Researchers build AI that unmixes hyperspectral images without knowing the mixing rules

A new machine learning approach can separate out the individual materials in a satellite image without being told how they were mixed together in the first place. This matters because current methods require scientists to guess the mixing model beforehand, which fails when the mixing isn't linear or when the mixing model is unknown.

Hyperspectral imaging is used to identify materials in satellite photos — minerals, crops, contaminants, vegetation health. Every current method assumes you know the mathematical model of how light from different materials gets combined in a single pixel. This paper's approach learns the unmixing process from data instead, which could work on real satellite imagery where the mixing is messier and more complex than the math assumes. If this generalizes, it removes a major constraint on what remote sensing analysis can do.