Based on a 2-layer soil-vegetation or land surface model, a rather general 4DVAR data assimilation framework for estimating the state variables of the model is developed. The adjoint method was used to minimize the error of surface ground temperature predictions subject to constraints imposed by the model system. We performed retrieval experiments for soil prognostic variables and verified the results against Oklahoma Atmospheric Surface Layer Instrumentation system (OASIS) observations. A recently implemented revision to the force-restore soil temperature scheme was found vital for the performance of the retrieval scheme as applied to real data. Using synthetic data, we systematically documented the following basic issues of land data assimilation: robustness of the retrieval scheme as related to information redundancy; temporal measurement sparsity; and the physical implications of the adjoint variables and their usage in sensitivity studies. Through the sensitivity analysis, we conclude that whether or not initial soil moisture condition that optimizes the superficial soil moisture description leads to an optimal estimation of the surface fluxes actually depends upon the vegetation coverage and growth conditions. In the OASIS measurements assimilation, we explained why the daytime period is the most informative period for the performance of the retrieval scheme. It is also shown that filter-based methods such as fitting an error covariance structure to interpolate model-data misfit at one time level to other levels is not essential for successful land data assimilation under incomplete measurements. The longer the assimilation window is the more accurate are the retrieved initial soil moisture conditions. This again underlines the importance of information redundancy, especially for schemes assimilating noisy observations.