基于机器学习的土壤水分地面观测站数据修复方法研究  

Research on Data Restoration Method of Soil Moisture Ground Observation Station Based on Machine Learning

作　　者：李孟 宋承运[1,2] 孙时雨 LI Meng;SONG Cheng-yun;SUN Shi-yu(School of Geomatics,Anhui University of Science and Technology,Huainan 232001,Anhui Province,China;Key Laboratory of Aviation-Aerospace-Ground Cooperative Monitoring and Early Warning of Coal Mining-induced Disasters of Anhui Higher Education Institutes,Huainan 232001,Anhui Province,China)

机构地区：[1]安徽理工大学空间信息与测绘工程学院,安徽淮南232001 [2]矿山采动灾害空天地协同监测与预警安徽普通高校重点实验室,安徽淮南232001

出　　处：《节水灌溉》2024年第11期89-96,共8页Water Saving Irrigation

基　　金：矿山采动灾害空天地协同监测与预警安徽普通高校重点实验室(安徽理工大学)开放基金项目(KLAHEI202205);安徽理工大学研究生创新基金项目(2023cx2174)。

摘　　要：土壤水分站点易受人为破坏、自然灾害和设备故障等因素的影响,出现数据不同程度的缺失,直接影响了其在农业生产、气象监测、生态环境等学科领域的深入应用。利用闪电河流域土壤温湿度无线传感器网络的土壤水分观测数据,采用机器学习方法,以MODIS地表温度、植被指数、地表反照率、热惯量以及地表高程等为参量,研究土壤水分站点缺失数据修复方法。研究结果表明:(1)随机森林表现出较高的精度(相关系数r=0.95,均方根误差RMSE=0.023 m^(3)/m^(3),无偏均方根误差ubRMSE=0.023 m^(3)/m^(3),偏差Bias=-0.001)优于广义回归神经网络;(2)在低植被覆盖区域,模型拟合效果优于高植被覆盖区域;(3)随机森林模型在站点不同缺失情况下,表现出了较高的精度(相关系数r>0.8,均方根误差RMSE≤0.038 m^(3)/m^(3),无偏均方根误差ubRMSE≤0.038 m^(3)/m^(3),偏差Bias≤0.018),且在部分时间段缺失下模型精度更高,很好地反映了土壤水分随季节变化的趋势。该研究为地面站点数据的修复以及地面站点的布设提供参考与支持。Soil moisture monitoring sites are vulnerable to factors such as human damage,natural disasters and equipment failures,resulting in varying degrees of data loss,which directly affects their application in agricultural production,meteorological monitoring,and ecological environment.This study uses the soil moisture observation data of a wireless sensor network of soil temperature and humidity in the Shandian River Basin,adopts machine learning methods,and uses MODIS surface temperature,vegetation index,surface albedo,thermal inertia,and surface elevation as parameters to study the method of repairing missing data of soil moisture stations.The results show that:①Random forests have higher accuracy(correlation coefficient r=0.95,root mean square error RMSE=0.023 m^(3)/m^(3),unbiased root mean square error ubRMSE=0.023 m^(3)/m^(3),bias Bias=-0.001)than generalized regression neural networks;②In areas with low vegetation coverage,the model fitting effect is better than that in areas with high vegetation coverage;③The random forest model has higher accuracy when the site data is missing at different levels(correlation coefficient r>0.8,root mean square error RMSE≤0.038 m^(3)/m^(3),unbiased root mean square error ubRMSE≤0.038 m^(3)/m^(3),Bias≤0.018),showing even higher accuracy during periods of partial data loss,which well reflects the trend of soil moisture changing with seasons.This study provides reference and support for the restoration of ground site data and the layout of ground sites.

关 键 词：站点土壤水分 机器学习 数据修复 

分 类 号：S152.7[农业科学—土壤学] TP79[农业科学—农业基础科学]