Abstract:
Profile soil moisture is an important variable for weather and climate prediction, flood forecasting, or the determination of groundwater recharge. Passive microwave remote sensing in the L-band (1.4 GHz) can provide information about the physical temperature and the dielectric properties of the land surface which are, in turn, related to soil moisture in the top few centimeters. In this paper we describe a four-dimensional data assimilation (4DDA) algorithm which uses a physically-based model of soil moisture and heat transport in order to extract information about soil moisture profiles and land-atmosphere fluxes from L-band microwave measurements.

The coupled soil moisture and temperature model which forms the basis of the 4DDA algorithm is designed to capture the key physical processes while remaining computationally efficient. We divide the computational region into one-dimensional vertical cells (or pixels). Moisture transport in each pixel is described with Richards' equation while energy transport is described with a force-restore model. We account for model errors by treating the surface forcings and parameters in different pixels as random fields which are correlated over time and space.

The meteorological and soil parameter inputs to the model are available at a finer scale than the brightness measurements. The measurement operator of the 4DDA algorithm accounts for this difference in scales as well as for the nonlinear relationship between soil moisture, soil temperature, and brightness temperature. This makes it possible to estimate soil moisture profiles at a finer scale than the resolution of the brightness data (down-scaling).

The estimates are derived from a variational least-squares algorithm. Through its implicit propagation of the error covariances, the algorithm is very efficient and thus able to provide optimal estimates without the simplifications that are needed in large-scale Kalman filtering applications. Variational assimilation methods interpolate and extrapolate the data in a dynamically consistent way.

Our algorithm is tested on the data set which has recently been collected during the 1997 Southern Great Plains (SGP97) experiment in central Oklahoma. Those data provide the first series of large-scale L-band images available. The paper presents preliminary results as well as an assessment of the computational and operational feasibility of the approach.