基于"点-面-时空"多源数据融合的作物病虫害多目标决策方法

Multi-objective Decision-making Method for Crop Diseases and Pests Based on Multi-source Data Fusion of "Point-Surface-Spatiotemporal" Framework

摘要:

针对作物病虫害智能防控中多源数据融合不足、多模型协同困难及工程化应用能力有限等问题，提出一种基于植物电子病历（PEMRs）的多源数据融合多目标决策方法。以电子病历为核心数据载体，融合结构化数据、文本数据与图像数据，构建"点-面-时空"多层次分析框架，实现从离散"点"诊断到区域"面"分析再到"时空"预测的连续决策支持。从"点"角度，针对结构化数据，构建Shared-MMoE多任务模型实现诊断与处方推荐协同优化;针对文本数据，基于BERT-CNN模型实现文本语义诊断;针对图像数据，结合EfficientNet-B3与ECA注意力机制构建轻量化病害识别模型;并结合知识图谱与大模型实现知识表达与解释推理。从"面"角度，基于电子病历数据进行时段筛选与区域聚合，通过热力图等方式展示病害空间分布。从"时空"角度，构建KAST-Graph时空预测模型，实现多区域作物病虫害发生趋势预测。在此基础上，开发面向实际应用的作物病虫害多目标决策平台，以微信小程序形式实现作物病虫害智能决策。试验结果表明：结构化诊断AUC达到96.33%，文本诊断准确率为93.13%，验证集病虫害识别准确率为85.95%;时空预测模型在MAE、RMSE和MAPE指标上均为最优;高频业务平均响应时间小于0.5s，多模态诊断一致性达到94.0%。研究结果表明，本文方法能够在轻量化部署条件下稳定运行，并有效支撑作物病虫害防控中的多目标决策需求。

Abstract:

Aiming to address the limitations of insufficient multi-source data fusion, weak multi-model collaboration, and limited engineering applicability in intelligent crop disease and pest management, a multi-objective decision-making method based on plant electronic medical records (PEMRs) was proposed.Taking PEMRs as the core data carrier, structured, textual, and image data were integrated to construct a "point-surface-spatiotemporal" framework, enabling continuous decision support from individual diagnosis to regional analysis and spatiotemporal prediction.At the "point" level, a Shared-MMoE model was developed for structured data to jointly optimize diagnosis and prescription recommendation;a BERT-CNN model was employed for semantic diagnosis of textual data;and a lightweight image recognition model based on EfficientNet-B3 with ECA attention was constructed.In addition, a knowledge graph combined with large language models was introduced for knowledge representation and interpretable reasoning.At the "surface" level, temporal filtering and regional aggregation of PEMR data were conducted to visualize spatial disease distribution.At the "spatiotemporal" level, a KAST-Graph model was developed to predict disease trends across multiple regions.Based on these, a practical multi-objective decision-making platform for crop pest and disease management was developed as a WeChat mini-program.Experimental results showed that the structured diagnosis achieved an AUC of 96.33%, text-based diagnosis accuracy reached 93.13%, and in the validation set pest and disease recognition accuracy of 85.95%.The proposed model outperformed baseline methods in MAE, RMSE, and MAPE.The system achieved an average response time of less than 0.5s, with a multi-modal diagnostic consistency of 94.0%.These results demonstrated that the proposed method operated stably under lightweight deployment conditions and effectively supported multi-objective decision-making in crop disease and pest management.