LiteViT: Leveraging the Power of Transformers for Edge AI in Crop Disease Classification

Download PDF
Open Webpage

Sai Mahesh Mudavat, Alakananda Mitra, Saraju P. Mohanty, Elias Kougianos

Published: 2025, Last Modified: 06 May 2026ISVLSI 2025EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Plant diseases have a significant impact on crop yield, presenting a serious challenge to food security. Despite advances in Artificial Intelligence (AI) for improved disease detection, real-world implementation remains limited due to high computational demands. LiteViT bridges this gap through a proposed Knowledge distillation framework that transforms powerful but computationally heavy Vision transformers (ViTs) into a field-ready tool by distilling the knowledge of a $\mathbf{3 0 0}$ million parameter ViT large teacher into a lightweight MobileViT-XXS (extra-extrasmall) model of size 3.8 MB, achieving 99.3% accuracy while retaining nearly identical performance compared to the teacher’s $\mathbf{9 9. 7 \%}$. The framework integrates a multimodal explainability framework that visually interprets model predictions to enhance interpretability. This framework demonstrates accurate and explainable plant disease detection, suitable for edge devices; thereby bridging the gap between laboratory and field-level deployment and advancing smart agriculture.