We present a description, an evaluation, and a comparison of four methods designed to produce objective and physically consistent maps of ozone concentration fields. These methods are based on the use of a chemistry-transport model (CTM) and available observations. In most existing analysis systems, the error covariance is modeled using assumptions of homogeneity and isotropy. However, these assumptions may fail in the case of strongly heterogeneous terrain or emission patterns. Therefore we propose a simple method for specifying an anisotropic and heterogeneous background error covariance model in a statistical interpolation. We illustrate the positive impact of the implementation of this model. Since the covariance model is independent of the state of the atmosphere and invariant in time, we simultaneously test kriging techniques that are generally used for spatial interpolations. Kriging is applied to the observed values or to the differences between simulated and observed concentrations, namely the innovations. We perform an objective statistical validation of all methods for data recorded over an entire summer season, in contrast to other evaluations of data assimilation methods made in air quality modeling. We demonstrate that the RMS error of the ozone analyses at the surface is 30--50% smaller than the one from simulations, regardless of the method used. Most of the time, the kriging method applied to the innovations gives results equivalent to that of the anisotropic statistical interpolation, in spite of very different formulations. In addition, we show that this method outperforms the classical kriging method applied to only observations. The information provided by the CTM is therefore essential to a good-quality representation of ozone patterns such as city plumes or urban/suburban gradients. We also use the Atmospheric Pollution Over the Paris Area (ESQUIF) airborne measurements to demonstrate that the anisotropic method efficiently corrects CTM fields in altitude.