Previous studies have shown that land surface conditions such as soil temperature, soil moisture, soil type, vegetation cover amount, and vegetation type strongly influence the values of near-surface atmospheric variables. The near-surface temperature, dew point, and wind fields are all affected by differences in land surface conditions, as are precipitation and convective initiation. The land surface conditions listed above also play an important role in partitioning the net radiation received at the surface into the various components of the surface energy balance. Those components include sensible, latent, and ground (soil) heat fluxes. Observations of surface fluxes could potentially improve analyses initializing atmospheric forecast models if used to update various model fields.

This hypothesis is tested by examining results from assimilating Oklahoma Mesonet and Oklahoma Atmospheric Surface-layer Instrumentation System (OASIS) network observations into the MM5 model over a 24 h period, from 12 UTC 01 Aug 2004 – 12 UTC 02 Aug 2004. Three different 51 member ensemble forecast runs are compared: one simple forecast over the period with no data assimilation, one run with assimilation of Mesonet observations of 1.5 m temperature, 1.5 m relative humidity, 10 m wind, and surface pressure every hour, and a third run adding hourly assimilation of sensible and ground heat flux data observed by stations in the OASIS network. The residual component of the surface energy balance is used to estimate latent heat flux at sites where it is available, and is assimilated along with the sensible and ground heat flux observations. These runs are carried out on a domain the size of the body of the state of Oklahoma. Bias, mean absolute error, and root mean square error statistics generated by comparing the model analyses to National Weather Service (NWS) automated observations are calculated for each run.

An ensemble Kalman filter (EnKF) data assimilation method is chosen to incorporate the observations into MM5. The EnKF procedure involves creating an ensemble of forecasts and estimating the background error covariance from that ensemble. The implementation of the EnKF procedure used allows for serial processing of the observations and avoids the formation of huge covariance matrices. The EnKF approach has shown improvement over other data assimilation methods such as 3D-Var, especially when the true forecast uncertainty is strongly flow-dependent.