The Silhouette Index and the K-Harmonic Means algorithm for Multispectral Satellite Images Clustering
Habib Mahi, Nezha Farhi, Kaouter Labed, Dalila Benhamed · 2018 International Conference on Applied Smart Systems (ICASS) · 2018
In this paper, we investigate the performance of the Silhouette index by applying the K-Harmonic Means algorithm to cluster different types of remote sensing datasets. The preliminary experimental results showed that the junction between the K-Harmonic Means algorithm and the Silhouette index is suitable for remote sensing images which present an insignificant degree of overlapping between clusters. Indeed, the Silhouette index returns a clear maximum which is synonym of the correct number of cluster. In contrast, the Silhouette index fails to return the real number of clusters when we deal with remote sensing data of which the clusters present a high degree of overlapping. To cope with this drawback, the angle based method was introduced to detect the correct number of clusters; the results prove the efficiency of the proposed process. Also, two comparisons were conducted, the first one between the well-known K-Means algorithm and the K-Harmonic Means and the second between the Silhouette index and the WB index in order to demonstrate the effectiveness of K-Harmonic Means algorithm and also the Silhouette index.