<![CDATA[Oil palm disease classification is essential for precision agriculture, but its performance is often limited by class imbalance, noisy samples, and overlapping feature distributions. Conventional Convolutional Neural Network (CNN)-based models provide strong feature extraction capabilities, yet they tend to be biased toward majority classes and are sensitive to unreliable training data. Oversampling methods such as the Synthetic Minority Over-sampling Technique (SMOTE) improve class balance, but may amplify noise because synthetic samples are generated without considering sample reliability. This study proposes an uncertainty-guided noise-aware oversampling framework integrated with a deep CNN for imbalanced oil palm disease classification. Aleatoric uncertainty is used to identify and remove noisy samples, while epistemic uncertainty guides synthetic data generation toward informative regions near decision boundaries. A sensitivity analysis identifies the most effective configuration, consisting of 20 forward passes, a dropout rate of 0.5, an aleatoric uncertainty threshold of 0.4, and prediction variance as the epistemic uncertainty measure. Experimental results show that the proposed framework achieves the highest validation accuracy of 92.88% and consistently outperforms CAE, LCNN, CNN, Deep SVM, and SCNN across Accuracy, Precision, Recall, Specificity, F1-Score, Balanced Accuracy, and Cohen’s Kappa. The largest improvements are observed in the minority classes, particularly Ganoderma Initial Infection and Ganoderma Infected. Wilcoxon signed-rank testing confirms that all improvements are statistically significant (p < 0.05). Computational analysis indicates only a moderate and practically insignificant increase in training time, memory usage, and computational complexity. These findings demonstrate that the proposed framework provides an effective and robust solution for imbalanced agricultural image classification.]]>
Copyrights © 2026
