Section: Scientific Foundations

Environmental image and data processing

Data with image nature, and especially satellite data, represent a huge amount of observations which is up to now largely unexploited by the environmental numerical forecast models. The operational state-of-the-art is mainly the assimilation of satellite data on a pixel basis: each pixel constitutes an independent information, expressed as a more or less complex function of the model's state variables. The challenge is to exploit the structure of the image observation by defining Image Assimilation methods: how to assimilate data with spatial and temporal coherency, such as observations of evolving fronts or eddies? Different issues are considered:

• Rewriting ill-posed image processing problems, usually addressed using numerical regularization techniques, through Image models. The Image Model describes the dynamics of the image sequence and makes it possible to formulate a data assimilation problem, where image observations are assimilated within the Image Model. This approach constitutes a relevant way to solve image processing problems, in which difficult issues such as occlusions or missing data are considered in a natural way. The usual spatial regularization is replaced by the temporal evolution laws for solving the underdetermination issue.

• Definition of Physical Image Models coupling variables image domain and from the forecasting model (in the same spirit than the qualitative Conceptual Models developed by meteorologists to describe specific phenomena and their signature on image data). The assimilation is then performed in two steps: first, in the Physical Image Model to yield “bogus” observations of the forecasting model's state variables, then directly in the forecasting model.

• Learning Image Models from image data. The aim is to define reduced basis, on which projecting the Navier-Stokes equations, to express the dynamics of the image sequence.

• Correcting location of structures from image data. The objective is to define data assimilation methods to modify the position of structures in case of a wrong location in the model representation.