Abstract : Understanding and forecasting the evolution of geophysical fluids is a major scientific and societal challenge. A good forecast must take into account all available information on the studied system. These informations include models, observations and a priori knowledge. Data assimilation techniques combine all these informations in a consistent way to produce model inputs. During the last decades, many satellites were launched to increase the knowledge of earth. They produce, among others, image sequences showing the dynamical evolution of geophysical processes such as depressions and fronts. These images sequences are currently under-utilized in data assimilation. This thesis presents a consistent approach for taking into account image sequences in variational data assimilation. After a presentation of images, their current use and its limitation, we introduce the concepts of interpretation level, image space and image operator used for direct image sequences assimilation. We also propose a new approach of regularization based on generalized diffusion for ill-posed inverse problems. Preliminary results on image processing and image sequences assimilation show a promising approach that solve most of the problems encountered with classical approaches of regularization.