The analysis of soil moisture for the initialization of a mesoscale numerical weather-prediction (NWP) model is considered subject to operational constraints, both in terms of computational cost and data availability. A variational technique is used to analyse the soil moisture by assimilating screen-level observations of temperature and relative humidity. We consider a simplified bi-dimensional (z and t) variational approach (simplified 2D-VAR), where the estimate of the observation operator is obtained from extra integration(s) of the numerical model. The fundamental assumptions of the method are first evaluated: linearity of the observation operator, horizontal decoupling between grid points, and truncation of the control variable space (variable decoupling), that allow the simplified 2D formalism. Thus, the variational method is applied at each grid point separately and the gain matrix is computed from finite differences given the small dimension involved. The 2D-VAR technique keeps count of the full physics of the model, so the corrections applied to the control variable are adapted to the current meteorological conditions and the grid-point characteristics (texture and vegetation), as well as to the previous soil state. The linear estimate of the observation operator is studied in detail to optimize its evaluation. The validation of the method is shown with simulated observations, and the assimilation of real observations is performed with different time-windows. A sequential assimilation cycle on a 6-hour time-window allows the comparison with the optimum interpolation technique, while a 24-hour window is considered to extend the temporal consistency of the assimilated observations in the analysis. Results from the performed analyses with the simplified 2D-VAR method show a good retrieval of soil moisture, and a comparison with other initialization methods is also provided. Copyright © 2004 Royal Meteorological Society.