Research On Grain Yield Prediction Based On Machine Learning

Our country is the largest grain producer in the world.And grain self-reliance is the guarantee of people’s well-being.Grain security is a fundamental problem related to human survival,and has an important impact on national economic development and stability of society.The world is now in a period of major changes,the risks and uncertainties in the grain market have increased obviously.Grain yield prediction can provide a decision base for countries to formulate grain security policies.And it is also of great importance for improving the agricultural production environment and grain yield.In this thesis,grain yield prediction models are built based on machine learning algorithms and conducts research on grain yield prediction.The main research object of the research is grain yield in Hubei Province,the data of grain yield and influencing factors in Hubei province from 1992 to 2021 and from 1978 to 2021 are used to do comparative analysis of model training and testing.The results show that using the dataset of grain yield and its influencing factors in Hubei province from 1992 to 2021 for model training and testing has better results.When using the comparison data set of the grain production data and influencing factors in Hubei province from 1978 to 2021 for training and testing of the model,the BP neural network model and support vector regression model perform not very well in predicting.The random forest model performs well in training and testing on both datasets.In this thesis,single models such as Random Forest(RF),BP Neural Network(BPNN)and Support Vector Regression(SVR)have been proposed to use the Stacking ensemble algorithm and the average method ensemble algorithm for model fusion,and four ensemble models are constructed.Random forest and BP neural network model use the the Stacking integration algorithm to build the RF-BPNN-Stacking model.Random forest,BP neural network and support vector regression model use the Stacking integration algorithm to build the RF-BPNN-SVR-Stacking model.Random forest and BP neural network model use average method integration algorithm to build the RF-BPNN-AVG model.Random forest,BP neural network and support vector regression model use average method integration algorithm to build RF-BPNN-SVR-AVG model.And the prediction effect of the four integrated models and the single model was compared and analyzed.The results show that the experimental predictive effects of the four integrated models constructed in this thesis are both better than that of single models.The RF-BPNN-Stacking integrated model has the best prediction effect in the 5-fold cross-validation and is better than the two integrated models using the average method.The RF-BPNN-SVR-Stacking integrated model has the best prediction effect in the 10-fold cross-validation and is better than the two integrated models using the average method.The two integrated models built using the Stacking algorithm are better than the two integrated models built by the commonly used model combination strategy averaging method in the two cross-validation methods respectively,and the application of these two Stacking integrated models can improve the accuracy of grain yield prediction.It provides a new concept for grain yield prediction.