An adaptive ensemble Kalman filter for soil moisture data assimilation

In a 19- year twin experiment for the Red- Arkansas river basin we assimilate synthetic surface soil moisture retrievals into the NASA Catchment land surface model. We demonstrate how poorly specified model and observation error parameters affect the quality of the assimilation products. In particular, soil moisture estimates from data assimilation are sensitive to observation and model error variances and, for very poor input error parameters, may even be worse than model estimates without data assimilation. Estimates of surface heat fluxes and runoff are at best marginally improved through the assimilation of surface soil moisture and tend to have large errors when the assimilation system operates with poor input error parameters. We present a computationally affordable, adaptive assimilation system that continually adjusts model and observation error parameters in response to internal diagnostics. The adaptive filter can identify model and observation error variances and provide generally improved assimilation estimates when compared to the non- adaptive system.