Orbital covariance propagation plays a vital role in the space situation awareness tasks such as the space conjunction analysis and space mission planning. Orbital errors, usually expressed by the orbital covariance matrix, can be propagated with linear models or non-linear models. Linear models are analytical and computationally fast since only the state transition matrix is needed. However, the propagation accuracy declines quickly with the time due to the linearization of highly nonlinear orbital dynamic systems. The non-linear models are accurate but computationally intensive, and thus are less attractive in the orbital covariance propagation for massive space debris. On the basis of the analysis of the propagated orbital errors, this paper proposes a dynamic calibration method to improve the accuracy of orbit error propagation. The method has three steps:the generation of realistic initial orbit covariance matrix, the linear propagation of the covariance matrix, and dynamic calibration of the propagated covariance using past observations. Experiments of real data processing show that the accuracy of calibrated orbit covariance can be improved by more than 60% compared to the linearly propagated covariance, and can provide more accurate propagated covariance for many space applications require high-accuracy such as the space collision probability computation.