<![CDATA[The decline in coconut production due to Weligama Coconut Leaf Wilt Disease (WCLWD) and Coconut Caterpillar Infestation (CCI) can be detected using deep learning. However, previous studies have largely ignored extreme data imbalance ratios, leaving models vulnerable to pseudo-accuracy and failure in recognizing minority classes. Furthermore, no existing studies on coconut disease classification have specifically evaluated model robustness against visual anomalies and background bias. To fill this gap, this study not only integrates the Xception architecture with the SMOTE oversampling technique to overcome imbalanced data but also conducts comprehensive stress testing. Using 5,139 images distributed in a 70:15:15 ratio, SMOTE was specifically applied to the training data. The model was optimized using a 299x299 resolution, a learning rate of 0.00001, and a 0.5 Dropout layer. Testing demonstrated optimal results with an overall accuracy of 99%. The implementation of SMOTE successfully handled data imbalance without sacrificing the sensitivity of the minority class (healthy leaves), evidenced by a 0.95 Recall and 0.82 F1-Score. Moreover, as a novel evaluation, testing using anomalous Out-of-Distribution images revealed a background bias in the CCI class. Nevertheless, the low predictive confidence level (43.06%) confirms that the model's regularization effectively prevents overconfident predictions and optimally calibrates visual uncertainty.]]>
Copyrights © 2026
