Evaluation of CHM Morphological Processing and U-Net Deep Learning for Citrus Tree Canopy Delineation from UAV Imagery

Abstract

Accurate delineation of individual tree crowns from UAV imagery is essential for precision citrus management but remains difficult in orchards with overlapping crowns and inter-row weeds. This study compared two contrasting approaches for citrus crown delineation: a canopy height model (CHM)-based morphological pipeline and a U-Net deep-learning model applied to RGB imagery. Both used an identical marker-controlled Watershed step for individual crown separation, so that the comparison reflected the segmentation stage alone. RGB imagery was acquired with a DJI Mavic 3M UAV over two structurally contrasting citrus orchards in the Beni Mellal-Khenifra region of Morocco, planted with the Maroc Late and Sidi Aissa varieties, and performance was assessed at the pixel and object levels. In the structurally simple Maroc Late orchard the two methods were effectively equivalent, both achieving high object-level agreement (detection score, the Jaccard index of the detected and reference crown sets, about 94%) and recovering crown area almost perfectly. The methods diverged sharply in the dense, weed-affected Sidi Aissa orchard: the U-Net sustained robust performance (detection score 93.6%; crown-area R² = 0.93), whereas the CHM-based pipeline, although it still detected roughly two-thirds of the reference crowns, was heavily penalised by false positives from weeds and merged crowns, lowering its detection score to 46.9% (crown-area R² = 0.64). Because the separation step was held constant, this divergence is attributable to the quality of the upstream mask rather than to the delineation algorithm: learned spectral and textural features proved decisive where the height threshold lost discriminative power. The results indicate that method choice should be matched to orchard structure, a lightweight height-threshold approach being sufficient in clean orchards and an RGB-trained model preferable under structural complexity. These findings, obtained from two citrus varieties at a single site and date, require multi-site validation before broader generalization.