Reliable mapping of sago stands is challenging in heterogeneous landscapes where other vegetation types exhibit similar spectral responses. This study compared support vector machine (SVM), random forest (RF), and gradient tree boosting (GTB) for binary sago stand mapping in the Arui watershed, West Papua, Indonesia, using 18 predictors derived from Sentinel-2 imagery and topographic data. This study represents the first application in Eastern Indonesia of Sentinel-2 imagery combined with SVM, RF, and GTB specifically for mapping sago palm stands across a watershed. Model selection used 1,934 pixels from 30 locations, with fivefold location-grouped validation applied consistently across algorithms. Predictor dependence was examined using permutation importance and fixed-parameter ablation, while additional analyses assessed location-allocation sensitivity, reference-label revisions, and spatial agreement among classification maps. SVM achieved the highest mean development F1 of 94.10%, followed by GTB (90.04%) and RF (88.55%), although its historical-test F1 was lower at 86.96%. Permuting B6 reduced SVM F1 by 53.39 % points, whereas removing B6 and retraining reduced it by only 1.16 % points, indicating that permutation sensitivity did not represent its unique contribution after retraining. In the map assessment, reference-label revisions increased SVM F1 from 75.00% to 86.27% without changing model predictions. The horizontal co-registration RMSE between Sentinel-2 imagery and the UAV orthomosaic was 9.18 m (0.92 of a 10 m pixel). The SVM map identified 803.57 ha of sago within the common valid area of 22,236.25 ha, while the three models disagreed across 512.65 ha. Together, these analyses show how validation design, predictor redundancy, reference-label sensitivity, and algorithm choice influence the interpretation of machine-learning-based sago distribution maps. However, limited geographic reference coverage, positional uncertainty, and the absence of independent accuracy and area-of-applicability assessments constrain conclusions about watershed-wide mapping accuracy. Mapped areas therefore represent unadjusted pixel-count summaries rather than statistically supported estimates of actual sago area.