84th AMS Annual Meeting

Wednesday, 14 January 2004: 4:30 PM
Reanalysis Land Data Assimilation Using Ensemble Techniques (INVITED)
Room 6E
Susan C. Dunne, MIT, Cambridge, MA; and D. Entekhabi
Poster PDF (92.4 kB)
The objective of this research is to develop a land surface data assimilation that is directed towards re-analysis rather than filtering. Land data assimilation using filters, such as extended or ensemble Kalman filters, ingests the data sequentially as they become available. This framework is ideally suited for forecasting problems where the observations up to the current time are used to update the initial conditions for forecasts into the future. This is the design of Kalman filters. Nevertheless in the land data assimilation problem the objective is often not forecasting but reanalysis. The goal is to use all available observations to produce a consistent data set of land surface states and fluxes. Thus observations before as well as after a given time are useful in the estimation. Kalman filters are not capable of this operation. Here we introduce so-called smoother estimators that use all the observations in a given time period to develop reanalysis estimates at all times within the period. Traditional smoothers such as the Rauch-Tung-Striebel (RTS) smoothers are optimal batch estimators as Kalman filters are optimal sequential estimators. Nevertheless in their traditional form both are limited to linear systems. Linearization of either the Kalman filter or the RTS smoother is seriously prone to unstable growth of the covariance matrices. Any artificial limits on the propagation of the covariance matrix results in suboptimal filters and poor estimation. Ensemble techniques have been developed for Kalman filters in order to avoid the linearization of the system equation. Several land data assimilation studies have used this powerful technique that is based on Monte Carlo estimation of sample covariances at estimation times. We present ensemble smoothers that also use random replicates of the land surface model to estimate statistics of the system at estimation times. In this study the batch smoother is an estimator that uses all observations in a given window to estimate the land surface states and fluxes at all times in that window. It is thus a reanalysis land data assimilation system well suited for estimating consistent states and fluxes based on all observations at all available times in the window. Smoothing combines future and past data to make an estimate rather than filtering sequentially through the dataset. This additional information on how the system evolves yields improved estimates of the state at the present time. We focus on the use of L-band microwave brightness temperatures at various temporal and spatial resolutions that will be assimilated into a mainstream land surface model (e.g. NCAR Land Surface Model) to obtain consistent estimates of soil moisture and surface fluxes. The hypothesis of this research is that by building an ensemble smoothing algorithm from a successful ensemble filter algorithm, we can extract more information from the observational data available. This is especially relevant for estimating profile soil moisture below the penetration depth of observations. The correct estimation of the profile will significantly impact the surface evaporation estimates where vegetation is present. In this study a smoothing algorithm has been developed from a successful ensemble Kalman filter algorithm. Before the application of the ensemble smoother for land data assimilation we tested the algorithm against the optimum smoother for an linear AR(1) model. The analytic solution for this case exists. Next results will be presented of its application with the NCAR Land Surface Model to the soil moisture estimation problem using field data (e.g. from SGP97,SGP99,SMEX02). The soil moisture and surface flux estimates obtained using the smoother will be compared to those obtained using the filter and to ground truth observations. The added value of including future observations will be weighed against the increased computational burden.

Supplementary URL: