Species-Specific Crown Segmentation of Araucaria angustifolia in UAV Orthomosaics: Semantic and Instance Pipelines with Operational Tiled Inference

Reliable individual-tree mapping from unmanned aerial vehicle (UAV) orthomosaics requires segmentation and reconstruction across overlapping tiles. This study compares five semantic and instance segmentation pipelines for mapping Araucaria angustifolia in Paraná, Brazil, using 20,108 orthomosaic reference polygons. Implementations were evaluated as configured, assessing tile-level accuracy, reconstructed canopy extent, and individual-crown agreement across five held-out test surveys. YOLACT++ achieved a pooled individual-crown F1 of 79.87% and a maximum survey-level F1 of 83.22%. In the validation ablation, cross-tile mask fusion consolidated repeated predictions, increasing precision by 29.17–31.74 percentage points across the three instance pipelines, with recall losses of 0.84–1.59 percentage points and improved F1. Tile-level performance did not fully reflect reconstructed crown agreement, while survey variation reinforced the need for evaluation under intended mapping conditions. Photograph supplementation provided no consistent improvement. All pipelines processed the benchmark orthomosaic and exported georeferenced predictions in under one minute using consumer-grade hardware. This assessment quantifies the contribution of spatial reconstruction to crown mapping and supports workflow selection for preliminary inventories and targeted inspection.