResNet50 and ResNet101) and transfer learning strategies from general (COCO) and domain-specific datasets (OAM-TCD and Detectree2). Four model variants were tested through 5-fold cross-validation and independent test data evaluation. Results demonstrate that the model with a ResNet50 backbone pre-trained on OAM-TCD achieved the best performance (AP 49.56%, F1-score 0.722, precision 0.720, recall 0.723), while also being computationally more efficient than ResNet101. Although Wilcoxon signed-rank tests indicated that the improvements were not statistically significant (p > 0.05), consistent performance trends still favored the OAM-TCD–based model. These findings highlight the importance of domain-specific pretraining and computational efficiency for practical large-scale UAV-based urban vegetation monitoring.