We consider the problem of recovering a moving visual scene recorded through a semi-reflective device, and extend the Sparse Component Analysis (SCA) method to three-dimensional data sources, using two different mixtures of the dynamic image source and the superimposed reflection, recorded at different polarizations, without having any a-priori knowledge about the structure and/or the statistics of the sources, or of the mixing matrix. We first apply our method on a simple physical example of separation of dynamic reflections, such as video signals recorded through the windshield of a car. In this example the required assumptions of linearity and stationarity are valid. A more interesting application deals with dynamic scenarios recorded from aircrafts or satellites, where it is desirable to extract a clear landscape view by separating it from a thin semi-transparent layers of clouds superimposed on the desired dynamic image. This is a more complex problem, since the mixtures are not stationary in space and the mixing coefficients vary in the presence of clouds. Further, the mixtures are not strictly linear and involve also multiplicative and convolutive components. We apply the 3D-SCA method in simulations of linearly mixed moving landscape contaminated by clouds.