A Dynamic Approach to Addressing Observation-Minus-Forecast Mean Differences in a Land Surface Skin Temperature Data Assimilation System

In land data assimilation, bias in the observation-minus-forecast (O-F) residuals is typically removed from the observations prior to assimilation by rescaling the observations to have the same long-term mean (and higher-order moments) as the corresponding model forecasts. Such observation rescaling approaches require a long record of observed and forecast estimates, and an assumption that the O-F mean differences are stationary. A two-stage observation bias and state estimation filter is presented, as an alternative to observation rescaling that does not require a long data record or assume stationary O-F mean differences. The two-stage filter removes dynamic (nonstationary) estimates of the seasonal scale O-F mean difference from the assimilated observations, allowing the assimilation to correct the model for synoptic-scale errors without adverse effects from observation biases. The two-stage filter is demonstrated by assimilating geostationary skin temperature (Tsk) observations into the Catchment land surface model. Global maps of the O-F mean differences are presented, and the two-stage filter is evaluated for one year over the Americas. The two-stage filter effectively removed the Tsk O-F mean differences, for example the GOES-West O-F mean difference at 21:00 UTC was reduced from 5.1 K for a bias-blind assimilation to 0.3 K. Compared to independent in situ and remotely sensed Tsk observations, the two-stage assimilation reduced the unbiased Root Mean Square Difference (ubRMSD) of the modeled Tsk by 10 of the open-loop values