A methodology is proposed in order to combine information from radar precipitation with other observations (e.g. screen-level temperature and relative humidity) in a soil analysis scheme based on an extended Kalman filter. A preliminary study is performed over the Czech Republic for one month in July 2008 using three hour rainfall accumulations derived from two C-band radars and a land-surface scheme forced by short-range forecasts from a limited-area model.
The Jacobian matrix of the observation operator is examined to make optimal choices for the estimation of the Kalman gain matrix. It is shown that the size of the perturbation for computing Jacobian matrix elements with finite differences has to be carefully chosen, since too small values lead to unphysical negative elements whereas too large values reduce the spatial variability considerably.
After a log-transform of the precipitation field, the corresponding errors are more compatible with the Gaussian hypothesis of the Kalman filter. However, at locations where model rainfall is underestimated, positive soil moisture increments are much too low.
Finally, two methods for combining the assimilation of screen-level observations with radar precipitation are compared. A first evaluation shows more accurate soil analyses (leading to reduced screen-level parameter forecast errors) when both sources of information are considered to correct soil moisture contents.
Avenues for improving the specification of observation and model errors of the precipitation field are also discussed.