In environmental sciences, modeling is used to provide, at a given date (past or future), the state of the environment. This is done through the development and application of appropriate models. In the case of atmospheric modeling, studying interaction of underlying events (meteorology, atmospheric chemistry in the various phases of matter, radiative transfer) is a scientific objective in itself (notably for describing aerosols) and it leads to complex models. For these models, the problem of parameterization (change of scale, description of phenomena at subgrid scales) is very important.
In the meantime, the question of models' representativeness is surfacing: how to build models with low degrees of freedom from the description of these complex processes ? How to make these models useful for impact studies and able to perform data assimilation ?
Models alone are generally insufficient because they need unknown
input (initial and boundary conditions, parameters, ...). These
unknown inputs require complimentary data. Consequently, there is
presently a growing interweaving between models and observation
systems (data providing systems). Among these observation systems,
satellite data will play a major role in the short term, through
acquisitions from all the environmental missions started these latter
years. Data assimilation procedures, whatever their methodological
nature, allow this data and models coupling.
In such a context, the whole data/models/output chain must be considered, with many resulting problematics:
Satellite data is in different forms (static, dynamic, multispectral, ...). Objects have particular characteristics in terms of motion, temporal evolution (topological changes, fragmentation, ...), geometry (vortices, ...). These characteristics necessitate appropriate mathematical modeling adapted to the resolution of problems arising from the analysis of these data. In other respects, satellite measure is obtained from a known physical process modelled by radiative transfer, and it makes sense to use the knowledge of data acquisition inside processing, with the goal of setting up image processing algorithms taking into account physical models. For example, in the domain of motion estimation in meteorological images, one can write down temporal conservation hypothesis applied to atmospheric physical parameters obtained by inverting a radiative transfer model, instead of the raw measurements.
Although methods presently exist (notably data assimilation with the variational approach), many questions are still open:
Beyond listed themes above, the development of software chains that allow closely coupling of models and data, raise specific problematics with respect to input (databases) and output (visualization of results).
We think that these two points are crucial in the context of environmental applications. For example, in atmospheric chemistry, databases needed by models are very heterogeneous and distributed. Moreover, model outputs fields can be of large dimensions (3D spatial fields evolving in time, with potentially dozens of chemical species); presently, models output representation is poor and with no comparison with potentially available data.
Resorting to visualization tools allows the handling of complexity (e.g. virtual reality), which can be an interesting domain to explore in conjunction with involved INRIA projects, be it geared towards the scientific community (how to better represent phenomena), towards applied partners (better visualize results), or towards decision makers (having potential impacts scenarios).
All these questions underpin our scientific project, defining its long-term strategy.