Improving accuracy of reproduction by numerical model is needed to elucidate the mechanism and prediction of local heavy rainfall. However, it is difficult to reproduce local heavy rainfall in numerical weather prediction model, because its 　spatial scale is too small. One of a source of the reason is the lack of accuracy of the initial value. Partly because of its high spatial and temporal variability, the moisture field is one of the less well described variables in initial conditions of numerical weather prediction system.

The aim of this study is clarifying sensitivities of data assimilation for the reproduction of more accurate local heavy rainfall, and considering the availability of the weather forecast.

This study investigated the impacts of GPS-derived precipitable water vapor (here after GPS-PWV) and Doppler radar-derived horizontal wind (here after HW) data assimilation on precipitation prediction with 1-km horizontal resolution. Three-Dimensional VARiational Data Assimilation (3D-VAR DA) system was used to assimilate GPS-PWV. And HW was assimilated by nudging method. Because of poor reproducibility in the experiment with 1 km horizontal resolution, this study conducted many sensitivity experiments about heavy rainfall event occurred in the southern Gifu Prefecture on 15 July 2010. This study used three kinds of GPS analysis data with different analysis methods such as Near Real Time (NRT) analysis data, reanalysis data and F3 data. This study also examined the sensitivity of these GPS data.

First, this study carried out experiments of data assimilation only the initial value. Some comparisons between observed and model-reproduced precipitation regions help to define the impacts of assimilating GPS-PWV and HW into the model. Assimilating GPS-PWV or HW reproduced precipitation area in the correct position and removed incorrect precipitation. Assimilating both GPS-PWV and HW, the position of the precipitation region was also correctly, and it reproduced more precipitation. However, accumulated precipitation was less than half of the observation. Therefore, sequential assimilation experiment with nudging was introduced. If applied sequential nudging only when water vapor of analytical value is moister than that of model, model-reproduced precipitation was the most closest to the observation.

Then this study carried out comparing time series of one hour precipitation between model and AMeDAS point of Tajimi. Sequential assimilation experiment with reanalysis data simulated one hour precipitation similar to that of observation. Those results indicate that reproduction precipitation of the case which locally and strongly precipitation maintains for a long time needs to improve water vapor fields sequentially.

Assimilated reanalysis GPS data, the result of reproducing precipitation was better than assimilated NRT data or F3 data. Contrast to other GPS data, because reanalysis data was analyzed coordinate value and zenith total delay (ZTD) separately, more accurate ZTD than the others was calculated. Therefore, it is important for improving precipitation prediction that removing error occurred when performing the GPS analysis.

To investigate whether sequential assimilation is useful or not for forecast, in this experiment, sequential assimilation was applied over only three hours from initial time. The result of this experiment, the time series of one hour precipitation in the southern Gifu is about the same as experiment which sequential assimilation was applied over twelve hours. This result means that sequential assimilation method is also useful for forecast experiment at least three hours after analysis nudging.

In the future, through this study, there is a possibility to improve the forecast by geting rid of the error when analyzing NRT data. And to improve the moisture field more accurately, it is necessary to realize more fine scale data assimilation and installation of marine type GPS.