In order to explore the feasibility of using equidistant split and random forest algorithm to predict the weather of the initial flowering period of winter wheat in northern Anhui,the in situ observation phenological data and the parallel observation meteorological data of the initial flowering period of winter wheat from seven agrometeorological observation stations in northern Anhui from 1980 to 2019 were used to screen the characteristic variables that affect the early or late flowering period by using the correlation coefficient method,and the training set and test set were separated by using the orderly equidistant sampling method. Based on the random forest algorithm(RF),from 10th to 15th of April,one forecasting model was trained every day to realize the daily rolling weather forecast of the initial flowering period of wheat,and was compared with the forecasting model trained by four algorithms: analog neural network(ANN),linear support vector machine(LSVM),multiple regression(RG) and support vector machine(SVM). The results showed that the 40 key meteorological factors,which were composed of three kinds of meteorological elements: average temperature,maximum temperature and sunshine hours,were closely related to the early or late flowering of wheat. Among the six daily weather forecasting models for the initial flowering period,40 characteristic variables were selected for the five models from 10th to 14th of April,and 39 variables for the model from 15th of April. The average accuracy,mean root mean square error(RMSE),and mean determination coefficient(r2) of the six weather forecast models for training set and test set were 93.3% and 80.4%, 1.860-1.960 and 2.510-2.709,and 0.944 and 0.841,respectively. The three test indices of the forecast model based on RF algorithm training are better than those of ANN,LSVM,RG,and SVM algorithm training. The initial flowering period of 2020 and 2021 can be accurately forecasted 7 to 9 days in advance,using RF algorithm model. The weather forecast model of wheat flowering in northern Anhui province based on RF algorithm was constructed by splitting the training set by equidistance sampling,which can achieve higher forecast accuracy. |