Explainable Disease Classification in Tomato Leaves Using Deep Learning

Abstract

This thesis addresses the challenge of detecting tomato leaf diseases using deep learning combined with Explainable Artificial Intelligence (XAI) techniques, and the development of a mobile application for practical field use. A convo- lutional neural network (CNN) model is trained on annotated image datasets to accurately classify multiple foliar diseases. To enhance the interpretability of the model’s predictions, we implement a modified version of the Score-CAM technique adapted for deployment on lightweight mobile environments, such as TensorFlow Lite. Due to technical constraints of on-device inference, our approach approximates the importance of activation maps by using their maxi- mum activation values as proxies, enabling efficient heatmap generation without the need for intermediate layer access or repeated inference. This pragmatic adaptation balances performance and explainability within the limitations of mobile platforms. The resulting system integrates the trained model into a user-friendly mobile application built with Flutter, offering offline disease de- tection and visual explanations. This work contributes a novel methodology for practical and interpretable plant disease diagnosis on resource-constrained devices, facilitating early detection and management in agriculture.