Abstract:
The first flowering stage of winter wheat depends strongly on meteorological factors, particularly on climate-sensitive elements. Therefore, it is remarkably significant to develop prediction models using machine learning for precise control of wheat scab. In this study, Random Forest (RF), Back Propagation neural network (BP), and Multiple Linear Regression (MLR) were integrated to establish precise prediction models for the first flowering period. The winter wheat phenology and daily meteorology were collected in 10 observation points of Jiangsu Province, China, during the period from 1980 to 2019. Five kinds of parameters were set in the RF and BP. The optimal model of each parameter was achieved after hundreds of times of automatic learning. The determination coefficient, the root mean square error, and the prediction accuracy were applied as the evaluation indicators to evaluate the forecast ability of the models. The results showed that there were obviously interannual fluctuations in the first flowering period of winter wheat, where most regions tended to be ahead of time. There were also great differences in the first flowering period among different regions. The specific difference was more than 21 d between the latest and the earliest days for the first flowering stage of winter wheat. The influence of temperature factors on the first flowering stage was more important than that of precipitation and sunshine factors. The five most important factors were ranked as follows, the active accumulated temperature of daily average temperature ≥0 ℃ from December of last year to March of that year, the average temperature from December of last year to March of that year, the accumulated days of daily minimum temperature ≤0 ℃ from December of last year to March of that year, the average temperature in March of that year, and the active accumulated temperature of daily average temperature ≥0 ℃ from December of last year to February of that year. Furthermore, the important characteristic variables were selected to predict the first flowering stage in early April. Correspondingly, RF, BP, and MLR were utilized to predict 5 days ahead at the shortest, while 32 days in advance at the longest. The integrated prediction of the optimal RF model corresponding to the five Mtry parameters was better than that of the single optimal BP model. The highest prediction accuracy was the RF, followed by the BP, and the MLR was relatively low. The RF and BP normally considered the nonlinearity between the predictors, but the MLR could not. Additionally, fewer predictors of MLR could better characterize the impact on the first flowering stage. The RF simulation value for the extreme years at the first flowering stage was smaller than the actual observation value, whereas, the BP presented the excessive simulation of fluctuation range at the first flowering stage. Consequently, the RF can accurately simulate the fluctuation trend at the first flowering period, where the prediction accuracy was over 85.0% among most stations. This finding demonstrates that the RF has high reliability and potential capacity for the forecast application of winter wheat at the first flowering stage.