Comparison of adaptive filtering techniques for land surface data assimilation

The accurate specification of modeling and observational error information required by data assimilation algorithms is a major obstacle to the successful application of a land surface data assimilation system. The source and statistical structure of these errors are often unknown, and poor assumptions concerning the relative magnitude of modeling and observation uncertainty degrade the quality of land data assimilation products. In theory, adaptive filtering approaches are capable of estimating model and observation error covariance information during the online cycling of a data assimilation system. To date, however, these approaches have not been widely applied to land surface models. Here, we implement and compare four separate adaptive filtering schemes in a data assimilation system designed to ingest remotely sensed surface soil moisture retrievals. Upon testing of each scheme via a synthetic twin data assimilation experiment, three of the four adaptive approaches are found to provide substantially improved soil moisture estimates. However, the specific model and observation characteristics of satellite-based surface soil moisture retrievals contribute to the relatively slow convergence of all schemes. Overall, results highlight the need to consider unique aspects of the land data assimilation problem when designing and/or evaluating the relative performance of adaptive filtering algorithms.