Estimation Of Winter Wheat Yield Based On Multi-spectral UAV Data

The food issue has always been the focus of worldwide attention,which is related to the stability,prosperity and development of the society7 In recent years,UAV systems(UASs)have developed rapidly and can provide key technologies and data support for ensuring food safety production and evaluation7 To estimate crop yield accuratelly,this study studied the yield estimation by taking winter wheat for example7 Based on multispectral UAV data,vegetation indices and texture features were extracted and fused at multiple growth stages7 Then,Stacking method was used to integrate multiple machine learning models7 Finally,a regional prediction and technical method for winter wheat yield was constructed7 The specific research contents are as follows:(1)Exploring the performance and differences of different machine learning models in yield prediction of winter wheat at different growth stages7 In this work,four machine learning methods,including Random Forest Regression(RFR),e Xtreme Gradient Boosting(XGBoost),Least absolute shrinkage and selection operator(Lasso),and Decision Tree(DT),were used to build vegetation index yield prediction models7The accuracy of the winter wheat yield prediction models decreases from the jointing stage to the filling stage,and then improves gradually,which shows the best prediction performance during the filling stage7 Among them,the Lasso model has the highest accuracy(~20078,RMSE054722 g/8)~2,MAE044781 g/8)~2)7(2)Modification of yield prediction models based on single growth stage multi feature parameters fusion7 This study constructed yield prediction models that combined vegetation indices and texture features by incorporating selected texture features7 The results show that the yield prediction models at the jointing stage,heading stage,flowering stage,and filling stage have all improved in effectiveness7 Compared to the vegetation index yield prediction models,the accuracy of the XGBoot model during the filling stage has been significantly improved,with an increase of 0702 in~2,a decrease of 270 g/8)~2 in RMSE,and a decrease of 0748 g/8)~2 in MAE7 It proves the effectiveness of incorporating texture features in improving model prediction accuracy7(3)Construction of yield prediction model based on Stacking method for primary learning method fusion7 In view of the poor anti-interference ability of a single learner and the problem of overfitting or under fitting issues,this study used Stacking method to fuse primary learning method7 In addition,Stacking models was developed by combining primary learning methods with Support Vector Regression(SVR),Lasso,and Ridge models as secondary learning method,respectively7 It can be seen that the accuracy of the Stacking models for predicting winter wheat yield have been improved in various growth stages,and good results have been achieved during the filling period7Compared with single learners,selecting RFR,XGBoost,and Lasso models as primary learners,while Ridge model as a secondary learner has the highest prediction accuracy,with~200789,RMSE053748 g/8)~2,MAE044748 g/8)~27(4)Studying of yield prediction method based on the fusion of feature parameters at multiple growth periods7 Regarding the issue of low sensitivity of crop yield to single growth period characteristics,this work added the vegetation indices and texture features of the subsequent growth stage as new yield estimation features to the yield estimation features of the previous growth stage,forming joint yield estimation features for multiple growth stages7 By utilizing the yield estimation factors of multiple growth stages,the accuracy of the yield prediction models are improved7 Through the combination of growth stages and the screening of yield estimation features,it can be seen that the RFR model,XGBoost model,and Lasso model have improved accuracy during the combined flowering and filling stages,and the Lasso model has the highest accuracy7 The predictiong accuracy is of~200789,RMSE053733 g/8)~2,MAE044730g/8)~2,which proves that combining multiple growth stages of yield estimation features can improve the prediction accuracy of the models.