Answers are grounded in this paper and cite their sources.
Automated tree-level inventories are essential for precision orchard management, yet unmanned aerial vehicle (UAV)–deep learning frameworks often evaluate cross-site performance only at crown mapping. This study introduces the Durian-specific Pipeline for Uncertainty-Led Structural Examination (D-PULSE), which integrates Mask R-CNN crown delineation, crown-geometry extraction, digital surface model apex–ground height estimation, and field-calibrated D30 allometry, where D30 is trunk diameter measured 30 cm above ground. D-PULSE also evaluates cross-site D2–D3 model divergence and height-error sensitivity. Testing across orchards in Melaka, Johor, and Pahang showed strongly site-dependent crown delineation. Pahang achieved F1-scores of 98.30% at intersection over union (IoU) ≥ 0.50 and 96.32% at IoU ≥ 0.75. Melaka and Johor achieved 77.54% and 76.03%, respectively, at IoU ≥ 0.50, and 47.08% and 52.56% at IoU ≥ 0.75, mainly because of small-crown omission. The apex–ground height method achieved R² = 0.586 and RMSE = 0.90 m. Among eight D30 equations, the crown area–height interaction model performed best, achieving R² = 0.896 and RMSE = 1.22 cm under nonlinear least squares and R² = 0.891 and RMSE = 1.25 cm under Bayesian calibration. Mean absolute percentage divergence (MAPD) between the crown-area-only D2 model and height-integrated D3 model was 9.62% in Melaka, with a 5,000-iteration bootstrap interval of 8.79%–10.50%. Johor’s deterministic divergence fell within this interval but increased under height-error sensitivity analysis, whereas Pahang remained above it. Crown-mapping performance and downstream allometric model behaviour therefore varied independently, supporting D-PULSE as a diagnostic framework for prioritizing local field validation.
No takes yet. Share an insight, caveat, or question.
Aldaeri et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: