Research On Winter Wheat Yield Estimation Model Based On Machine Learning

It is of great significance to conduct non-destructive yield estimation before harvest in wheat breeding high-generation plot yield comparison experiments which can improve wheat breeding efficiency.Therefore,with the rapid development of remote sensing technology,yield estimation in breeding high-generation plots based on this has gradually attracted widespread attention from teachers and students in the field of wheat genetics and breeding.Our team has been conducting long-term wheat genetic breeding work,requiring yield comparison experiments on hundreds of wheat varieties every year.However,due to limitations in agricultural machinery,equipment,and personnel,it is impossible to appropriately expand the selection quantity,which affects breeding efficiency somehow(crop variety selection is a probability event,so on a certain probability basis,the larger population selection range,the higher probability of excellent new strains appearing).Therefore,this study selected Agricultural Station 1 in Yangling Demonstration Area,Shaanxi Province as the research area,and used the UAV platform equipped with a five-channel multispectral sensor to collect data from 290 varieties in the experimental site from 2021 to 2022 for four growth stages(jointing,heading,filling,and maturity)of wheat in the experimental site.The widely used traditional linear model(LR)and five different algorithms in machine learning models,including Partial Least Squares Regression(PLSR)based on linear regression models and Support Vector Machine(SVM)adapted for small sample training,were selected,There are also three types of models based on decision tree algorithm[Decision Tree(DT),random forest(RF)and Gradient Boosting Regression Tree(GBRT)],which are intended to fit a scientific and accurate winter wheat yield estimation model based on UAV multispectral remote sensing.The test results are as follows:(1)The DJI M300 RTK quadrotor UAV was selected to carry the Red-Edge MX multispectral sensor to obtain the multispectral data of four key growth periods in each plot of winter wheat in the test area,and the wheat yield estimation models with different algorithms were established based on the characteristics of multiple multispectral images.Firstly,extract five band information(Red,Green,Blue,Near Infrared,and Red-Edge)from UAV images,40 texture features,and calculate 40 vegetation indices related to wheat growth and yield.Analyze the correlation between each feature and yield and model weights.Combine the R,B,G,N bands,MCARI,RENDVI,VREI1,and ARI2 vegetation indices with N＿Dis、N＿Var、N＿Con texture feature combination is the optimal feature combination,and winter wheat yield estimation models for each period are established based on LR,PLSR,SVM,DT,RF,and GBRT algorithms.Among them,the RF model has the highest accuracy(R~2=0.75)and the RMSE is the lowest;The results showed that the RF wheat yield estimation model had the highest accuracy at a altitude of 30m with the filling period data.(2)For multiple growth stages and different model input parameters,research has found that the RF algorithm model has the highest determination coefficient R~2=0.808 and the lowest RMSE when combined with data information on heading,filling,and maturity stages.Secondly,based on the optimal combination of growth stages and three types of feature parameters,the GBRT algorithm constructed the model with the highest coefficient of determination R~2=0.927 and RMSE=0.091.Finally,based on the impact of different ROI scales on model accuracy,it was found that the model performs best at 80%sampling scale(R~2=0.934,RMSE=0.085).