Mapping sago stands reliably is difficult where other vegetation produces similar spectral responses. We examined this problem in the Arui watershed of West Papua, Indonesia, by comparing support vector machine (SVM), random forest (RF), and gradient tree boosting (GTB) classifiers. The predictors comprised ten Sentinel-2 bands, six spectral indices, elevation, and slope, all represented on a 10 m grid. Model selection drew on 1,934 pixels from 30 locations, and every algorithm was evaluated with the same five folds grouped by location. Predictor importance was assessed with 30 permutations for each predictor in every fold. We also removed predictors and retrained the models without changing their parameter settings. Mean F1 for the sago class was 94.10% for SVM, 90.04% for GTB, and 88.55% for RF. When B6 was permuted, SVM F1 fell by 53.39 percentage points; after B6 was removed and the model retrained, the decrease was only 1.16 points. Thus, although the fitted SVM relied heavily on B6, much of its information was recoverable from the remaining inputs. The SVM recorded an F1 of 86.96% on the historical test. In a separate assessment of 644 pixels, auditing the reference labels increased its F1 from 75.00% to 86.27% without changing any prediction; consequently, this audit was not treated as a new independent test. Within the raster area shared by the three models, the unadjusted SVM map identified 803.57 ha of sago from pixel counts. All models agreed on 643.83 ha, but their classifications differed over 512.65 ha. Under the local selection procedure, SVM gave the best result. Its wider use should nevertheless be considered cautiously because the predictors were redundant, some reference data had been examined previously, and the sampled locations covered only part of the watershed.