Presenter/Author Information

Isabelle Herlin
Dominique Bereziat
Etienne Huot

Keywords

data assimilation, motion, satellite images

Start Date

1-7-2012 12:00 AM

Description

Simulation models and image data are simultaneously available for numerousscientific domains, such as oceanography and meteorology. They are indeed two complementarydescriptions of the same complex system. Data Assimilation is a well-knownmathematical technique used, in environmental sciences, to improve forecasts obtainedfrom the simulation models, thanks to the observation data. One class of data assimilationalgorithms, named 4D-Var, globally adjusts the model output to the observations, thatare available over a period of time. The question of how to derive accurate characteristicfeatures from images, with an optimal use of the simulation model, is of major interestfor the image processing community. In this article, we consider applying data assimilationmethods for motion estimation on a sequence of satellite images acquired over theocean. We describe various strategies that can be derived in the framework of variationaldata assimilation (4D-Var). They mostly depend on the choice of the state vector itself.According to this definition, the dynamics has to be described and observation operatorsspecified in order to characterize the information displayed by the image sequence. Wedetail the mathematical setting of these strategies and analyze their properties. Resultsare provided on twin experiments to quantify the methods and on satellite acquisitionsacquired over the Black Sea.

Share

COinS
 
Jul 1st, 12:00 AM

Image Assimilation and Motion Estimation of Geophysical Fluids

Simulation models and image data are simultaneously available for numerousscientific domains, such as oceanography and meteorology. They are indeed two complementarydescriptions of the same complex system. Data Assimilation is a well-knownmathematical technique used, in environmental sciences, to improve forecasts obtainedfrom the simulation models, thanks to the observation data. One class of data assimilationalgorithms, named 4D-Var, globally adjusts the model output to the observations, thatare available over a period of time. The question of how to derive accurate characteristicfeatures from images, with an optimal use of the simulation model, is of major interestfor the image processing community. In this article, we consider applying data assimilationmethods for motion estimation on a sequence of satellite images acquired over theocean. We describe various strategies that can be derived in the framework of variationaldata assimilation (4D-Var). They mostly depend on the choice of the state vector itself.According to this definition, the dynamics has to be described and observation operatorsspecified in order to characterize the information displayed by the image sequence. Wedetail the mathematical setting of these strategies and analyze their properties. Resultsare provided on twin experiments to quantify the methods and on satellite acquisitionsacquired over the Black Sea.