Automatic front detection using a sequence of SST images

Recent and past algorithms designed to detect fronts in satellite-derived sea surface temperature fields do not perform as well as human experts. One reason for this performance gap is that the algorithms rely almost exclusively on the information contained in a single image, while human experts rely on information acquired by viewing many images when they analyze one image. This manuscript presents an algorithm which uses a sequence of images to detect temperature fronts in one sea surface temperature image. The multi-image edge detection algorithm starts by applying a single-image edge detection algorithm to the sequence of images under study. Next, fronts or portion of fronts from neighboring images, which were detected by the single-image algorithm and which match features in the current image, are identified as persistent. The coordinates of these persistent fronts are then passed to the single-image edge detection algorithm so that additional fronts can be detected without decreasing reliability. Because clouds affect the edge detection process, the identification of cloudy areas must be an integral part of that process. A multi-image cloud detection algorithm was devised to complement the modified version of an existing single-image cloud detection algorithm. In the region off Cape Hatteras, an additional processing step allows for the automated extraction of the Gulf Stream northern edge from all the other detected fronts. The performance of the multi-image edge detection algorithm, of various single-image algorithms and of a human expert are evaluated on a set of 98 images. For that purpose the location of the fronts obtained by applying various methods to the SST images is compared to the in-situ measures of the Gulf Stream position. With respect to both quality and the number of the detected edges, the multi-image edge detection algorithm is the only automated method which achieves results comparable to those obtained by a human expert