Machine Learning and Deep Learning Techniques for Betel Leaf Disease Detection in Imbalanced Datasets

Abstract

The cultivation of betel leaf (Piper betle) is an important supplementary livelihood for many farmers in Sri Lanka’s wet zone, particularly in the districts of Gampaha, Kalutara, and Colombo. However, the productivity and quality of betel leaves are frequently threatened by diseases such as bacterial leaf blight, fungal infections, pest attacks, and mite infestations. Timely and accurate identification of these issues is critical, yet traditional diagnostic methods used by farmers are often outdated, inefficient, or inaccessible. This research project addresses these challenges by developing an automated image classification system using both machine learning and deep learning approaches. A dataset of betel leaf images was collected and categorized into four classes: bacterial leaf blight, pest attack, mites attack, and healthy leaves. Using this dataset, five traditional machine learning models (Decision Tree, LightGBM, Random Forest, Support Vector Machine, and XGBoost) and four deep learning models (VGG16, VGG19, ResNet50, and ResNet101) were trained using consistent preprocessing, augmentation, and training strategies. Model performance was evaluated using balanced accuracy, macro F1-score, and AUC-ROC to ensure fair and robust comparison. Among all models tested, ResNet50 emerged as the top performer, achieving the highest scores across all evaluation metrics. This best-performing model was then integrated into a user-friendly web application, allowing real-time disease prediction from uploaded leaf images. The study demonstrates how deep convo lutional neural networks, when paired with proper preprocessing techniques, can significantly enhance plant disease classification accuracy. The proposed solution offers a scalable and practical tool to support early detection and disease management, thereby promoting precision agriculture in Sri Lanka’s betel leaf farming communities.