Land use / land cover (LULC) classification answers one question: what's actually covering the ground in this image, pixel by pixel? Done by eye, it's slow and subjective — two people looking at the same scene will draw the boundaries differently. Done with plain unsupervised clustering, it's fast but incomplete: a clustering algorithm will happily group pixels by spectral similarity, but it has no idea which cluster is "water" and which is "vegetation." Someone still has to look at cluster #3 and decide what it actually represents.
The part clustering alone doesn't do
K-means and similar unsupervised methods are genuinely good at their actual job: grouping pixels that look spectrally alike. What they don't do is attach real-world meaning to a group. "Cluster 3" and "Cluster 7" mean nothing on their own — someone has to inspect each one and manually assign a label, and that step is where consistency (and a lot of time) usually gets lost, especially across a large scene with dozens of clusters.
Automatic baseline, or add your own samples
The automatic path is genuinely unsupervised — no training data required, ready as soon as your imagery is uploaded. If a class needs strengthening, or you need one the automatic baseline didn't capture, you can add your own samples on top of it. Both the automatic and the sample-based classes use the exact same underlying spectral signature method, so refining doesn't mean switching to a different system — it's the same approach, with more input.
What imagery you need
Red, Green, and Near-Infrared (NIR) bands at minimum — that's what the NDVI and NDWI calculations behind the automatic labeling actually rely on. A SWIR band, common in Landsat-style 6-band imagery, adds a real built-up/bare-soil distinction via NDBI; without it, that specific distinction falls back to a coarser brightness-based estimate rather than disappearing entirely.
It doesn't need to already be one combined file, either. A real USGS EarthExplorer Landsat download, or a Copernicus Sentinel-2 download, typically arrives as separate single-band files — upload them separately and pick your imagery source (Landsat 4/5, Landsat 7, Landsat 8/9, or Sentinel-2) for correctly-labeled upload slots, since Landsat's older and newer sensors genuinely use different band numbers for the same colors. No detour through QGIS or GDAL to combine them first.
How to do it
-
Upload your imagery
One pre-stacked multi-band GeoTIFF works, or separate single-band files (a real Landsat/Sentinel-2 download) — pick your imagery source for correctly-labeled upload slots, or assign bands manually for anything else.
-
Choose automatic or refine
Run the automatic baseline on its own, or add your own samples to strengthen a class or define a new one — both use the same underlying method.
-
Build your classes
If refining, click a handful of sample points to start a class, then quickly confirm or reject a few spatially-spread candidate pixels per round — optional after the first pass.
-
Export your outputs
Choose any combination of a classified raster, labeled vector polygons, and a branded PDF report with category-wise area — all in one download.
Ready to classify your own imagery?
Open Hybrid LULC Classification →