<![CDATA[Conventional Red-Green-Blue (RGB) unmanned aerial vehicle (UAV) imagery offers a cost-effective alternative for precision agriculture; however, its limited spectral separability constrains classification when target classes exhibit high visual similarity, particularly in non-cultivated areas with dense vegetation. Maturity assessment of sago palm (Metroxylon sagu Rottb.) exemplifies this challenge: identification of the harvestable stage remains dependent on labour-intensive field inspection, and delayed identification causes stem mortality and yield loss. This study evaluates an Object-Based Image Analysis (OBIA) framework for discriminating three sago maturity levels (Young, Harvestable, Overripe) from RGB orthomosaics and a Digital Surface Model over a 9-hectare non-cultivated site in Wailawi, North Luwu, Indonesia, using a DJI Phantom 4 Pro at 50 m altitude. Multi-resolution segmentation generated 6,210 crown-level objects, classified by Random Forest under two configurations: a five-feature set and a seventeen-feature set optimised through Recursive Feature Elimination. Evaluation used 600 independent objects (200 per class) from the testing partition through stratified random sampling, re-labelled by visual interpretation, with 95% confidence intervals from 1,000 bootstrap resamples. The seventeen-feature model outperformed the baseline, yielding Overall Accuracy of 93.00% (95% CI: 91.00–94.83) versus 89.00% (86.83–91.50), Macro-F1 of 92.92% versus 88.92%, and Cohen's Kappa of 0.895 (0.860–0.922) against 0.835 (0.795–0.873). Classification uncertainty concentrated at the Young-Harvestable boundary, whereas Overripe was consistently discriminated (F1 = 98.77%). Visible-band spectral statistics, GLCM and GLDV texture descriptors, and DSM-derived structural features contributed most to accuracy, while geometric descriptors showed marginal influence. The framework establishes a robust and economically accessible pathway for operational sago maturity monitoring.]]>
