<![CDATA[Leaf-fall disease caused by Pestalotiopsis sp. is a major threat to rubber (Hevea brasiliensis) plantations because it suppresses photosynthetic activity, accelerates defoliation, and reduces latex productivity. In operational practice, severity assessment is still dominated by visual field inspection, which is subjective, time-consuming, costly, and difficult to standardize across large plantation areas. This study develops a disease severity classification model for Pestalotiopsis sp. using a Convolutional Neural Network (CNN) based on vegetation-index features derived from UAV multispectral imagery. The model classifies disease severity into four levels: L1 (Light Infection), L2 (Moderate Infection), L3 (Severe Infection), and L4 (Very Severe Infection). To represent temporal and biological variability in disease expression, multispectral data were collected from multiple rubber clones over two observation periods. Feature construction focused on NDRE, LCI, CI, NDVI_NDRE_Interaction, and GCI_Ratio, which capture chlorophyll-related and canopy condition responses to infection. Because severity classes were imbalanced, the Synthetic Minority Over-sampling Technique (SMOTE) was applied before model training. A one-dimensional CNN was then trained to learn nonlinear patterns among index-based predictors for multilevel severity classification. Hyperparameter tuning improved overall accuracy from 85.30% to 90.00%. Class-wise F1-scores changed from 0.91 to 0.94 (L1), 0.83 to 0.84 (L2), 0.75 to 0.88 (L3), and 0.97 to 0.84 (L4), with the largest improvement in L3 recall (0.67 to 0.94). These results indicate that the selected vegetation indices and interaction terms are informative predictors for objective and scalable disease severity classification under heterogeneous plantation conditions.]]>
Copyrights © 2026
