Segmentation of Upwelling Regions in Sea Surface Temperature Images via Unsupervised Fuzzy Clustering

In this work the Anomalous Pattern algorithm is explored as an initialization strategy to the Fuzzy K-Means (FCM), with the sequential extraction of clusters, one by one, that simultaneously allows to determine the number of clusters. The composed algorithm, Anomalous Pattern Fuzzy Clustering (AP-FCM), is applied in the segmentation of Sea Surface Temperature (SST) images for the identification of Coastal Upwelling. Two independent data samples of two upwelling seasons, in a total of 61 SST images covering large diversity of upwelling situations, are analysed. Results show that by tuning the AP-FCM stop conditions it fits a good number of clusters providing an effective segmentation of the SST images whose spatial visualization of fuzzy membership closely reproduces the original images. Comparing the AP-FCM with the FCM using several validation indices to determine the number of clusters shows the advantage of the AP-FCM since FCM typically leads to under or over-segmented images. Quantitative assessment of the segmentations is accomplished through ROC analysis with ground-truth maps constructed from the Oceanographers' annotations. Compared to FCM, the number of iterations of the AP-FCM is significantly decreased.