NON-GAUSSIAN ENSEMBLE FILTERING AND ADAPTIVE INFLATION FOR SOIL MOISTURE DATA ASSIMILATION

Abstract

The forecast error distribution in modern day land data assimilation systems is typically modeled as a Gaussian. The explicit tracking of only the first two moments can be problematic when trying to assimilate bounded quantities like soil moisture that are more accurately described using more general parameterizations. Given this issue, it is worthwhile to test how performance of land models is affected when the accompanying data assimilation system abides by a relatively more relaxed set of underlying assumptions. To study this problem, we perform experiments using the ensemble Kalman filter (EnKF) and rank histogram filter (RHF) to assimilate surface soil moisture content observations into the NASA Catchment land surface model. The EnKF acts as the traditional (Gaussian) standard of comparison whereas the RHF represents the novel and more general data assimilation method. An additional parameter of our tests is the usage of an adaptive inflation scheme that is only applied to the ensemble prior. This is done in an attempt to mitigate the negative effects of systematic deficiencies not accounted for by either filter. The examinations were carried out at a number of globally-distributed test locations, deliberately coinciding with sites used to validate NASA SMAP soil moisture retrieval products. Initial comparisons of the two filtering approaches in a perfect model context show both filters to provide significant benefits to the soil moisture modeling problem, with the RHF edging out the EnKF as the more performant filter. The relative performance gain of the RHF was most noticeable with respect to bias mitigation metrics and to the surface-level anomaly correlation scores, an interesting result given that neither filter is formulated to explicitly accommodate a systematic bias. When additionally applying adaptive inflation, both filters showed improvement in skill but such improvements were not significant. The use of synthetic observations and lack of a bias correction implementation may have led to exaggerated results. To address this concern, the experiments were performed again but using real observations from SMAP soil moisture retrievals, with in situ validation data proxying as truth. A robust bias correction scheme was used as well to more closely approximate practices used in operational settings. The RHF continues to show better metrics than the EnKF, but no longer in a statistically significant sense. A similar result was noted with respect to inflation usage. The most likely reason for this outcome is the low observation count. The findings obtained from the data assimilation experiments in this dissertation offer insight on how best to focus development efforts in soil moisture modeling and land data assimilation.