| The most challenging task in the agricultural field is how to accurately predict crop yields.Wild blueberry is one of the important fruit crops for trade in northern China,and its fruit-set is a key factor in determining the yield.In the past,typical machine learning algorithms used to collect a large amount of real data to build prediction models.However,agriculture in the north is underdeveloped and research on wild blueberries is very limited,which limits the possibility of directly building prediction models.Based on the above,the purpose of this thesis is to establish an optimal wild blueberry fruit-set prediction meta-model by using the combination of wild blueberry pollination simulation model and machine learning,verify the possibility of the meta-model for cross-regional prediction,and apply it.This thesis uses transfer learning methods to solve the problem of insufficient data for training machine learning models.The main contributions of this thesis are as follows:Firstly,aiming at the problem that the collection of real agricultural data often has a long period and high cost,and there is no guarantee that the data will be able to effectively train machine learning models,this thesis combines computer simulation and machine learning,and uses wild blueberry pollination simulation models to design simulation experiments to generate simulations data set,the pollination model is verified through field observation and experimental data collected in Maine,USA in the past 30 years.The results show that meta-modeling overcomes the limitations of traditional use of empirical models or computer simulation methods to predict crop fruit-set,and can achieve rapid prediction.Secondly,in order to reduce model input,this thesis selects the best feature subset before model training.Nine regression prediction modeling methods are compared by cross validation: multiple linear regression,support vector machines,K nearest neighbors,decision tree,random forest,gradient boosting decision tree,adaptive boosting,artificial neural network,extreme gradient boosting.The results show that the prediction performance of the extreme gradient boosting meta-model is better than other models.In addition,this study uses grid search to optimize the optimal meta-model,and analyzes the important factors that affect the output of the meta-model.At last,this thesis verifies the rationality of the optimal meta-model for the prediction of wild blueberry fruit-set in northern China,and successfully applies the transfer learning method to solve the problem of insufficient data for training machine learning models.Based on the temperature and rainfall days of wild blueberry in 151 districts and counties in northern China in 2019,the extreme gradient boosting meta-model was used to design prediction experiments to study the effects of different weather changes and different bee densities during the flowering period on the fruit-set of wild blueberry in the study area.The results show that the increase in rainy days and the increase in temperature pose a serious threat to the fruit-set in most areas in northern China;and under different weather trends,the highest fruit-set in northern China is located at high altitudes around 50°N.At the same time,under different weather conditions,the increase of osmia and bumblebee density has the most significant impact on the positive trend of blueberry fruit-set which can alleviate the negative impact of bad weather on blueberry fruit-set. |
