Unlike conventional models, ChatLD relies solely on textual descriptions of disease symptoms, achieving remarkable accuracy (88.9%) in classifying six tomato diseases. This innovation opens the door to scalable, cost-efficient, and cross-crop disease identification for real-world agricultural applications.
Crop diseases and pests are responsible for up to half of yield losses in developing regions, threatening global food supply chains. While deep learning has improved image-based disease recognition, its dependence on massive labeled datasets restricts its adaptability to diverse crop types and field conditions. Moreover, fine-tuning these models for each new scenario is time-consuming and data-intensive. Large language models such as GPT-4 and Gemini have demonstrated strong reasoning and generalization skills in fields like medicine and finance, but their potential in agriculture remains underexplored. Prompt engineering—especially the CoT method—has emerged as a powerful approach to guide LLM reasoning without retraining. Motivated by these challenges, the research team sought to design a framework that leverages LLMs’ zero-shot reasoning to classify crop diseases from images using only textual knowledge, without the need for retraining or large datasets.
A study (DOI: 10.1016/j.plaphe.2025.100094) published in Plant Phenomics on 7 August 2025 by Tao Lin’s team, Zhejiang University, introduces ChatLeafDisease, a training-free large language model framework that achieves high-accuracy, cross-crop disease classification, offering a scalable and data-efficient solution for intelligent agricultural disease diagnosis.
The study employed a training-free large language model (LLM)-based framework, named ChatLD, built on GPT-4o with Chain-of-Thought (CoT) prompting to classify crop diseases from leaf images using only textual disease descriptions. The method combined a disease description database and a CoT-guided reasoning agent that scored the correspondence between image patterns and disease features, allowing step-by-step logical classification without retraining or fine-tuning. Comparative experiments evaluated ChatLD against GPT-4o, Gemini-1.5-pro, and CLIP models on tomato disease datasets. The ChatLD framework achieved the highest and most stable accuracy of 88.9%, significantly outperforming GPT-4o (45.9%), Gemini (56.1%), and CLIP (64.3%). The CoT-based scoring rules enhanced reasoning ability, reducing misclassification among visually similar diseases. For instance, ChatLD correctly identified more than 88% of Late Blight, Mosaic Virus, and Yellow Leaf Curl Virus samples, though similarities between Early Blight and Late Blight still caused minor confusion. Ablation experiments confirmed that removing scoring rules dropped accuracy from 90.3% to 51.8%, emphasizing their role in logical reasoning, while concise and clear disease descriptions improved accuracy by over 40%. Beyond tomato, ChatLD demonstrated remarkable zero-shot generalization, reaching 94.4% average accuracy across new crops such as grape, strawberry, and pepper—exceeding the fine-tuned CLIP model trained on up to 50 samples per class. When tested on real-world field data (PlantSeg dataset), ChatLD maintained solid performance with 77.3% accuracy, despite challenges from complex backgrounds and overlapping leaves. Overall, the method’s integration of CoT reasoning, structured scoring, and refined textual inputs enabled ChatLD to outperform traditional deep-learning models in both accuracy and scalability, confirming its potential as a universal, data-efficient tool for intelligent crop disease classification.
ChatLD represents a major step toward data-efficient digital agriculture. Its training-free, text-driven architecture drastically reduces the need for labeled image datasets, making it ideal for developing regions where data scarcity is a barrier. The model can be seamlessly scaled to new crops by adding textual disease descriptions, offering immediate diagnostic support for farmers and agronomists. Beyond classification, ChatLD could serve as a foundation for intelligent disease management systems, integrating multimodal inputs such as environmental data and time-series crop growth indicators. This would allow comprehensive analysis of disease causes, enabling real-time monitoring and precision treatment recommendations.
###
References
DOI
10.1016/j.plaphe.2025.100094
Original URL
https://doi.org/10.1016/j.plaphe.2025.100094
Funding information
This work was supported by Key R&D Program of Zhejiang Province under Grant Number: 2022C02003 and China National Key Research and Development Plan under Grant Number: 2022YFD2002303.
About Plant Phenomics
Plant Phenomics is dedicated to publishing novel research that will advance all aspects of plant phenotyping from the cell to the plant population levels using innovative combinations of sensor systems and data analytics. Plant Phenomics aims also to connect phenomics to other science domains, such as genomics, genetics, physiology, molecular biology, bioinformatics, statistics, mathematics, and computer sciences. Plant Phenomics should thus contribute to advance plant sciences and agriculture/forestry/horticulture by addressing key scientific challenges in the area of plant phenomics.