A multi-source fusion and feedback-optimized intelligent agent for crop disease and pest diagnosis and treatment

Crop diseases and pests pose a critical threat to global food security and agricultural sustainability. Traditional control methods are often limited by delayed diagnosis and a lack of capability for personalized solutions.

To address these challenges, we developed a knowledge-enhanced diagnostic and treatment agent, optimized through multi-source knowledge fusion and farmer feedback. The agent integrates disease identification results from a visual model, multi-factor contextual parameters, and a crop knowledge graph. These components form a unified multi-source knowledge representation. The system converts multi-criteria farmer evaluations into reward signals, enabling continuous optimization of action strategies through interaction with real-world environments. Under the combined guidance of multi-source knowledge and farmer feedback-driven reinforcement learning, the agent can generate accurate and practically applicable treatment recommendations without additional task-specific fine-tuning of the generative model.

Experiments on multiple baseline models demonstrate that combining these components consistently achieves the best performance. BERTScore increases by 25.23% on average, accuracy based on large language model evaluation improves by 30.27%, and practicality increases by 37.67%.

These results validate the effectiveness and generalization capability of the proposed method.