A Practical Vision Transformer and Retrieval based Advisory System for Rice Leaf Disease Management

Through this work, we present a practical AI-based advisory system for rice leaf disease management that combines a lightweight Vision Transformer (DeiT) for image-based disease classification with a retrieval-based knowledge module for farmer guidance. The system is designed to operate under real field conditions, where im-ages may contain background noise, lighting variation, and partial occlusions. Given an input leaf image, the model predicts the disease class along with a confidence score and optionally retrieves up-to-date agronomic information to generate actionable recommendations. Experiments con-ducted on a mixed dataset of curated and real-field images demonstrate strong classification performance, achieving 98.9% test accuracy. A qualitative evaluation of generated advisories with domain experts further indicates that the system provides contextually relevant and actionable guidance for smallholder farmers.