Abstract: Based on data observed by Doppler weather radars deployed at Longyan, Zhangzhou, Quanzhou and dual polarization Doppler radar deployed at Haicang of Xiamen, the dual-radar three-dimensional wind retrieval and mosaic technique in the dynamic earth coordinate and the WRF (Weather Research and Forecasting) model and its data assimilation system are used to conduct a set of numerical experiments to investigate the impacts of three-dimensional variational assimilation of dual-radar retrieved winds upon the detailed heavy precipitation forecast caused by the landfall typhoon "Meranti" (1614) during the period of 14-15 September 2016 in Fujian Province. The results are as follows:(1) Dual-radar retrieved winds in the dynamic earth coordinate can reasonably describe realistic wind fields, and the errors are relatively small. Comparisons between retrieved winds and wind profiler radar detections during the "Meranti" landfall show that the mean absolute errors of retrieved wind direction and speed are 7.8° and 2.6 m/s, respectively. Comparisons between retrieved winds and Xiamen second radio-soundings show errors of 3.4° and 1.1 m/s, respectively. (2) Horizontal thinning of retrieved winds have obvious impacts upon the assimilation and forecast results. Overly dense retrievals of winds have negative effects on the assimilation and forecast results. A triply nested domain with resolutions of 18, 6, 2 km is used in the present study, and assimilation has been applied to all the domains and only to the innermost domain, respectively. For the above two situations, the typhoon circulation structures from analysis and from forecast both are negatively impacted when the horizontal resolution of retrieved winds reaches 0.1° resolution, and the negative effects are more obvious when the horizontal resolution of the retrieved winds becomes higher. Sensitivity experiment results indicate that numerical forecasts are the best when using 0.2° resolution data. (3) Using NCEP/GFS (National Centers for Environmental Prediction/Global Forecast System) 0.5°×0.5° analysis data as model initial fields, three experiments are carried out at different initialization time, i.e., 14:00, 20:00 BT 14 and 02:00 BT 15 September 2016. It is found that the results with different initialization time are obviously different considering hourly evolution of simulated typhoon inner core and spiral rain band in Fujian Province, typhoon track and intensity, the threat scores and spatial correlation coefficients of hourly precipitation. The simulation initialized at 14:00 BT 14 yields the best result. The simulation initialized at 20:00 BT 14 yields the worst, which is due to the overestimation of initial typhoon maximum wind core. (4) Based on the above three experiments initialized at different time, radar retrieval winds are assimilated by the WRF three-dimensional variational assimilation system. Simulations of hourly surface wind field, typhoon inner core and spiral rain band in Fujian Province, and the threat scores and spatial correlation coefficients of hourly precipitation, as well as typhoon circulation and vertical profiles of zonal and meridional winds all are improved significantly, and the valid time of positive effect is up to 24 h. However, for the simulation of typhoon track, the valid time of positive effect is only six hours.