Abstract: In order to construct a county-scale dynamic refined simulation model for longan yield, authors analyze the influence of meteorological factors on longan yield using the longan production data of Maoming from 1990 to 2020 and the daily meteorological data of the national meteorological observatory, establish and carry out the comparative analysis on the dynamic simulation model of longan yield in Huazhou, Gaozhou and Xinyi based on the random forest regression method and stepwise regression method with different data schemes. The results showed that the longan yield in Maoming is closely related to meteorological factors, and the minimum temperature and the relative humidity during the growth period have the greatest influence on the longan yield in Maoming, with 15 and 14 meteorological factors selected respectively, and their maximum correlation coefficients being −0.31 and 0.43, respectively. Compared with the multiple stepwise regression method, the accuracy of longan yield simulation model constructed by the random forest regression method is higher. The model determination coefficient (R2) is 0.97, which increases by 7%, the mean absolute error (MAE) is 210.16kg·ha−1, which decreases by 52%, and the root mean square error (RMSE) is 289.62kg·ha−1, which decreases by 46%. When the data of similar climate characteristic areas outside the simulation target region is introduced, the simulation result of the random forest regression model is significantly improved, with R2 increases by 3%, MAE decreases by 32%, and RMSE decreases by 31%, while the simulation result of the stepwise regression model has no significant change. The longan yield simulation model based on the random forest regression method is reliable, which can meet the demand for refined meteorological service of longan.

Key words: Yield simulation, Longan, Refined simulation, Random forest regression