Hybrid Large Language Model-quantitative model fusion framework for consistency tuning of open-field white radish planting-to-harvesting

In recent years, the mechanization of open-field white radish production in China has advanced significantly. Nevertheless, fully mechanical operations still encounter critical challenges in the planting-to-harvesting stage, particularly insufficient cross-stage adaptability and poor coordination of operational parameters. To overcome these limitations, taking open-field white radish as an example, a hybrid planting-harvesting consistency tuning framework integrated with a Large Language Model (LLM, and Qwen3–8B as the base model in this work) and several Quantitative Models (QMs) is proposed. Retrieval Augmented Generation (RAG) mechanism is used for LLM tuning, and a Consistency-Guided RAG (CG-RAG) architecture is proposed to reduce the hallucinations generated by the LLM. To provide the support for the feature extraction through satellite maps, an improved Unet model EA-MSAtt-UNet is proposed for high-precision image segmentation and feature extraction. All the information provided by the fine tuned CG-RAG LLM are delivered to four QMs through our carefully designed interface, and finally the consistency decisions are provided by the designed four QMs - a ridge-line planning model, a machinery-agronomy matching model, a planting-harvesting consistency model, and a maximum profit model - strictly constrained by the knowledge by LLM outputs. This hybrid mechanism and framework substantially enhanced decision reliability and mitigates hallucination risks inherent to standalone LLMs. Experimental results demonstrate that the CG-RAG LLM delivers superior performance on Qwen3–8B, attaining an LLM-Metric of 95.12 ± 0.88 and an Expert Acceptability Rate (EAR) of 91/100. Following integration of the QMs, the system achieves scores of 38, 35, and 49 (on a 50-question test set) for machinery-agronomy executability, planting-harvesting consistency, and profit calculation accuracy, respectively. A seeding-stage field-based calculation example further demonstrates the preliminary applicability of the framework for field identification, ridge-line planning, machinery recommendation, and projected profit estimation. Historical comparison with previous-year field data under similar conditions indicates an approximate 3% error level in the calculation chain, providing a reference for its practical reliability. It indicates that the proposed LLM-QM fusion approach provides robust technical support for collaborative planting-harvesting decision-making, paving the way for more reliable and efficient mechanical production systems in open-field white radish cultivation.