K-Means Clustering

K-Means Clustering#

K-Means groups every pixel into one of K clusters by spectral similarity, producing an unsupervised classification of the scene.

How it works#

Each pixel’s spectrum is a point in band-space. K-Means picks K initial centroids, then repeats two steps until the centroids stop moving (or a limit is hit): assign each pixel to its nearest centroid (Euclidean distance), then recompute each centroid as the mean of its assigned pixels.

The output is a single-band label image where each pixel’s value is its cluster index (0…K−1); nodata pixels are set to −1. The final cluster centroid spectra are stored and viewable. Bad bands are excluded from the distance computation.

The result is added as a new dataset named K-Means Labels (k=K): <source>.

Using the tool#

Choose an Input Dataset, enter K clusters, and click OK. The run proceeds in the background. Click View Centroids to open any stored result and plot its centroid spectra.

Advanced Options#

Expand Advanced Options to control the fit (all optional; sensible defaults are used when blank):

  • Initialization Method — k-means++ (smart default), random, or manual (pick your own starting spectra; this hides the seed/initializations fields since the start is fixed).

  • Number of Initializations — how many times to re-run with different starts, keeping the best.

  • Max Iterations — iteration cap per run.

  • Convergence Tolerance — stop once centroids move less than this.

  • Random Seed — fix for reproducible results.

  • Algorithm — lloyd (classic) or elkan (faster on well-separated clusters, more memory).