Abstract: Abstract: Water and heat fluxes exchange between biosphere and bottom atmosphere are indispensable parts in understanding the surface energy conversion and hydrological cycle processes happening on the land surface. Estimation and prediction of fluxes have immense research significance in fields of environmental protection, agricultural production and climate prediction. Land surface model is a powerful tool to obtain space-time continuous fluxes despite its poor simulation accuracy. The state-of-the-art data assimilation method provides a way to solve this problem. With the help of the offline version of Community Land Model CLM3.0 as a dynamic framework, we use the Ensemble Kalman Filter assimilation algorithm to build a single-site surface water and heat fluxes assimilation system. The algorithm perform an ensemble simulation to estimate initial condition error covariance and observational error covariance for the objective dynamic model and analyze background diagnostic outputs by calculating a weighted mean with observations. Perturbations on surface initial condition, atmospheric forcing data and observations are generated by a random sampling strategy based on the supposition of normal distribution with a priori mean and standard deviation for all variables. Data from three flux observing sites from Ameriflux flux observational network (Chestnut Ridge, ARM SGP Main and Tonzi Ranch) which stand for three different land surface conditions are engaged in parallel experiments to test the system and evaluate the effectiveness of flux assimilation under the framework of land model. Before processing further experiment, an optimal ensemble size was selected by evaluating outputs of latent heat from models with different ensemble size RMSE. The results of parallel experiments showed that direct assimilation of sensible and latent heat fluxes can improve the estimates of total surface sensible and latent heat fluxes in all three types of underlying land surface condition. In ARM SGP Main site, a typical case for cropland ground type, RMSE of sensible heat flux decreased from 67.49W/m2 to 14.07 W/m2 and that of latent heat flux decreased from 70.07 W/m2 to 14.35 W/m2. In Chestnut Ridge site that stands for forestry, RMSE of sensible heat flux dropped from 82.56 W/m2 to 48.56 W/m2 and that of latent heat flux fell from 42.99 W/m2 to 38.92 W/m2, respectively. Tonzi Ranch, a grassland site, RMSE in is also diminished by assimilating in situ observations with decrements of 45.14W/m2 for sensible heat flux and 8.75 W/m2 for latent heat flux. Furthermore, by comparing the results we gained above with mainstream study that focusing on assimilation of surface temperature and humidity to indirectly improve the fluxes prediction, we conclude that under the dynamic framework of community land model the flux outputs from direct assimilation model are better than those from surface state assimilation model. It is noteworthy that the accuracy of observational error estimation will directly affect the assimilation results though errors from initial condition, observation and atmospheric forcing will make contributions simultaneously.