On Clustering Performance Indices for Multispectral Images

摘要

Clustering of multispectral image pixels can be a exploratory tool to analyze the contents of the image in the absence of ground truth information. The validity of the clustering algorithms can be quantified computing several performance indices. Each performance index enhances some statistical property of the obtained data partitions. Performance indices are not equivalent, and they can even lead to quite different conclusions from the same data partitions. To show this, we have applied two well known clustering algorithms (K-means, Fuzzy c-means) and some supervised classification algorithms to a well known multispectral image. We compare the ground truth partition with the ones found by the clustering and supervised algorithms The values of the diverse performance indices over the same partitions vary and can lead to quite different conclusions.