Satellite and ground-based instruments for Earth observation produce large quantities of multi-source observations with different coverage, resolution, and quality. Utilizing such multi-source data for retrieval in the best way is still a largely unresolved question. The traditional techniques from spatial statistics and data fusion are not satisfactory, due either to computational constraints, difficulties in modeling and parameter estimation, or inability to provide uncertainty estimates. This talk will introduce the Gaussian conditional random field model as an alternative to the existing approaches. It has ability to model complex relationships in multi-source spatial-temporal data, it is conceptually simple, flexible, and computationally efficient. The potential of the proposed approach will be illustrated on the problem of aerosol retrieval. In the preliminary experiments, 28,374 collocated MODIS and AERONET observations collected at 217 AERONET sites during 2005 and 2006 were used. The results showed that the Gaussian conditional random fields were more accurate than the neural networks and the operational C005 retrieval algorithm. The results indicated that the proposed model can successfully use multi-source data, integrate multiple retrieval algorithms, and exploit spatio-temporal correlations in data.