HR: 1340h
AN: H13C-0429    [Abstracts]
TI: Application of Unscented Kalman Filter for Higher Order Accuracy in the Assimilation of Near Surface Soil Moisture
AU: * Chintalapati, S
EM: chintala@uiuc.edu
AF: Department of Civil and Environmental Engineering, University of Illinois at Urbana-Champaign, 205 N Mathews Avenue, Urbana, IL 61801 United States
AU: Kumar, P
EM: kumar1@UIUC.EDU
AF: Department of Civil and Environmental Engineering, University of Illinois at Urbana-Champaign, 205 N Mathews Avenue, Urbana, IL 61801 United States
AB: Coupled land-atmosphere models are increasingly focused on using assimilated near-surface soil-moisture to improve the prediction of moisture and heat fluxes. Any assimilation scheme requires efficient handling of uncertainties, manifested through observational errors and errors in the land surface model (accounting for inaccuracies in specification of initial conditions, surface boundary parameters, forcing data, underlying model physics etc.), and better representation of inherent non-linearity in the system. The extended Kalman filter (EKF) is one of the well known algorithms for data assimilation in nonlinear systems. Unfortunately, EKF approximates the state variable as a Gaussian random variable (GRV), which then is propagated through the first-order linearization of the nonlinear system. This can seriously affect the accuracy or even lead to the divergence of the nonlinear system, often requiring resetting the state covariance to its positive definiteness. Ensemble Kalman filter (EnKF), which is easier to implement in complex models, provides a sub-optimal alternative. EnKF propagates the probability distribution of the state variable using a randomly selected set of realizations (ensemble) through the nonlinear model. The necessary statistics are then obtained from the ensemble analysis. But the dependence of its performance on random sampling (some statistical features may be lost) and requiring large number of ensemble members for non-Gaussian distributions can make EnKF computationally expensive and sometimes infeasible. The unscented Kalman filter (UKF) addresses this problem through an unscented transformation, wherein the state variable is approximated as a random variable, but it is now represented using a minimal set of sample points. These minimal sample points are deterministically chosen, rather than random sampling (as in Mote-Carlo techniques, EnKF) and completely capture the mean and covariance accurately to at least the 3rd order for any nonlinearity and the higher moments to at least 2nd order, with the same computational efficiency as EKF. The present study evaluates the performance of UKF for the prediction of soil moisture and associated fluxes, and their prediction errors, through implementation in NCAR's Land Surface Model.
DE: 3322 Land/atmosphere interactions
DE: 3360 Remote sensing
DE: 1800 HYDROLOGY
DE: 1833 Hydroclimatology
DE: 1866 Soil moisture
SC: Hydrology [H]
MN: 2004 AGU Fall Meeting