China has unveiled its first open-source large language model dedicated to crop protection, marking a significant step in applying artificial intelligence to agricultural safety and productivity. The model, named Green Shield, was developed by Nanjing Agricultural University in collaboration with the National Key Laboratory of Agricultural Biosafety and more than 30 industry partners, according to a report published by Science and Technology Daily.
The launch comes as China faces recurring pest outbreaks and growing challenges linked to pesticide resistance, issues that have increasingly strained agricultural output and sustainability. Project leader Dong Shameng, who also serves as vice dean of the university’s College of Plant Protection, noted that farmers urgently require accurate, professional guidance at the grassroots level. He pointed out that general-purpose large language models often fail to deliver reliable answers in specialized agricultural contexts and may even produce inconsistent or potentially unsafe pesticide recommendations.
To address these shortcomings, the research team constructed a highly specialized dataset comprising more than 2.5 billion tokens drawn from academic literature, patents, national standards, and field reports. The dataset spans major crop categories such as rice, wheat, soybeans, vegetables, and fruit trees, while integrating pest monitoring data, environmentally friendly control measures, and pesticide registration information.
According to Wang Dongbo, a professor at the university’s College of Information Management, the model is capable of accurately identifying crop types, growth stages, and disease symptoms before generating comprehensive treatment strategies. Through targeted training, the system achieves strong convergence and demonstrates high precision in pest recognition, enabling more reliable decision-making in the field.
A key safety feature of Green Shield lies in its automated compliance mechanism. Before issuing any recommendation, the model cross-checks inputs against the national pesticide registration database to ensure that suggested chemicals meet regulatory requirements, including approved usage, crop compatibility, and dosage limits. Any non-compliant output is automatically blocked and corrected, reducing the risk of pesticide misuse at the source.
Wang Yuanchao, vice president of Nanjing Agricultural University, said the institution will continue refining the model through field testing and iterative upgrades. The goal is to build an intelligent agricultural tool that is accessible, practical, and effective for farmers, ultimately supporting the digital transformation of modern agriculture across the entire production chain.