ChatLeafDisease: a chain-of-thought prompting approach for crop disease classification using large language models

Area of Science:

• Agricultural Science
• Artificial Intelligence
• Plant Pathology

Background:

• Accurate crop disease classification is vital for food security and effective disease management.
• Deep learning models require substantial training data, limiting their application for diverse crops.
• Large language models (LLMs) offer potential for zero-shot learning but their use in crop disease classification is underexplored.

Purpose of the Study:

• To develop a novel, training-free framework for crop disease classification utilizing LLMs.
• To evaluate the efficacy of the proposed framework against existing models for tomato disease identification.
• To assess the scalability and performance of the framework for new crop diseases.

Main Methods:

• Developed ChatLeafDisease (ChatLD), a framework based on GPT-4o with chain-of-thought (CoT) prompting.
• Integrated a disease description database and a classification agent guided by CoT prompts.
• Compared ChatLD performance against GPT-4o, Gemini, and Contrastive Language-Image Pre-training (CLIP) models.

Main Results:

• ChatLD achieved 88.9% accuracy for six tomato diseases, significantly outperforming GPT-4o (45.9%), Gemini (56.1%), and CLIP (64.3%).
• Scoring rules within CoT prompts were crucial for capturing disease-specific differences and enhancing accuracy.
• Condensed disease descriptions improved classification performance, and ChatLD demonstrated high accuracy for new crop diseases, indicating scalability.

Conclusions:

• The ChatLD framework offers a highly accurate and scalable LLM-based solution for crop disease classification.
• This training-free approach, leveraging textual descriptions, overcomes data limitations of traditional deep learning methods.
• ChatLD presents a promising alternative for agricultural disease identification, supporting global food security efforts.