Abstract:

Based on the CMPAS (CMA Multi-source merged Precipitation Analysis System) algorithm developed by National Meteorological Information Center(NMIC), using the ground observational precipitation data and radar quantitative precipitation estimation data, Sichuan provincial precipitation products of 0.01°/1 h (SC) during April to September, 2021, were produced.Taking the hourly precipitation observation data as true value, SC and real-time precipitation products of 0.01°/1 h (RT) developed by NMIC were evaluated.The evaluation was divided into two categories according to whether the observation data is modified or not.On this basis, using the machine learning support vector regression algorithm (SVR), the improvement method of RT was developed.The results show that, in the case of observation data is not modified, RT is prone to overestimation and SC is prone to underestimation.These two kinds of products have similar performance in the basin area.RT has a better performance in Panzhihua, Liangshan, Ganzi and Aba region.In the case in which the observation data are modified, no matter the whole province or each region within the province, SC is clearly better than RT.RT is significantly larger than the observation value.This implies that the wrong observational data before modification have a significant impact on the performance of RT.The mean absolute error of the improved products is reduced by 29.9% and the root mean square error is reduced by 41.1% and the correlation coefficient is increased by 4.94% compared with that before the improvement.

Key words: Merged precipitation products, Support vector regression (SVR), Fine assessment