基于深度学习的小麦籽粒锌含量预测及安全利用分区

李清彩; 陈娟; 赵庆令; 蔡图; 韩文撑; 褚琳琳

文章摘要

李清彩,陈娟,赵庆令,蔡图,韩文撑,褚琳琳.基于深度学习的小麦籽粒锌含量预测及安全利用分区[J].农业环境科学学报,2024,43(10):2248-2259.

基于深度学习的小麦籽粒锌含量预测及安全利用分区

Prediction and safe utilization zoning of zinc content of wheat kernels based on deep learning

投稿时间：2024-02-02

DOI：10.11654/jaes.2024-0127

中文关键词: 深度学习多层感知机神经网络随机森林小麦锌安全利用

英文关键词: deep learning multi-layer perception neural networks（MLPNN） random forest （RF） wheat zinc safe utilization

基金项目:山东省重点研发计划项目(2020CXGC011403,2023CXGC010904)

作者	单位	E-mail
李清彩	山东省鲁南地质工程勘察院(山东省地质矿产勘查开发局第二地质大队), 山东济宁 272100 自然资源部采煤沉陷区综合治理与生态修复工程技术创新中心, 山东济宁 272100 山东农业大学资源与环境学院, 山东泰安 271018
陈娟	山东省鲁南地质工程勘察院(山东省地质矿产勘查开发局第二地质大队), 山东济宁 272100 自然资源部采煤沉陷区综合治理与生态修复工程技术创新中心, 山东济宁 272100
赵庆令	山东省鲁南地质工程勘察院(山东省地质矿产勘查开发局第二地质大队), 山东济宁 272100 自然资源部采煤沉陷区综合治理与生态修复工程技术创新中心, 山东济宁 272100 山东农业大学资源与环境学院, 山东泰安 271018	zqlzb@126.com
蔡图	济宁市自然资源和规划局, 济宁 272000
韩文撑	山东省鲁南地质工程勘察院(山东省地质矿产勘查开发局第二地质大队), 山东济宁 272100 自然资源部采煤沉陷区综合治理与生态修复工程技术创新中心, 山东济宁 272100
褚琳琳	山东省威海生态环境监测中心, 山东威海 264200

摘要点击次数: 1009

全文下载次数: 983

中文摘要:

为实现对小麦籽粒Zn含量的精准预测及安全利用分区，以济宁南部小麦种植区为研究对象，采集并测定了小麦籽粒中Zn及根际土壤样品中SiO₂、Fe₂O₃、MgO、CaO、Na₂O、K₂O、OrgC、P、N、S、Zn和pH等12种理化指标的含量，系统研究了小麦籽粒中Zn含量及其根际土壤理化指标含量特征，利用多层感知机神经网络和随机森林模型对小麦籽粒Zn含量变化特征进行预测，选择最优模型预测出济宁南部区域小麦籽粒Zn含量，并结合GIS技术划分了贫锌、缺锌、足锌和富锌农田。结果表明：济宁南部区域小麦籽粒中Zn含量平均值（39.7 mg·kg^-1）与富锌小麦籽粒推荐值基本持平，超出黄淮麦区小麦籽粒Zn平均含量1.32倍；经相关分析和聚类分析得出，小麦籽粒Zn与根际土壤理化指标之间相互作用、相互耦合，存在着较为复杂的非线性关系；多层感知机神经网络预测模型的R² （0.999）、RMSE（0.194）和MAE（0.146）等评价指标均优于随机森林模型；根际土壤中P、pH、OrgC和N指标是影响多层感知机神经网络预测相对重要的特征变量；研究区以足锌农田和缺锌农田为主，面积占比分别为57.47%和33.97%，谨慎利用贫锌区和安全利用富锌区农田面积占比分别为6.05%和2.51%。通过深度学习与农业地质相结合，利用多层感知机神经网络实现了通过简单土壤理化指标精准预测小麦籽粒锌含量。

英文摘要:

In order to achieve accurate prediction of Zn content in wheat grains and safe utilization zoning. Zn content of wheat grains and 12 physical and chemical indexes of rhizosphere soil samples including SiO₂, Fe₂O₃, MgO, CaO, Na₂O, K₂O, OrgC, P, N, S, Zn, and pH were determined in the wheat planting area of southern Jining. The characteristics of Zn content in wheat grains and the physical and chemical indexes of rhizosphere soils were systematically studied. The multi-layer perception neural networks（MLPNN）and random forest（RF） models were used to predict the variation characteristics of wheat grain Zn content. The optimal model parameters were selected to predict the Zn content of wheat grains. Zn-poor, Zn-deficient, Zn-sufficient, and Zn-rich farmlands were divided by GIS technology. The results showed that the average Zn content（39.7 mg·kg^-1）of wheat grains in the southern region of Jining was similar to the recommended value of Zn-rich wheat grains, which was 1.32 times higher than the average Zn content of wheat grains in the Huang-Huai wheat region. Through correlation analysis and cluster analysis, there was a complex nonlinear relationship between the interaction and coupling between Zn in wheat grain and the physical and chemical indexes of rhizosphere soil. R²（0.999）, RMSE（0.194）, and MAE（0.146）of the MLPNN prediction model were better than those of the RF model. P, pH, OrgC, and N in rhizosphere soil were relatively important characteristic variables affecting MLPNN prediction. The study area is dominated by Zn-sufficient and Zn-deficient farmlands, accounting for 57.47% and 33.97%, respectively. The proportions of Zn-poor farmland and safe use of Zn-rich area were 6.05% and 2.51%, respectively.

HTML 查看全文查看/发表评论下载PDF阅读器