In part 1 of this study (Prigent et al., 2004), in situ measurements were used to analyze and describe the sensitivities of satellite measurements (i.e., active and passive microwave observations and surface skin temperature diurnal cycle amplitude) to the soil moisture variations to describe the complex relationships that exist between them. Soil moisture was considered in the first 10-cm layer on a 0.25° equal-area grid and a monthly timescale. In this study, the lessons from the first paper are exploited to document the sensitivity of the satellite data to the global large-scale variations of soil moisture. A statistical model based on neural networks is developed to link the satellite observations and soil moisture estimates. Given the lack of available in situ soil moisture measurements on a global basis, National Centers for Environmental Prediction (NCEP) and European Centre for Medium-Range Weather Forecasts (ECMWF) soil moisture reanalyses are used as a realistic global indicator of soil moisture. As a consequence, the statistical model cannot be considered as a retrieval scheme per se, but it shows the feasibility of such an approach. It also quantifies the information content that can be expected from the satellite observations. Applications of such a statistical model include checking the consistency of surface model, and as the basis for variational assimilation of satellite observations into a numerical surface model.