A version of the Seeded Region Growing approach for the automatic recognition of coastal upwelling from Sea Surface Temperature (SST) images is proposed. Our algorithm, derived from an approximation clustering model derives a homogeneity criterion in the format of a product rather than the conventional difference between a pixel value and the mean of values over the region of interest. It involves a boundary-oriented pixel labelling so that the cluster growing is performed by expanding its boundary iteratively. We introduce a self-tuning version of the algorithm in which the homogeneity threshold is locally derived from the approximation criterion over a window around the pixel under consideration. The window serves as a boundary regularizer. The algorithm has been applied to a set of 28 SST images of the western coast of mainland Portugal, and compared against a supervised version fine-tuned by maximizing the F-measure with respect to manually labelled ground-truth maps. The areas built by the unsupervised version of our algorithm are significantly coincident over the ground-truth regions in the cases at which the upwelling areas consist of a single continuous fragment of the SST map.
|Title of host publication||2015 Conference of the International Federation of Classification Societies (IFCS)|
|Publication status||E-pub ahead of print - 2015|
|Event||Conference of the International Federation of Classification Societies (IFCS 2015) - Bologna, Italy|
Duration: 6 Jul 2015 → 8 Jul 2015
|Conference||Conference of the International Federation of Classification Societies (IFCS 2015)|
|Abbreviated title||IFCS 2015|
|Period||6/07/15 → 8/07/15|