To address the problem of large bias in the quantitative estimation of heavy precipitation by traditional methods using weather radar, the thesis uses the 1-hour cumulative rainfall as the estimation object, a new model for radar precipitation estimation based on radar mosaic data and XGBoost (eXtreme Gradient Boosting) algorithm. The model is designed with the radar combined reflectance factor of the previous hour as the input factor, and further employs several rejection strategies of anomalous samples to effectively remove some of the noise from the modelling samples, thus better constructing a non-linear mapping relationship between the radar combined reflectance and the estimated object. The root mean square error (RMSE) is 6.04 mm, the mean absolute error (MAE) is 3.50 mm, and the forecast bias (BIAS) is 1.05 for the 320,000 independently tested samples; compared to the Z-R(300,1.4) relational method currently used on operational systems, the RMSE and MAE of the former decrease by 20.6% and 10.3% respectively, while the BIAS indicators show a significant underestimation of precipitation magnitude by the latter. For samples with hourly rainfall intensity greater than 10 mm, further statistical results show that the new scheme’s RMSE, MAE and TS scores are substantially better than the Z-R (300,1.4) relational method for practical operational applications.