Abstract: Accurate background error covariance is the foundation for all advanced data assimilation systems. For four dimensions data assimilation (4D-Var), assimilating the observation data is converted to a question of cost function minimization which is constricted by atmosphere dynamic model. By adjusting the control vectors, the distance between model trajectory and real time observations reached its minimal value over whole assimilation time window. As background error covariance evolves according to the adjoint and tangent linear model, it can adapt to rapid development weather. However, most of operational 4D-Var systems still adopt simi-climatic background error covariance model compromised by huge dimensionality, which can't be exactly defi ned with all available information. As the rapid development of computer science, the problem of dimensionality can be released by ensemble method. Ensemble four dimensionality data assimilation (En4DVar) employed several independent perturbed analysis forecast cycles to remedy the limited information synchronously. In this scheme, fl ow-dependent background error covariance can be estimated from the differences between ensemble members. Several famous numeric prediction centers, such as ECMWF,Mete-France, adopted it to provide fl ow-depended background error covariance for the high-resolution determined 4D-Var system. In this thesis, the basic theory of the En4DVar method is demonstrated briefl y, followed by a description of currently application at ECMWF, and focusing on the disturbing, filtering, calibration as well as other key techniques for helping to improve the precision of estimates. The last part presents an investigation of some issues in current operation and possibly future research fi elds in the En4DVar.