The impact of incorrect model error assumptions on the assimilation of remotely sensed surface soil moisture

Data assimilation approaches require some type of state
forecast error covariance information in order to optimally merge model predictions
with observations. The ensemble Kalman filter (EnKF) dynamically derives such
information through a Monte Carlo approach and the introduction of random noise in
model states, fluxes, and/or forcing data. However, in land data assimilation,
relatively little guidance exists concerning strategies for selecting the appropriate
magnitude and/or type of introduced model noise. In addition, little is known about
the sensitivity of filter prediction accuracy to (potentially) inappropriate assumptions
concerning the source and magnitude of modeling error. Using a series of synthetic
identical twin experiments, this analysis explores the consequences of making
incorrect assumptions concerning the source and magnitude of model error on the
efficiency of assimilating surface soil moisture observations to constrain deeper root-
zone soil moisture predictions made by a land surface model. Results suggest that
inappropriate model error assumptions can lead to circumstances in which the
assimilation of surface soil moisture observations actually degrades the performance
of a land surface model (relative to open-loop assimilations that lack a data
assimilation component). Prospects for diagnosing such circumstances and
adaptively correcting the culpable model error assumptions using filter innovations
are discussed. The dual assimilation of both runoff (from streamflow) and surface soil
moisture observations appears to offer a more robust assimilation framework where
incorrect model error assumptions are more readily diagnosed via filter innovations.