Motion analysis in oceanographic satellite images using multiscale methods and the energy cascade

Motion analysis of complex signals is a particularly important and difficult topic, as classical Computer Vision and Image Processing methodologies, either based on some extended conservation hypothesis or regularity conditions, may show their inherent limitations. An important example of such signals are those coming from the remote sensing of the oceans. In those signals, the inherent complexities of the acquired phenomenon (a fluid in the regime of fully developed turbulence—FDT) are made even more fraught through the alterations coming from the acquisition process (sun glint, haze, missing data etc.). The importance of understanding and computing vector fields associated to motion in the oceans or in the atmosphere (e.g.: cloud motion) raises some fundamental questions and the need for derivating motion analysis and understanding algorithms that match the physical characteristics of the acquired signals. Among these questions, one of the most fundamental is to understand what classical methodologies (e.g.: such as the various implementations of the optical flow) are missing, and how their drawbacks can be mitigated. In this paper, we show that the fundamental problem of motion evaluation in complex and turbulent acquisitions can be tackled using new multiscale characterizations of transition fronts. The use of appropriate paradigms coming from Statistical Physics can be combined with some specific Signal Processing evaluation of the microcanonical cascade associated to turbulence. This leads to radically new methods for computing motion fields in these signals. These methods are first assessed on the results of a 3D oceanic circulation model, and then applied on real data.