Efficient and Interpretable Deep Learning for Mango Leaf Disease Detection Using Knowledge Distillation and XAI

Abstract

Mango leaf diseases are a major cause of crop loss and can only be minimized through accurate and on-time disease identification. This paper proposes an effective and explainable deep learning model of mango leaf disease that makes use of logit-based Knowledge Distillation alongside temperature scaling and explainable artificial intelligence (XAI) strategies. The teacher network, DeiT-S, was pre-trained using a large collection of real-life and augmented images across eight disease categories with the validation accuracy of 99.14%. A small student model, MobileNetV3- small was then trained with a combined loss, where KL-divergence between the softened outputs of the teacher (soft targets, temperature T=4.0) and the ground-truth labels (hard targets) are combined with a cross-entropy loss. This method allowed the student to transfer the knowledge acquired by the teacher with high efficiency, so the training accuracy was 99.60%, and the validation was 99.46%. To allow better interpretability, Grad-CAM was used to visualize decision- making areas on leaf images. Lastly, an intuitive web-based application was created which could be used to upload leaf images and obtain predictions about the disease along with visual descriptions. The proposed framework is highly accurate and computationally efficient and transparent, which suggests its applicability in the practical implementation in agricultural disease management.