Analysis The Uncertainty Of Soybean LAI Estimation Model Based On Hyperspectral

Soybean is one of the important oil crops,leaf area index(LAI)reflect photosynthetic rate and yield prediction in breeding high-yielding varieties.The test area is located in Jiaxiang County,Jining City,Shandong Province,The experiment data were obtained from the UHD185 carried by unmanned aerial vehicles,the data include126 multi-strains of soybean hyperspectral data and ground measured LAI data were obtained from the flowering period,bearing pod period,full pod periode,beginning seed period and mature period.Due to the model estimation accuracy is affected by the selected band,model calibration set and model form.Based on this,the model uncertainty research was carried out.The article divided uncertainties of remote sensing technology in estimating LAI into three aspects.Firstly,we use Least Absolute Shrinkage and Selection Operator(LASSO),maximum determination coefficient(MDC)and optimum index factor(OIF)to select the sensitive bands.Secondly,we use simple random sampling SRS and stratified type random(STR)doing model calibration set analysis.Finally,we construct six models and compare with each other to choose the suitable model.The specific contents and results are as follows:(1)Parameter uncertainty of model,on the base of previous studies LASSO,MDC and OIF are used to select the hyperspectral sensitivity bands of each growth period as the input parameters.The results showed that the susceptible bands were 540nm,760nm and 852nm for flowering period,464nm,628nm and 856nm for bearing pod period,556nm,644nm and 796nm for full pod period,572nm,756nm,848nm for beginning seed period,556nm,628nm and 908nm of mature period,548nm,760nm and 800nm for the whole growth period.(2)The uncertainty of input model,we use SRS and STR to extract model calibration sets,using Random forests(RF),artificial neural networks(ANN)and Support vector machines(SVM)for 100 times in modeling and prediction for the whole birth and single growth period,then using the average performance of model evaluation as index.The RF model SD_R~2 based on SRS and STR for the whole growth period is0.042 and 0.031,respectively,single growth period is 0.137 and 0.13.The SVM model based on the two sampling methods have the same trend.The result shows that model based on STR have a stable prediction accuracy.(3)The uncertainty of model structure,we use conventional univariate and multivariate models for different growth stages of the model establishment,and analysis modeling and prediction accuracy.Model utilize the least squares is suitable for balanced LAI data(eg,in beginning seed period the unary linearity model R~2=0.64,RMSE=0.685,PRD=1.628,and RRMSE=0.163,respectively.The exponential model R~2=0.644,RMSE=0.706,PRD=1.581,and RRMSE=0.168,respectively);RF model is suitable for sample size and variation is relatively large(eg.whole growth period R~2=0.533,RMSE=1.283,PRD=1.459,RRMSE=0.26);the multivariate estimation model has high accuracy and high precision than univariate model each growth model.