Citation

Al-Daeri, Abdulrahman Sufyan Taha Mohammed (2026) Deep learning-driven UAV remote sensing for crown delineation and parameters extraction in durian orchards. Masters thesis, Multimedia University.

Full text not available from this repository.

Abstract

Reliable tree-level structural inventories in managed durian orchards require accurate crown delineation and dependable extraction of crown geometry, tree height, and trunk diameter. This study evaluated D-PULSE (Durian-specific Pipeline for UAV-based Learning and Structural Examination), an integrated research workflow combining unmanned aerial vehicle (UAV) imagery, existing deep-learning instancesegmentation models, photogrammetric height estimation, field-calibrated trunk diameter at 30 cm above ground (D30) modelling, and cross-site diagnostics. Four orchard blocks in Melaka, Johor, and Pahang provided 3,628 manually delineated durian crowns, while 370 field-measured trees at Melaka A supported height evaluation and D30 calibration. Three segmentation paradigms—one-stage YOLO26s-seg, two-stage Detectree2-based Mask R-CNN, and query-based Mask2Former—were evaluated using different test sets; therefore, their results do not establish a controlled head-tohead ranking. Mask R-CNN achieved an overall F1-score of 84.55% at an intersectionover-union threshold of 0.50, while YOLO26s-seg achieved 84.58% in its separate evaluation. Mask2Former recovered more small crowns in Melaka and Johor, with some loss of precision. Crown area was more consistently associated with detectability than tree height. At Melaka A, ordinary least-squares correction of DSM apex–ground heights reduced the held-out height RMSE to 0.568 m. Among eight candidate allometric equations, the crown area–height interaction model achieved the best nonlinear least-squares fit for D30 (R² = 0.896; RMSE = 1.22 cm). Cross-site behaviour was examined using mean absolute percentage divergence (MAPD) between this height-integrated model (D3) and a crown-area-only model (D2). Melaka had a MAPD of 9.62% (95% bootstrap reference interval: 8.81–10.53%). Johor was within this interval under deterministic application but exceeded it after height-uncertainty propagation, whereas Pahang remained outside it. Since independent field D30 measurements were unavailable in Johor and Pahang, these findings indicate differences in model behaviour rather than externally validated D30 accuracy. Overall, D-PULSE links crown-mapping quality with structural estimation and helps identify orchards where additional field validation or recalibration may be warranted.

Item Type:	Thesis (Masters)
Additional Information:	Call No.: S494.5.R4 A43 2026
Uncontrolled Keywords:	Agriculture—Remote sensing
Subjects:	S Agriculture > S Agriculture (General)
Divisions:	Faculty of Engineering and Technology (FET)
Depositing User:	Ms Nurul Iqtiani Ahmad
Date Deposited:	05 Nov 2026 13:35
Last Modified:	11 Nov 2026 02:51
URII:	http://shdl.mmu.edu.my/id/eprint/17074

Downloads

Edit (login required)