Spectral Unmixing Algorithm: AI Core Brain Enabling Precise Material Identification for Multispectral Sensing

Spectra collected by multispectral sensors aremixed spectrasuperposed by target materials, ambient light, reflection, shadows and noise, leading to massive errors in direct identification.Spectral Unmixingextracts standard spectral fingerprints of pure target materials by stripping interference signals via AI algorithms, the “software brain” enabling material identification for multispectral hardware.

The mainstream commercial solution adopts theLinear Spectral Unmixing (LSU)model with three core steps. First,Endmember Extraction: a standard material spectral library is built with pure spectra of plastics, metals, skin, blood, road surfaces and vegetation recorded. Second,linear unmixing calculationdecomposes mixed spectra into proportions of each standard material. Third,noise and outlier filteringeliminates interference from ambient light, reflection and dynamics to output precise material classification results. Algorithm optimization directions vary by scenario: industrial inspection focuses onprecise extraction of defect endmembersto identify tiny corrosion and impurities; consumer electronics emphasizesbiological spectral unmixingto distinguish skin, blood and living tissues for anti‑spoofing and health detection; automotive perception prioritizesreal‑time dynamic spectral unmixingadapting to dynamic environments of rain, fog, intense light and backlighting. Currently, Sony, Imec and Tesla are equipped with self‑developed spectral unmixing algorithms. Hardware‑algorithm synergy delivers a complete solution for large‑scale multispectral commercialization.