Crown dieback and dead trees detection from digital aerial images: comparison of artificial intelligence–based deep learning models across forest sites from a wide geographic gradient

Abstract Effective forest health monitoring requires timely, accurate detection of crown dieback and dead trees. Artificial intelligence and particularly deep learning (DL) approaches integrated with high-resolution RGB aerial imagery overcome past remote sensing limitations, related to model transferability and spatial scale by enabling automated, tree-level monitoring across large forest landscapes. This study evaluated four DL models (YOLO11s-SAM, YOLO11s-seg, DeepLabV3+, and SegFormer) for mapping dieback and dead tree crowns using imagery from five ecologically distinct forest sites in Iran, Germany, Luxembourg, and the USA. Two training strategies were evaluated: models trained separately on individual sites (site-specific training) and models trained on combined data from multiple sites (integrated training), with one site held out as an independent test set to assess transferability. In site-specific experiments, YOLO11s-SAM achieved the highest point-based (verifying the correct localization of dead and dieback trees) accuracy (94.81%–98.51%) and the best polygon-based performance (verifying the correct delineation of the affected areas using the metric segments shape similarity) in three of four sites, while SegFormer led in the Luxembourg Forest. Under integrated training, SegFormer achieved higher point-based accuracy (95.55%–97.77%) and polygon-based accuracy (90.50%–94.54%) in most sites, whereas YOLO11s-SAM remained superior only in the Alang Darreh Forest. Both models generalized well to the independent test site, with YOLO11s-SAM and SegFormer achieving point-based accuracies of 94.66% and 93.33%, and polygon-based accuracies of 87.38% and 86.46%, respectively. YOLO11s-SAM produced the most segments shape similarity but tended to underestimate crown extent, whereas SegFormer consistently overestimated it. These findings highlight the robustness and generalizability of YOLO11s-SAM and SegFormer DL models for forest health monitoring. They enable a fully automated, spatially transferable workflow using widely available RGB aerial imagery, demonstrating strong potential for developing a global foundational model to accurately detect canopy dieback and dead tree at scale.