Toward Consistent Canopy Characterization From Low-Cost UAV Imagery

Monitoring tropical forest structure at landscape scale requires cost-effective methods capable of bridging the gap between field inventories and satellite remote sensing products. However, the diversity of available individual tree crown (ITC) segmentation algorithms raises questions about the consistency and reliability of derived structural metrics-a critical issue for any ecological application relying on these outputs. We evaluated three ITC segmentation algorithms-Detectree2, SAM, and the hybrid Detectree2SAM (D2S)-applied to very high-resolution RGB orthomosaics (5 cm resolution) acquired over ~200 ha of the Luki Biosphere Reserve (Democratic Republic of Congo) using a low-cost drone. Algorithms were validated at the individual scale against a photointerpretation reference of 1882 manually delineated crowns and at the plot scale against a field inventory of 360 trees across 18 plots. At the individual scale, IoU-based F1 scores ranged from 0.57 (D2S) to 0.67 (SAM), revealing clear precision-recall trade-offs. At the plot scale, D2S provided the most accurate crown area estimates (RMSD = 26%-29%), while SAM best reproduced canopy density (RMSD = 22%). All three algorithms systematically overestimated aggregated metrics; a linear correction substantially reduced prediction errors and improved cross-algorithm convergence, yielding consistent estimates-median crown area of the 20 largest trees (~148 m²), total crown area (~3850 m² per plot), and canopy density (~80 ind/ha). Agreement maps revealed systematic spatial divergences, including edge artifacts, SAM's tendency to underestimate canopy density, and localized overestimation of crown areas in disturbed stands. Raw ITC outputs carry substantial systematic biases requiring explicit field-calibrated correction before ecological use. We propose a three-step operational pipeline-(i) plot-scale field validation, (ii) per-algorithm linear bias correction, and (iii) landscape-scale aggregation-that reduces inter-algorithm divergence and delivers consistent, ecologically interpretable structural estimates from low-cost UAV imagery.