Application of Gradient Boost Decision Tree in Radar Quantitative Precipitation Estimation

Based on gradient boost decision tree (GBDT), a radar quantitative precipitation estimation (QPE) model is established through using the three-dimensional CINRAD (China New Generation of Weather Radar) radar echo data and the precipitation data from the regional automatic weather stations (AWSs) in Guangdong Province. In order to evaluate the GBDT scheme, both the Z-I relations and dynamic Z-I relations are used to estimate precipitation. The comparison experiment results show that the established radar quantitative precipitation estimation of GBDT scheme is better than Z-I relations and dynamic Z-I relations both in accuracy and stability, and can reflect the real situation of rainfall better. According to analysis indices such as mean error, relative error and root mean square error, the GBDT scheme can estimate precipitation fairly accurately. It improves the estimation of short-term precipitation for torrential rain, especially the precipitation greater than 30mm/h. Rainfall estimation of GBDT scheme is in good consistency with the observation by rain-gauge and can truly reflect the precipitation on ground surface.