PCA before clustering

K-means sees distances, not labels. Six clusters is a declared benchmark choice, because there are six known classes — not a discovery of how many land-cover types exist.

Labels enter only afterwards, through the adjusted Rand index (ARI): a permutation-invariant agreement score, 1 for exact agreement and about 0 at chance. Reference ARI is about 0.522 for all features, 0.508 for two PCs and 0.518 for twelve; silhouettes, measured in one common standardized space, are about 0.350, 0.346 and 0.351.

A dramatic reduction in dimension gives similar clustering scores — and PCA does not make the clusters coincide with the classes.