Assimilation of image data

This is a newborn exploratory topic. Most of the data assimilation theory consider observations as mathematical data vectors without any structure, all the structural information is handled through specification of the background and observation error covariances matrices. In other words, observations are used as collection of pixels, and the relationship between neighbouring pixels only appears rather through a rather cumbersome modelling of spatial error correlation structures.

Although this approach has work rather well through data assimilation of raw data in the 4D-Var algorithm (where much of the structural information is provided by the numerical model itself), it is very computer intensive. At the same time, structural information in remote-sensed images (satellite and radars) is still rather poorly used, for various reasons.

The aim of this project is to explore new ways of extracting significant data from images (in the broad sense of the term: it can be a sequence of multispectral images) in order to inject important information into numerical models of the atmosphere; a process known as "bogusing" in NWP. The key step is the synthetic characterization of the considered image feature, called an "object", and its translation into pseudo-observations suitable for assimilation into 3D-Var or 4D-Var NWP system, to gether with more conventional observations.

The following weather phenomena are being considered for this approach:

The image data comes from satellites (multispectral IR and microwave radiances, and possibly scatterometre wind) and radars (3D maps of Doppler wind components and polarimetric reflectivity).

References