A hybrid four-dimensional variational（H4DVAR）data assimilation approach is proposed by combining the Ensemble Square Root Filter（EnSRF）with the Four-Dimensional Variational（4DVAR）data assimilation method， which is composed of two time windows with the first using EnSRF and the second using 4DVAR， and this combination can make good use of both EnSRF and 4DVAR. An Observing System Simulation Experiment（OSSE）is set up to investigate the ability to retrieve the true soil moisture profile with the new method by only assimilating the near-surface soil moisture observations into a land surface model. After comparing the performance of the three data assimilation schemes（i.e.，EnSRF，4DVAR，and H4DVAR），it is shown that the H4DVAR is superior to the rest two methods because it can quickly retrieve the soil moisture profile with less error. However， when small ensembles are used to calculate the background error covariance， the spurious long-range vertical error correlation between an observation and a state variable will have a bad influence on the estimation of soil moisture. Therefore the authors propose a method to tackle this issue by adding a correlation matrix with the elements defined by the Gaussian function into the original background error covariance. By this way， the rms error of the estimated soil moisture reduces from 0.036 cm3/cm3 to 0.016 cm3/cm3 with a relative reduction of 55.6%，and the most important is the large reduction of the errors in some soil moisture estimates， for example，the error at the depth of 90 cm reducing from 0.085 cm3/cm3 to 0.024 cm3/cm3 with a relative reduction of 71.8%.