University of Konstanz
Graduiertenkolleg / PhD Program
Computer and Information Science

PhD Program Spring School 2006


Statistical Approach to Evaluation of Thinning Algorithms used in Assimilation of Meteorological Data

speaker Vladimir Bondarenko
 
date March 10, 2006
 
abstract Mathematically, data assimilation can be seen as a large multidimensional optimization problem. Its solution is a model state vector (Analysis) accurately describing the state of the earth atmosphere at certain date given a vector of available measurements data (or observations vector), and another model vect or (Background) containing the prediction of the forecast model for this date. Neither measurements data nor model prediction is error free. In order to ge t an optimal Analysis, exact correlation coefficients for observations and Background errors are necessary. Unfortunately, such information for measurement errors is usually not available. To decorelate the observations, a thinning procedure is applied prior to data assimilation at most weather prediction cen ters. Thinning decreases the observations density simply by deleting of some (redundant) measurements data. This reduces the effective error-correlation of nearby measurements. However, rejection of some measurements may lead to loss of potentially important information. Sophisticated thinning algorithms are necessary to find an optimal measurements subset, which would result in the most accurate forecast. Until now, there are several thinning algorithms implementing different thinning strategies. Due to very high computational costs of a forecast pro duction, it is hardly possible to compare these algorithms by evaluating their impact on the forecast quality directly. In the present work, another statis tical approach to evaluation of thinning algorithms was developed. It is based on random simulations of Background and measurements data and consequent sol ving of the assimilation optimization problem for an optimal Analysis. The quality of the Analysis (not the forecast) is then used as the evaluation criter ion. In the talk, the approach will be explained in details and the first experimental results will be presented.