关键词: 玉米, 精准施肥, 知识图谱, 大语言模型, 智能体

Abstract:

[Objective] Maize precision fertilization must consider target yield, soil nutrient supply, cultivar characteristics, regional conditions, and management practices. However, relevant knowledge is scattered across variety approval bulletins, technical guidelines, expert rules, and historical experiments. Data-driven models can estimate fertilizer rates, but they often lack explicit agronomic knowledge and traceable evidence. Traditional case-based reasoning can reuse similar cases, but its stability is limited when target fields differ from existing records. Therefore, this study aimed to construct a Knowledge-Enhanced Decision Agent Framework (KEDAF), integrating a knowledge graph, historical cases, and agronomic rules to improve the accuracy, adaptability, interpretability, and traceability of nitrogen (N), phosphorus (P), and potassium (K) recommendations. [Methods] Maize variety approval bulletins, provincial fertilization guidance documents, agronomic rules, expert experience, and historical fertilization cases were used as the main data sources. Historical cases were collected from public literature on maize fertilization experiments, nutrient management, soil testing, target-yield fertilization, and fertilizer treatment effects. After screening records with clear locations, cultivar information, soil indicators, yield information, and fertilizer application data, 815 cases covering the six major maize planting regions in China were retained. For knowledge graph construction, Qwen2.5-1.5B-Instruct was used as the base model, and Low-Rank Adaptation (LoRA) was adopted for entity extraction. Knowledge triples were generated through entity-type mapping. After cleaning, normalization, and fusion, the maize fertilization knowledge graph contained 63 096 nodes and 277 906 relations. In the decision stage, DeepSeek-V3.2 was used as the language reasoning and interaction engine. User input was parsed into a field state vector containing cultivar, province, ecological region, target yield, soil organic matter, alkali-hydrolysable nitrogen, available phosphorus, available potassium, and pH. Based on this state, KEDAF activated normative memory from the knowledge graph and regional rules, experiential memory from the historical case database, and safety memory from the agronomic rule base. Candidate schemes were generated from knowledge graph priors, adapted similar cases, and fused actions. Similar cases were retrieved by combining categorical semantic similarity and numerical similarity, and their fertilizer rates were revised according to differences in target yield, soil nutrients, organic matter, and pH. Candidate schemes were evaluated by an internal critic using knowledge consistency, case support, and agronomic risk. When fertilizer amounts or stage-wise ratios exceeded safe boundaries, projection-based correction was performed. Experiments used an 8:2 train-test split. Rule-based, traditional case-based reasoning (CBR), and CBR-ML methods were selected as baselines. [Results and Discussions] KEDAF achieved an overall MAPE of 5.337 0% and an R² of 0.882 0 for total N, P, and K prediction. For stage-wise allocation, the R² reached 0.944 4. At the nutrient level, phosphorus prediction was the most stable, with a MAPE of 4.046 9% and an R² of 0.888 8, followed by nitrogen with a MAPE of 4.203 4% and an R² of 0.882 4. Potassium prediction was relatively weaker, with a MAPE of 7.760 8% and an R² of 0.874 7. Compared with the rule-based method, KEDAF reduced the total MAPE from 9.957 0% to 5.337 0% and the stage-wise MAPE from 8.540 7% to 3.444 6%. Compared with traditional CBR and the CBR-ML hybrid model, KEDAF also performed better. Ablation experiments showed that the knowledge graph prior provided a stable decision anchor, case adaptation improved field-specific adjustment, and safety constraints enhanced recommendation feasibility. [Conclusions] KEDAF integrated structured knowledge, historical experience, and agronomic constraints into a traceable fertilization decision process. Offline experiments showed that it outperformed rule-based, CBR, and CBR-ML baselines. The framework provided evidence-supported recommendations for maize precision fertilization, but further field validation and integration of weather, soil monitoring, remote sensing, and management data are still needed.

Key words: maize, precision fertilization, knowledge graph, large language model, agent