Data Assimilation in the Earth Sciences: Building on and Going beyond Classic Estimation Theory

Dennis McLaughlin
Ralph Parsons Laboratory, Department of Civil and Environmental Engineering, Massachusetts Institute of Technology

The increase in data available from remote sensing platforms and in situ sensor networks together with recent improvements in earth science modeling capabilities have prompted much interest in the field of data assimilation. Data assimilation/data fusion methods characterize environmental systems by using distributed models to merge uncertain data taken over different scales, with different accuracies, frequencies, and coverage. Data assimilation problems can be conveniently posed as stochastic estimation problems. The details depend on the objectives of the characterization (e.g forecasting vs. retrospective analysis) and on the focus of the estimation procedure (e.g. parameter vs. state estimation). In any case, classical estimation theory provides an attractive way to formulate and solve problems in a wide range of disciplines. Most texts and papers in the field rely on the assumptions and perspectives of this theory, including i) a least-squares perspective which often implicitly assumes that system states are Gaussian (i.e. fully characterized by their first two moments) , ii) a reliance on linear or quasi-linear theory, iii) assumptions that model and measurement errors are additive and independent of the states and iv) a tendency to neglect the computational issues that arise when the system state size is very large (as is often the case in earth science applications). Since all of these assumptions are likely to be problematic in real applications there is a growing awareness in the field that we need to go beyond classical estimation theory and develop a new set of methods that are appropriate for problems of realistic size and complexity. This talk highlights some of the critical conceptual/methodological issues facing environmental data assimilation today and briefly discusses some promising directions for future research.

Last Update: September 27, 2005