Assimilation of Surface Soil Moisture Data into a Model

Jeffrey P Walker, Garry R Willgoose and Jetse D Kalma

Poster Presentation

Model estimates of soil moisture suffer from errors in the model physics, initialization and atmospheric forcing data. While remote sensing provides a measurement of the spatial distribution of surface soil moisture content, it does not provide any direct information on the soil moisture profile or its variation during the inter-observation period. In contrast, point measurements give a continuous estimate of the soil moisture profile at a point, but do not give the spatial distribution. Hence an optimal estimate of the soil moisture may be obtained by utilizing the complimentary attributes of the three approaches.

The usefulness of surface soil moisture data for estimating the spatial and temporal variation of soil moisture has been illustrated using field data and a soil moisture model. The field data consisted of atmospheric forcing, surface and profile soil moisture data collected for a 6ha catchment in a temperate region of Eastern Australia. Surface soil moisture data was measured on a 20m regular grid using 15cm TDR probes to replicate remote sensing observations. The soil moisture model was calibrated to 10 months of measured soil moisture profile data at 13 locations throughout the catchment. Then, starting from a poor initial condition, the soil moisture model was run for a separate 1 month period and the surface soil moisture observations assimilated using the Kalman filter. Model predictions were compared with the measured soil moisture profile data and predictions without assimilation.

This study has shown that assimilation of surface soil moisture data into a soil moisture model can yield a significant improvement in the soil moisture profile estimation. It has also shown that when using an appropriate assimilation scheme, accurate initialization of the forecasting model is not critical, yielding an average rms error in total soil moisture storage of 6% v/v as compared to 13% v/v when surface observations were not assimilated. Moreover, it has been shown that the updating interval is relatively unimportant when the forecasting model has a good calibration and forcing data has a high level of accuracy.