Background error covariance is an important part of variational data assimilation system，which is used to spread the observation information to other grid points and vertical levels of the model. In order to model the inhomogeneity and anisotropy in horizontal error functions of raw background error covariance，a new method is proposed to model horizontal error functions based on orthogonal wavelet transforms. The results of experiments show that the new method successfully modeled the heterogeneity and anisotropy included in the local correlation functions and depicted the structure and feature of raw covariance correctly.