An Explainable Artificial IntelligenceDriven Minimal Net for Tea Leaf Disease Detection

Abstract

Tea leaf diseases pose a significant threat to crop yield and quality, especially in regions where agriculture is a key economic sector. Accurate and real-time detection of tea leaf diseases remains a significant challenge due to limitations in existing automated approaches, including heavy model architectures, a lack of interpretability, and poor adaptability in low-resource environments. This study introduces a comprehensive and scalable framework for the automated detection of tea leaf diseases, integrating deep learning, explainable AI, and lightweight deployment mechanisms. Initially, five fine-tuned Convolutional Neural Networks (CNNs) and five transfer learning models were evaluated, and MobileNetV2 and VGG16 achieved the highest accuracies of 99.52% and 99.68%, respectively. However, VGG16 in the CNN stage and Xception in the transfer learning setting had relatively lower accuracy rates of 88.94% and 73.50%, respectively, which indicates the need for a more balanced and efficient approach. To address this, a lightweight yet high-performing CNN model, MNet, was developed. With only 49,292 parameters, M-Net achieved a competitive accuracy of 97.29%, along with 97.12% precision and recall, and a low false alarm rate of 0.54%, making it ideal for real-time field deployment. The model's transparency was enhanced through the integration of LIME and SHAP that offer clear visual justifications for predictions. Finally, M-Net was successfully deployed as both a mobile application and a Streamlit-based web app, each capable of detecting all disease classes with high confidence. This work demonstrates the feasibility of deploying intelligent, explainable, and resourceefficient AI solutions in agricultural diagnostics to support smart farming and timely disease management.