<![CDATA[Field-image classification of cassava leaf disease requires trustworthy probabilities and a mechanism for withholding uncertain decisions, rather than label accuracy alone. This study evaluates HSV, GLCM, and MobileNetV3-Small embedding fusion in a two-stage hierarchical Random Forest with probability calibration and selective prediction. A total of 7,508 CCMT images were evaluated using outer stratified 5-fold cross-validation. Model comparisons report mean ± standard deviation across five outer folds. End-to-end MobileNetV3-Small achieved 93.90±0.73% accuracy, 94.64±0.59% balanced accuracy, and 94.23±0.73% Macro F1, exceeding the Hierarchical Hybrid Random Forest at 88.28±1.12%, 89.66±0.76%, and 88.79±1.07%, respectively. The 1.03-percentage-point Macro F1 difference between the hierarchical and Flat Hybrid models was not significant at the 5% level (two-sided Wilcoxon, p=0.125). A separate calibration and selective-prediction experiment divided each outer-train fold into 70% model-fit, 15% calibrator-fit, and 15% threshold-selection partitions, all disjoint from the outer-test fold. Simplex-temperature yielded the lowest ECE (0.54%), whereas sigmoid performed better on log loss and Brier score. Thresholds selected on the held-out threshold-selection partition yielded 75.87% outer-test coverage, 95.40% accepted accuracy, and 95.58% accepted Macro F1. Selective accuracy applies only to accepted samples, not all data or guaranteed future performance; the contribution is evaluation of the coverage–risk trade-off, not improved intrinsic accuracy.]]>
