| This study focused on cotton as the research object.Firstly,multispectral images of the experimental field were acquired by DJI drones during the experiment period.Then,the images were corrected and preprocessed.After vegetation index calculation and feature selection using SHAP values,two ensemble models,Bayesian model averaging and Stacking,were used to construct the best SPAD estimation model and the optimal MDG classification model of cotton,respectively.The construction composition and characteristics of different models were explored.A precise spraying prescription map based on the mite inversion map generated by the optimal model was used to evaluate the precise spraying effect of the unmanned aerial vehicle plant protection.The main research results are as follows:(1)Regarding the feature variable selection of the cotton SPAD estimation model and the MDG classification model,the SHAP value theory was used to obtain the best input variables.The SHAP value theory can evaluate the contribution of variables to the model and screen the best variables suitable for the model.The results showed that the best vegetation indices for the regression model were MTCI and LCI,while the best vegetation indices for the classification model were VARI,B,SIPI,and MTCI.(2)In order to explore the best cotton SPAD estimation model,three linear models and three nonlinear models were used to model and analyze the cotton SPAD.The results showed that the accuracy of the three linear models was similar.The optimal nonlinear model was RF,with an R2 of 0.688.Bayesian model averaging and Stacking were then used to construct the cotton SPAD estimation model.The results showed that the BMA23 ensemble model constructed based on Bayesian model averaging was the best cotton SPAD estimation model,with EN,BR,and RF as the base models.The R2 was 0.700.The rules and characteristics of the ensemble model constructed by the basic regression model were discussed,and the proportion of the basic models with high accuracy in the Bayesian model averaging ensemble model was analyzed.It was concluded that the combination of basic models should satisfy the principle of heterogeneity,that is,the combination of linear and non-linear models.The performance of the ensemble model composed of powerful base models will not improve as their weights increase.(3)In order to explore the best cotton MDG classification model,seven classification models were used,and the correlation between the seven models was analyzed to select the optimal four models as the basic models of the ensemble model.Then,two ensemble models were used to construct the classification model.The results showed that the best classification model was the S23 ensemble model constructed based on the Stacking ensemble model,with RF and XGB as the base models and SVM,KNN,or LR as the meta-model.The overall classification accuracy was 0.874.By exploring the composition of the two-layer basic models of the Stacking ensemble model,it was concluded that the first-layer model of the Stacking ensemble model should usually select powerful basic models,and the second-layer model should usually select general linear models.The performance of the ensemble model will not improve as the number of first-layer models increases.If the first and second-layer models are of the same type of algorithm,they will weaken each other’s model performance.(4)The S23 ensemble model was determined to be the best model for monitoring the level of cotton mite damage,and the optimal spraying parameters were obtained by combining the resampled precise spraying prescription map generated by the optimal MDG classification model with the pre-experiment of drone mist distribution.On the 5th and 10 th days after spraying,unmanned aerial vehicle images and ground data were collected to evaluate the precise spraying effect.The experimental results showed that the amount of pesticide used decreased by 6.4% compared with the previous... |
PDF Full Text Request