Front detection on satellite images based on wavelet and evidence theory: Application to the sea breeze fronts

摘要

This paper is concerned with the detection of fronts in satellite images. We focus on some specific textured patterns of clouds that are visible on Meteosat Second Generation (MSG) images and generated at the so-called "sea breeze front". This is the limit of the penetration of the sea breeze inland. The sea breeze circulation is a phenomenon that arises when land and sea surface temperatures reveal strong variations. This generates a landward wind that creates a cloud-free area starting from the coast line and ending at the sea breeze front. With the new geostationary meteorological sensors like MSG, this band of cloud-free area can clearly be seen. The automatic analysis of the sea breeze front with such image sensors (instead of using local measurements) is then of great importance. It has the precious advantage to extract huge amount of data and to get rid of the use of local sensors. Unfortunately, from an image processing point of view, this front appears as the limit of a very textured area. It is sometimes disturbed by clouds located in higher layers of the atmosphere. Due to this complexity, conventional detection methods issued from computer vision are not adapted. In this paper we propose an approach that automatically detects fronts in images and we apply this framework to the sea breeze fronts. The methodology is based on the well-known active contour method (or "snake") issued from the computer vision community. The specific textures involved as well as the transparency phenomena are dealt with some properties of the wavelet decomposition of the images. This decomposition enables to compute several criteria related to the presence or not of a front that are combined using the Dempster-Shafer theory. The validation of our approach is done on synthetic and real data. It is important to outline that the presented theoretical framework is not only devoted to the detection of the sea breeze front but can also be used to detect any others textured patterns.