Abstract: To do a good job of refining low-temperature frost prediction of white loquat in Ninghai County, Zhejiang Province, a low-temperature frost prediction model based on machine learning was established．Numerical temperature forecasts and geographic information were used as input variables, actual temperature was used as output variables．Screening was carried out on nine different models, and it was determined that progressive gradient boosting regression tree was the most effective model for low-temperature forecasting．A low-temperature forecast model was built using by progressive gradient boosting regression tree with January, February, and December from 2021 to 2023 data as a training set, January and February 2024 data as a testing set．Results showed that model had a revised effect on numerical temperature forecasting, reduced average error of low temperature prediction at station by 0.29 °C, and refined spatial resolution of low-temperature prediction from 5 km×5 km to 200 m×200 m．Model effectively improves accuracy and fineness of white loquat frost prediction, and has good application value．