The forecast quality from Numerical Weather Prediction (NWP) models and climate models depends on accurate initialisation. Therefore variables such as latent (LE) and sensible (H) heat flux from the land surface, which provide the lower boundary condition for NWP, need to be as accurate as possible at the beginning of a forecast period. Land Surface Models (LSMs) such as the CSIRO Biosphere Model (CBM) represent the exchange of energy and water between the earth's surface and lower atmosphere and are used to calculate LE and H. Soil moisture and temperature states of these models help partition incoming energy to the earth's surface between LE and H. Producing accurate predictions of LE and H is hindered by inaccuracies in LSMs such as uncertain initial model state conditions, errors in model forcing data, errors in model physics and a lack of data for accurately parameterising models. Data assimilation blends observations of a model variable(s) with a model to update/correct the model and achieve more accurate predictions than by running the model offline. Assimilating soil moisture observations into LSMs is a proven technique for improving predictions of soil moisture and hence LE and H. Although, assimilating soil moisture may not necessarily lead to optimal LE and H predictions due to a complex and non-linear relationship between them. Assimilating LE and H observations has not been thoroughly explored in the scientific community and could potentially produce more accurate LE and H predictions. This study compares the assimilation of soil moisture observations with that of combined LE and H observations into the CBM with the resulting impacts on predictions of LE, H and root zone soil moisture and temperature examined. Assimilation experiments were performed with a 1-year series of data using the Ensemble Kalman Filter (EnKF) algorithm. Observations and model forcing data were measured on a one-dimensional point scale at a site in south-eastern Australia. Errors were prescribed to initial conditions and to meteorological forcing variables. Observations were assimilated on typical remote sensing timescales - every 3 days for soil moisture (SMOS satellite) and twice daily with minimal cloud cover for LE and H (MODIS). Both observation sets were able to improve predicted soil moisture when assimilated compared to the offline model, with the soil moisture assimilation producing better results. Soil temperature predictions from both assimilation runs were worse than from the offline model indicating a warm bias. LE and H predictions are improved overall by both assimilation runs with LE and H assimilation producing the best predictions. It is demonstrated here that while surface soil moisture assimilation can improve soil moisture predictions in a LSM and consequently improve LE and H predictions, assimilating LE and H observations can produce more accurate LE and H predictions. Therefore the assimilation of LE and H observations into LSMs has the potential to provide NWP models with optimal LE and H estimates for initialisation.
