Adjoint retrieval of prognostic land surface model variables for an NWP model: Assimilation of ground surface temperature

Based on a 2-layer land surface model, a rather general variational data assimilation framework for estimating model state variables is developed. The method minimizes the error of surface soil temperature predictions subject to constraints imposed by the prediction model. Retrieval experiments for soil prognostic variables are performed and the results verified against model simulated data as well as real observations for the Oklahoma Atmospheric Surface layer Instrumentation System (OASIS). The optimization scheme is robust with respect to a wide range of initial guess errors in surface soil temperature (as large as 30 K) and deep soil moisture (within the range between wilting point and saturation). When assimilating OASIS data, the scheme can reduce the initial guess error by more than 90%, while for Observing Simulation System Experiments (OSSEs), the initial guess error is usually reduced by over four orders of magnitude. Using synthetic data, the robustness of the retrieval scheme as related to information content of the data and the physical meaning of the adjoint variables and their use in sensitivity studies are investigated. Through sensitivity analysis, it is confirmed that the vegetation coverage and growth condition determine whether or not the optimally estimated initial soil moisture condition leads to an optimal estimation of the surface fluxes. This reconciles two recent studies. With the real data experiments, it is shown that observations during the daytime period are the most effective for the retrieval. Longer assimilation windows result in more accurate initial condition retrieval, underlining the importance of information quantity, especially for schemes assimilating noisy observations.