The paper describes a learning method on sliding windows for estimating apparent motion on long temporal satellite sequences acquired over oceans. A "full model", which is defined on the pixel grid, is chosen to describe the dynamics of motion fields and images, based on heuristics of divergence-free motion and advection of image brightness by the velocity. The image sequence is split into small temporal windows that half overlap in time. Image assimilation in the full model is applied on the first window to retrieve its motion field. This makes it possible to define subspaces of motion fields and images and a "reduced model" is defined by applying the Galerkin projection of the full model on these subspaces. Data assimilation in the reduced model is applied on this second window. The process is iterated for the next window until the end of the whole image sequence. Each reduced model is then learned from the previous one. The main advantage of the approach is the small computational requirements of the assimilation in the reduced models that make it feasible to process in quasi-real time image acquisitions. Twin experiments have been designed to quantify the full model and the learning method on sliding windows and demonstrate the quality of the motion fields estimated by the approach.