Land cover without training data: a physics-based pipeline
Most land-cover products are supervised classifiers: they need labelled training data, encode brittle correlations, and don't travel well between regions. There is another way that removes the label bottleneck entirely.
Two independent layers
The first layer clusters each pixel by spectral material identity — concrete, asphalt, metal roof, water, soil, vegetation — in an illumination-decoupled colour space, with cloud and shadow falling out as their own classes for free. The second characterises each pixel's spatial context: how diverse, how textured, and how large the surrounding patch is.
Built-up, roads, agriculture, solar farms, quarries and isolated structures then emerge as queries against that representation, not as separately-trained models. Default semantic labels come from OpenStreetMap as a majority vote — never as ground truth — so the map stays current and high-confidence disagreements surface as candidate change events.
Why training-free travels
Because the method encodes material physics and the structural signature of human-built environments — both invariant across geography — it runs anywhere with no local calibration: on full Sentinel-2 multispectral, on RGB basemaps, or on enhanced derived-resolution imagery. The same representation is the foundation beneath construction detection, corridor monitoring and agricultural products.