Wheat Growth Monitoring And Yield Prediction In The Hetao Irrigation District Based On UAV Remote Sensing And Machine Learning

The Hetao Irrigation District is one of the important spring wheat production bases in China.Rapid,efficient,and non-destructive acquisition of field growth information of spring wheat and yield prediction is of great significance for food security and economic development.This study aims to use unmanned aerial vehicle remote sensing technology and machine learning methods to carry out real-time,efficient,and accurate estimation and prediction of spring wheat growth and yield in the Hetao Irrigation District.Five field experiments were set up at three ecological sites in the Hetao Irrigation District with nitrogen application rates,base fertilizer ratios,irrigation volumes,plant densities,and varieties.A consumer-grade unmanned aerial vehicle equipped with multi-spectral and RGB sensors was used to collect multi-temporal remote sensing images of spring wheat at the jointing stage,heading stage,flowering stage,and grain-filling stage,and extract26 multispectral and 25 RGB vegetation indices of the spring wheat canopy as input variables for machine learning models.Through field sampling measurement and laboratory testing,the growth and development data and yield data of spring wheat were obtained as output variables and validation data for machine learning models.The spring wheat’s growth and yield were analyzed and predicted using regression analysis and 25machine learning models based on multispectral,RGB vegetation indices,and their combination.After selecting the highest accuracy model,the model accuracy was further improved using hyperparameter optimization technology.The main results of this study are as follows:(1)The important features for estimating SPAD based on MS multispectral vegetation indices are SCCCI and GNDVI,with the optimal model being Huber.The important features for estimating SPAD based on RGB vegetation indices are IPCA,r,and RBRI,with the optimal model being BR.The important features for jointly estimating SPAD based on the two types of vegetation indices are R,SCCCI,IKAW,IPCA,and GNDVI,with the optimal model being OMP.The accuracy ranking of the three estimation modes after hyperparameter optimization is MS+RGB>MS>RGB,with the highest accuracy in estimating SPAD using the OMP model under the combination of MS+RGB vegetation indices(test set MAE=4.056,MSE=24.704,RMSE=4.970,R2=0.820,RMSLE=0.175,MAPE=0.125).(2)The important features for estimating Fv/Fm using MS multispectral vegetation indices are MSAVI2,with the best model being LAR.The important features for estimating Fv/Fm using RGB visible light vegetation indices are RGBVI and Ex R,with the best model being GBR.The important features for estimating Fv/Fm using the combination of vegetation indices(MS+RGB)are SIPI,Ex R,and VEG.The accuracy ranking of the three estimation modes is MS+RGB>MS>RGB,with the highest accuracy in estimating Fv/Fm using the ARD model(Automatic Relevance Determination)under the combination of MS+RGB vegetation indices(test set MAE=0.019,MSE=0.001,RMSE=0.024,R2=0.925,RMSLE=0.014,MAPE=0.026).(3)The best model for estimating plant water content(PWC)using multispectral vegetation indices is KNN,with the important features being CL2 and MSAVI2.The best model for estimating PWC using RGB vegetation indices is ET,with the important features being GBRI and r.The best model for estimating PWC using the combination of vegetation indices(MS+RGB)is LR,with the important features being CL2,INT,GBRI,and WI.The accuracy ranking of the three estimation modes is MS+RGB>MS>RGB,with the highest accuracy in estimating wheat PWC using the LR model(Linear Regression)under the combination of MS+RGB vegetation indices(test set MAE=0.026,MSE=0.001,RMSE=0.032,R2=0.863,RMSLE=0.020,MAPE=0.046).(4)The important features for estimating wheat leaf area index(LAI)using multi-spectral vegetation indices are OSAVI and MSAVI1,and the best model is Ridge.The important features for estimating LAI using RGB vegetation indices are RBRI,Ex GR,IPCA,and b,and the best model is SVM.The important features for estimating LAI using a combination of vegetation indices(MS+RGB)are IPCA,OSAVI and b,and the best model is TR.The accuracy ranking of the three estimation modes is MS+RGB>MS>RGB,and the accuracy of estimating wheat LAI using the TR model(Theil Sen Regression)in the MS+RGB vegetation index combination is the highest(test set MAE=0.376,MSE=0.189,RMSE=0.434,R2=0.839,RMSLE=0.122,MAPE=0.148).(5)The important features for estimating wheat aboveground biomass(AGB)using multi-spectral vegetation indices are SCCCI,GNDVI,and DVI,and the best model is OMP.The important features for estimating AGB using RGB vegetation indices are RGBVI,b,and INT,and the best estimation model is ADA.The important features for estimating AGB using a combination of vegetation indices(MS+RGB)are GLI2,INT,RGBVI,and GNDVI,and the best model is LLAR.The accuracy ranking of the three estimation modes is MS+RGB>MS>RGB,and the accuracy of estimating wheat AGB using the LLAR model(Lasso Least Angle Regression)in the MS+RGB vegetation index combination is the highest(test set MAE=1.096,MSE=1.620,RMSE=1.273,R~2=0.805,RMSLE=0.143,MAPE=0.119).(6)During the filling period,the important features for estimating wheat yield using multi-spectral vegetation indices are GNDVI,LCI,and CL2,and the best model is Ridge.The important features for estimating wheat yield using RGB vegetation indices are WI and IPCA,and the best model is ADA.The important features for estimating wheat yield using a combination of vegetation indices(MS+RGB)are WI,LCI,and CL2,and the best model is BR.The accuracy of predicting wheat yield using the BR model(Bayesian Ridge)optimized by the combination of MS+RGB vegetation indices during the filling period is the highest(test set MAE=0.357,MSE=0.172,RMSE=0.414,R~2=0.824,RMSLE=0.056,MAPE=0.056).(7)The combination of UAV multispectral-visible remote sensing technology and machine learning methods,along with hyperparameter optimization techniques,have improved the estimation accuracy of various wheat growth parameters and yield prediction to varying degrees.The combination of UAV multispectral-visible remote sensing technology and machine learning methods can effectively reflect and predict the growth and yield of spring wheat with high accuracy and reliability.This study provides a new technical means and decision support for precise cultivation management of spring wheat in Hetao Irrigation District,and has important theoretical and practical significance in improving the production efficiency of spring wheat.