Abstract: The atmospheric water vapor can be retrieved exactly on certain scales by the deflection and delay of the Navigation Satcllite Timing And Ranging Global Position System(GPS) satellite signal along signal path. Based on the water vapor obscr-vations along the signal paths over the density ground-based GPS network,a new GPS tomography model has been established by using a Gauss function as horizontal constraint and the radiosondc(RS) observation vertical constraint, which includes RS mean values of 4 years as priori information and RS standard deviation as the weight matrix to get the regional high-accuracy three-dimensional (3D) distribution of the atmospheric water vapor. The observations on the ground surface have been used in the tomography formulation to improve the tomographic result accuracy in the overall region, especially the low-level. A 3D water vapor tomographic grid has been designed based on the GPS observation network with 22 GPS sites in the eastern part of Hubei Province in the middle reaches of the Yangtze River. The quasi-real-time retrieval system can provide the hourly output of 3D water vapor density. The 3D wet refraction and water vapor density can be retrieved from the observations of Slant Wet Delay(SWD) and Slant Water Vapor(SWV), respectively. The accuracy of the retrieval results in the 3D tomography domain of the overall, the low-level, the center and the edge has been analyzed by comparing to the RS observations during the flood season of 2010. The results show that for overall sample tests the 3D water vapor density average deviation is -0.63 g/m3, the standard deviation is 1.22 g/m3, and the correlation coefficient with RS is 0.98. The correlation between the retrieval results for water vapor density and soundings is better than that between those for the wet refraction and RS ohservations, and the correlation hctween RS ohscrvations and the results of water vapor density after joining the ground surface observations is better than those without joining the ground surface observations in the tomography. For the rich observations path of GPS, the accuracy at the center of the tomography domain is better than that in the overall and at the marginal domain. Combined with the ground surface observations, the 3D water vapor retrieval accuracy is improved with enhancing the correlation between tomographic results and RS observations, reducing the standards deviation of tomographic results in low-level of the region, and the average deviation of tomographic results in the center and the layer above 2 km of the region. The magnitude order difference between the elements in the signal length matrix due to the ground surface observation being used in tomography could disturb the stability of the observation equations and thus increase the overall standard deviation.