Variational Gravity Data Assimilation to Improve Soil Moisture Prediction in a Land Surface Model

Accurate prediction of soil moisture in a land surface model (LSM) is critical in improving land surface and atmosphere interactions in the atmospheric general circulation models used in numerical weather prediction and global climate models. Gravity is a relatively new source of remotely sensed data, available since the launch of the twin Gravity Recovery And Climate Experiment (GRACE) satellites in March 2002. After correction for other signals, gravity provides a measure of terrestrial water storage (TWS) that includes soil moisture and groundwater. However, assimilating gravity data into a LSM to improve soil moisture prediction has not been widely investigated. A variational approach to assimilating gravity data is presented for one temperate site in south-east Australia (-35◦ latitude) for the year 2005. Parameters for the LSM are tuned to the site and forcing is compiled from local atmospheric observations. Gravity observations (corresponding to GRACE products) are synthetically generated from soil moisture and groundwater observations at the site. The model independent parameter optimiser PEST is used to determine initial conditions of soil moisture for a (monthly) moving window by minimising the modelled and observed gravity residual. The assimilated gravity observation is the difference of monthly gravity averages. The ability of variational assimilation to temporally and spatially disaggregate the lumped soil moisture signal present in gravity data is shown. Variational gravity data assimilation reduces predicted soil moisture (and TWS) bias when compared to the open loop predictions made with degraded LSM forcing data (Table 1). As well as reducing soil moisture bias, gravity data assimilation also improves the predicted soil moisture (and TWS) variance bringing it closer to the observed variance. While variational gravity assimilation retrieves the true (observed) soil moisture variance well, it cannot effectively correct the soil moisture bias in deeper layers resulting from A horizon (average depth of 30 cm) soil parameters being used throughout the profile (to a depth of 4.6 m). It is believed that assimilation performance would improve significantly if more realistic model soil parameters were available. This paper is a proof of concept, that variational methods can be used to assimilate gravity into a land surface model and it does improve soil moisture predictions.