Automated modeling of virtual worlds remains an open problem in computer graphics.
One of the biggest challenge lies in the modeling of large open worlds with a high level of detail,
which in turn induces problems in the proper handling of massive data sets along with a drastic
increase in creation time for artists. Simulation techniques and interactive editing both fail to
handle the scaling problem, the massive amount of data required demanding costly processes while preventing
efficient and precise user control. On the contrary procedural methods allow for effective on-the-fly
generation of models, without costly storage needs; but they still suffer from certain limitations:  
independent processing of different kind of objects composing a virtual scene (vegetation, terrain),
lack of user control and lack of multiscale integration of natural phenomena. The overall goal of this
PhD thesis is to develop new models based on classes of parametrized functions, combined with sparse modeling
and frequency modeling, working at various levels of detail, yielding sufficient controllability and
representation of the main features that pertain to the studied phenomena.