Enhancing precision in rice leaf disease detection: A transformer model approach with attention mapping

Abstract

Advancements in image recognition technology have significantly impacted agricultural practices, especially in early detection of rice leaf diseases, which is crucial for maintaining crop health and yield. This paper introduces the Optimized BEiT model, a novel lightweight CNN and transformer architecture, specifically designed for this purpose. The model demonstrates high accuracy, outperforming traditional models with a precision of 0.92, recall of 0.91, and F1-score of 0.91. It was rigorously trained and validated on a comprehensive dataset featuring healthy and unhealthy rice leaf images. The use of attention mapping techniques such as GRAD-CAM has been pivotal in interpreting the model's predictions, enhancing the transparency and reliability of AI in agricultural diagnostics. These techniques provide clear, comprehensible insights into the predictive features of the model, ensuring its decisions are understandable and trustworthy. This research is a significant stride in precision agriculture, offering a robust tool for agricultural professionals. This model not only represents a technical achievement but also a practical solution for real-world agricultural challenges, demonstrating the potential of AI to enhance food security and sustainable farming practices.