Coastline extraction techniques from multispectral satellite images are of great interest for protection and monitoring of coastal areas. In this regard, the Sentinel-2 satellites can give a great contribution thanks to their wide coverage of the earth's surface. These images can be processed by GIS software, so as to detect the sea from all the rest. However, the traditional supervised classification requires the involvement of the operator to create suitable training sites: this approach, in addition to being associated to the operator's skill, often takes a long time to be completed. This contribution presents a study carried out on Sentinel-2 dataset and proposes the application of an unsupervised classification method, the k-means, on four different classification indices. The coastlines extracted by unsupervised classification are therefore compared with the coastline manually vectorized from the RGB composition. The results demonstrate the effectiveness of k-means for distinguishing, in the images produced by the indices application, two clusters (water / no-water) in a reduced time lapse if compared with the traditional supervised techniques.