Predicting Above-ground Biomass And Yield Of Winter Wheat Based On UAV Multi-source Spectral Remote Sensing Imagery

Above-ground biomass is the most commonly used physiological and biochemical parameter to respond to crop growth in the field,and the level of yield is a good or bad measure of yield status.Timely and accurate yield estimation and efficient monitoring of above-ground biomass play a crucial role in precise management and rapid decision-making in agriculture,and have important research implications for the formulation and implementation of national food security and agricultural policies.This study was conducted at the Qiliying experimental base of the Institute of Farmland Irrigation,Chinese Academy of Agricultural Sciences,Xinxiang,Henan Province,China,using two multi-rotor UAV platforms equipped with multispectral,thermal infrared and hyperspectral sensors,respectively.Canopy multispectral,thermal infrared and hyperspectral images of winter wheat planted in the main experimental area at tassel,flowering and grain filling stages were collected in 2019-2021.The main studies explored estimating winter wheat yield based on UAV multi-spectral,hyperspectral and thermal infrared,and estimating winter wheat above-ground biomass based on UAV multi-spectral.The main research and conclusions of the paper are as follows:(1)The construction of the vegetation index was based on multi-spectral UAV imagery acquired.Three basic learning models,(1)Cubist,(2)support vector machine(SVM)and(3)linear ridge regression(LRR),were constructed and a Stacking ensemble learning model was development combining the three basic learning models for above-ground biomass prediction at tassel and flowering stages.The correlation between vegetation index and above-ground biomass was mostly greater at flowering than at tassel stage,the LRR estimation model performed best at tassel(R2 = 0.58,RMSE = 1843.42 kg·hm-2,RPIQ = 2.25,RPD = 1.50)and the SVM model had the highest estimation accuracy within flowering(R2 = 0.65,RMSE = 2221.42 kg·hm-2,RPIQ = 2.46,RPD = 1.63).Compared to the base learning model for each period,the accuracy of the estimation model constructed based on the Stacking algorithm improved to a maximum R2 of 0.59 and 0.68 at the tassel and flowering stages,respectively.(2)Vegetation indices based on multispectral image data were used as input variables.Multiple linear regression(MLR),partial least squares regression(PLS),support vector machine(SVM)and multiple mixed linear regression(Cubist)as regression methods for yield prediction at tassel,flowering and filling stages.MLR and Cubist were also used as secondary learners,and the Stacking method was used to ensemble each primary learner,and a five-fold cross-validation method was used for model construction and accuracy evaluation.Under irrigation,the correlation between each vegetation index and yield gradually increased as the growth stages of winter wheat developed,reaching a maximum value of 0.67 at the filling stage.comparing the accuracy of the four base learner models,the Cubist model had the highest yield estimation accuracy at the tassel stage(R2 = 0.41),flowering stage(R2 = 0.45)and filling stage(R2 = 0.57).Compared to the R2 maxima of the base learner model at each fertility stage,the R2 of the secondary learner MLR model increased to 0.53,0.55 and 0.61,respectively,while the R2 of the secondary learner Cubist model increased to 0.54,0.58 and 0.61,respectively.The best results were obtained with the Cubist model.The fusion of models using the Stacking method also improved the yield estimation accuracy.(3)The construction of multiple vegetation indices significantly correlated with yield as well as the crop water stress index(CWSI)is based on multispectral and thermal infrared imagery from drones acquired at the tassel,flowering and grain filling stages.A recursive feature elimination(RFE)method was used to filter the best subset of vegetation indices in each period.Four yield estimation models were constructed using support vector machines(SVM)based on the vegetation indices and CWSI before and after filtering.The results showed that the vegetation indices and CWSI were strongly correlated with yield in all three fertility stages.Among the four yield estimation models,the model based on the best subset of vegetation indices combined with CWSI performed best,with R2 reaching 0.45,0.51 and 0.56 in the three fertility periods respectively.(4)Constructing multiple spectral indices based on hyperspectral imagery for yield prediction at flowering and grain filling stages.Elimination(RFE),Boruta and Pearson Correlation Coefficient(PCC),were used to filter high spectral indices to reduce the dimensionality of the data,four major basic learner models,namely Support Vector Machine(SVM),Gaussian Process(GP),Linear Ridge Regression(LRR)and Random Forest(RF),were also constructed,an ensemble machine learning model was developed by combining the four base learner.The results showed that the SVM yield prediction model constructed on the basis of the preferred features performed the best among the base learner models,with an R2 between 0.62 and 0.73,the accuracy of the proposed ensemble learner model was higher than that of each base learner model,the R2(0.78)of the yield prediction model based on Boruta’s preferred characteristics was greatest during the grain filling stage.(5)Combining the above multi-period yield prediction models constructed from multi-species spectral data.The DLF ensemble yield prediction model based on hyperspectral data performed the best in the flowering and filling stages,with R2 reaching 0.66 and 0.78 respectively,providing a theoretical reference for future yield prediction studies.