This study focuses on improvement to the modeling of the evolution of the model error variance in the problem of assimilating satellite observations of chemical species. The model error variance evolution equation for the assimilation of CO is described here with localized sources in addition to transport and error growth. The assimilation of carbon monoxide (CO) observations from MOPITT is performed using a sub-optimal Kalman filter in the MOZART-2 chemistry-transport model. It is shown that this new approach can dramatically improve the ability of the assimilation to diverge from erroneous model-generated features.