Clusterization is unsupervised classification — and the simplest form of machine learning in the platform: the algorithm learns the natural groupings in your data on its own and sorts every pixel into a class by spectral similarity, with no training data and nothing to label. It’s the fastest way to turn a continuous insight into a map of discrete units: alteration domains, lithological groupings, land-cover classes. (When you do have ground truth to teach from, that’s supervised learning — see Trained AI Models.)
Running a clusterization #
In Cloud → Exports, click the layers icon (Clusterize Image) on a file. It runs on files from your own saved Areas up to 200 km² and 5 million pixels (the tooltip explains any blocker). Parameters:
- Number of clusters — 2 to 9 (default 6). Start around 5–6: too few merges real units, too many fragments them.
- Method — K-Means (fast, compact clusters — the default), MiniBatchKMeans (faster on large files), or Gaussian Mixture Model (better when populations overlap).
- Clean clusters (on by default) — removes speckle noise and smooths the class map. Keep it on unless you specifically want the raw labels.
- Keep raw output — also saves the pre-cleaning raster, for comparison.
- Include preview image — adds a quick-look PNG.
What you get #
The result is a new classified GeoTIFF saved right next to the source file in your Cloud — it appears as a processed file attached to the original insight, and you can put it on the Globe with Visualize on Map like any other file.
The results view gives you the numbers behind the map:
- Cluster Centers — each class’s center value, pixel count, mean and spread, straight from the algorithm.
- Class Statistics — the cleaned totals: pixels, area per class (in your preferred unit) and ratios. The two tables can differ slightly — one is before cleanup, the other after.
- The Outputs list carries download links for every artifact. The links are temporary and show a live countdown — if one expires, re-run to regenerate.
- Export produces a PDF report or CSV of everything.
Vectorization: from classes to polygons #
On the clean clusters output, Download vector converts the class raster into polygons — delivered as a Shapefile ZIP (.shp, .shx, .dbf, .prj, .cpg) that includes a QGIS style file (.qml) with the classes already colored. Each polygon carries its class id and area in m². Open it in QGIS or any GIS and it’s ready to work with: measure class areas, intersect with tenements, plan ground follow-up.
Adjacent polygons of the same class are merged, and tiny slivers are dropped, so the output is clean enough to use directly.
When to reach for it #
- Domaining an alteration index — turn a continuous ratio into 5–6 mappable units before ground-truthing.
- Reducing a PCA export — cluster the components into a pseudo-geology map.
- Preparing vector deliverables — clusterize, vectorize, and hand the shapefile to the GIS team in one sitting.