<![CDATA[Pestalotiopsis sp.-induced leaf-fall disease threatens rubber (Hevea brasiliensis) productivity by reducing canopy density, degrading chlorophyll, and impairing photosynthetic capacity. This study developed an exploratory multiclass framework to classify disease severity in rubber plantations using UAV multispectral data and a Random Forest (RF) algorithm. The dataset comprised 144 ground-truth canopy-level samples labeled into four ordinal severity classes, L1 (mild), L2 (moderate), L3 (severe), and L4 (very severe), and 1760 additional UAV-derived observations for spatial inference. Spectral and structural predictors included the Leaf Chlorophyll Index (LCI), Normalized Difference Red Edge (NDRE), density_LCI, and density_NDRE, collected across three rubber clones, BPM 24, GT 1, and RRIC 100. Model development used an 80:20 stratified train–test split, stratified and nested cross-validation, class weighting, and SMOTE applied only within training folds to reduce imbalance and leakage. Performance was evaluated using accuracy, class-wise precision–recall–F1, and the multiclass Matthews Correlation Coefficient (MCC). Hyperparameter tuning (n_estimators = 150, max_depth = 7, min_samples_split = 9) increased MCC from 0.143 to 0.547 on independent test data and 0.586 under the all-labeled-data scenario, indicating moderate agreement. LCI emerged as the dominant predictor, supporting the physiological relevance of chlorophyll-sensitive red-edge information. However, low recall for L2 and unstable L4 performance reflected class imbalance, limited labeled samples, and overlapping spectral responses. The framework provides a more informative severity stratification than binary detection, but further ground-truth expansion, ordinal or cost-sensitive modeling, comparative benchmarking, and multi-season validation are required before operational deployment.]]>
