How AI and Adaptive Data Boost Betel Leaf Disease Detection to 98% Accuracy in Bangladesh

Dhaka, Bangladesh, Source:

Betel leaf stands as a vital pillar for traditional medicine and rural livelihoods across South and Southeast Asia. Yet this essential crop faces persistent threats from recurring diseases such as leaf rot and leaf spot. These pathogens inflict significant yield losses and economic damage, particularly affecting smallholder farmers in Bangladesh. While deep learning presents a promising avenue for automated disease detection, its practical application often encounters hurdles due to limited training data and class imbalance issues.

Traditional manual inspection methods remain labor-intensive and subjective. This reliance on human observation frequently leads to delayed diagnoses and inconsistent pesticide application. Previous attempts at using deep learning have struggled with small datasets. Many of these earlier models relied on fixed augmentation techniques that failed to adapt to model performance during the training process. This new approach reframes augmentation policy selection as a multi-armed bandit problem. A UCB controller dynamically selects between different RandAugment magnitudes and Mixup activations based on validation performance. This strategy balances exploration and exploitation to identify the most effective data transformation strategy.

The study employs a strict evaluation protocol where only original images are used for dataset partitioning. This method prevents data leakage that can artificially inflate performance metrics. The ConvNeXt-Tiny backbone was chosen for its efficiency and accuracy. It was tested against other models including ResNet, DenseNet, and EfficientNet. Results indicate that ConvNeXt-Tiny with 12-arm UCB-guided augmentation achieved an impressive 98.04% accuracy. The macro-F1 score reached 97.87%. This performance significantly outperforms fixed augmentation methods and other backbones.

Research also demonstrated robustness under increasing class imbalance. The model maintained high performance even when minority class samples were reduced. By integrating explainable AI techniques like Grad-CAM, the model provides visual evidence that it focuses on relevant leaf features rather than background artifacts. This framework enhances diagnostic accuracy while offering a scalable and efficient solution for resource-constrained agricultural environments. It paves the way for smarter and more resilient farming practices through advanced artificial intelligence.