基于无人机多光谱遥感的玉米根域土壤含水率研究

Retrieving Soil Moisture Content in Field Maize Root Zone Based on UAV Multispectral Remote Sensing

及时获取农田作物根域土壤墒情是实现精准灌溉的基础和关键。以内蒙古自治区达拉特旗昭君镇试验站大田玉米为研究对象,利用无人机遥感系统,分别在玉米营养生长期(Vegetative stage,V期)、生殖期(Reproductive stage,R期)和成熟期(Maturation stage,M期)获得7次玉米冠层多光谱正射影像,并同步采集玉米根域不同深度土壤含水率(Soil moisture content,SMC);然后,采用灰色关联法对提取的多种植被指数(Vegetation index,VI)进行筛选,选取与土壤含水率敏感的植被指数;最后,分别采用多元混合线性回归(Cubist)、反向传播神经网络(Back propagation neural network,BPNN)和支持向量机回归(Support vector machine regression,SVR)等机器学习方法,构建不同生育期的敏感植被指数与土壤含水率的关系模型。结果表明,3种机器学习方法中SVR模型在各生育期的建模与预测精度均最优,BPNN模型次之,Cubist模型最差;其中SVR模型在M期效果最优,其建模集和验证集R^2分别为0.851和0.875,均方根误差(Root mean square error,RMSE)均为0.7%,标准均方根误差(Normalized root mean square error,nRMSE)分别为8.17%和8.32%,R期效果最差,其建模集和验证集R^2分别为0.619和0.517。

Rapid acquisition of soil moisture content(SMC)in crop root zone is the key to drought supervision and precision irrigation.The relationship between the unmanned aerial vehicle(UAV)multispectral remote sensing and SMC was mainly studied based on the field maize data of experimental station in Zhaojun Town,Dalate Qi,Inner Mongolia.The canopy images of field maize with five irrigation treatments were obtained at different growth stages(vegetative stage,reproductive stage and maturation stage)through the six-rotor UAV equipped with 5-band multispectral camera,and the SMC values at corresponding time was acquired by drying method on the field at five soil depth(10 cm,20 cm,30 cm,45 cm and 60 cm).Then the spectral reflectance of field maize canopy was extracted to calculate a number of vegetation indexes(VIs).Firstly,data was adopted to analyze the grey relationships between SMC and the selected typical VIs,and the selected typical VIs were used to determine the sensitivity of different VIs to SMC at different growth stages.Secondly,machine learning models of Cubist,back propagation neural network(BPNN)and support vector machine regression(SVR)were constructed and verified.The result showed that the three machine learning models showed good performance on modeling and prediction at different growth stages.The effectiveness of the SVR model was optimal among the three machine methods.The effect of the BPNN model followed,and the Cubist model was relatively the worst.The optimal model was the SVR model at M stage,the modeling R^2 and validation R^2 for the SVR model were 0.851 and 0.875,and the root mean square error(RMSE)both were 0.7%,and the normalized root mean square error(nRMSE)were 8.17%and 8.32%,respectively.The inversion accuracy of the SVR model at R stage performed badly,the modeling R^2 and validation R 2 for the SVR model were 0.619 and 0.517,respectively.The research result was of great significance to monitor the soil moisture content in root area of crops and meaningful to precision irrigation.