Cotton Leaf Disease Detection Using Advanced Ensemble Learning Techniques

Abstract

Cotton is a major basic agricultural commodity in many countries. It is negatively affected by leaf diseases such as bacterial blight, curl virus, and fusarium wilt, which can reduce yields by up to 20–30%. Identifying these diseases easily, accurately, and promptly is critical to managing the crop. This study proposes advanced deep learning-based models for automated multi-class classification of cotton leaf diseases using image data. A custom CNN was implemented and evaluated, along with the MobileNetV2 and EfficientNetB0 architectures. EfficientNetB0 was used to assess the robustness of the prediction through test-time augmentation (TTA). A stacking ensemble model was also tested, which included the outputs of the three base models fed into logistic regression as a meta-learner. The proposed approach was evaluated using accuracy, precision, recall, and the F1 score. The custom CNN produced an accuracy of 94% with adequate class-wise performance for the three classes. MobileNetV2 and EfficientNetB0 achieved accuracies of 90% and 89%, respectively. The TTA improved the performance of EfficientNetB0 to 93.71%. The stacking ensemble achieved the best accuracy of 97.90%, with class-wise precision, recall, and F1 scores reaching near perfection, confirming that ensemble learning is beneficial for classifying plant diseases. In general, a high-accuracy, scalable approach was developed that could provide significant support to precision agriculture and reduce crop loss due to fungus-based diseases using US-based devices.