Research On Bayesian Regression Network Ensemble Learning Method And Its Application To Agricultural

Precision fertilization is an important part of precision agriculture and one of the cores of promoting agricultural modernization.For the complex nonlinear relationship between soil measured content,yield and fertilization amount,it is often difficult to obtain accurate target prediction values with traditional precision fertilization methods.In addition,the period required to obtain actual agricultural data is long,and the actual samples obtained are characterized by a small number and imbalance.In order to solve the above problems,this paper proposes homogeneous ensemble and heterogeneous ensemble learning methods based on Bayesian regression network.Bayesian analysis is a classical and practical statistical method,which is widely used in many fields.This method can combine prior information and data information of all samples for modeling,and can not only give the predicted mean value,but also the confidence level and uncertainty of the predicted value.Compared with other machine learning algorithms,the Bayesian regression network can adaptively obtain the required hyperparameters according to the data information.Combining the characteristics of Bayesian analysis with ensemble learning can lead to a model with more accurate prediction ability,stronger generalization ability,and more robust model.The main work is as follows.1)Using Gaussian process regression as the carrier of Bayesian nonparametric regression,the hyperparameters of the model are adaptively obtained by the maximum likelihood method.2)Using Bagging method to repeatedly sample the training samples to generate sub-sample spaces.The training model is used to generate several individual networks.Due to the difference of data distribution and the instability of Gaussian process in different subsample spaces,multiple individual networks with excellent predictive ability and degree of difference are generated.Using the idea of selective integration,a batch of individual networks with high prediction accuracy and obvious differences are selected from all individual learners through the prediction error ranking of the validation set and the Affinity Propagation Clustering algorithm to replace all the networks for integration,and the number of clusters is calculated.The impact of forecasting is explored.3)In order to further improve the generalization ability of the model,two combination strategies are proposed in the conclusion generation part.One is based on the accuracy of predict weighting method,and the weight is assigned according to the prediction accuracy given by the model in the validation set.The second is to use the characteristics of Bayesian analysis to assign weights directly according to the inverse of the prediction variance given by the model on the test set named the uncertainty of predict weighting method.And compared the accuracy of the best sub-network,the equal weight combination strategy and the above two combination strategies.4)Further combine the Bayesian regression network ensemble learning model based on the uncertainty of predict weighting method and random forest,gradient boosting regression tree,artificial neural network ensemble learning model for multi-model heterogeneous ensemble learning.The prediction errors of this model and the four base models are compared.In this paper,the proposed algorithm is applied to the actual agricultural data obtained from the corn test field in Yushu City,Jilin Province.The experimental results show that the Gaussian process regression network under the Bayesian nonparametric model has strong fitting ability and generalization ability to small amount sample data.In terms of combining strategies,the uncertainty of predict weighting method is better than that of the single optimal sub-network,the equal weight combination and the accuracy of predict weighting method.Compared with a single network,homogeneous integration and multi-model heterogeneous integration have improved accuracy and generalization ability.The multimodel heterogeneous integration further improves the accuracy and generalization ability of prediction,and the maximum prediction error does not exceed 1%.Fully meet the actual production needs.