Hyperspectral identification of mineral pigments in Thangka paintings for cultural heritage conservation

Accurate identification of pigments is vital for preserving Tibetan Thangka paintings, where mineral colors often occur in undocumented mixtures and layered structures. We introduce an automatic hyperspectral framework that unifies pigment classification and mixing-ratio estimation without prior knowledge. A standardized library of 12 pure pigments and 54 binary mixtures was built, and a multi-stage strategy coupling classification with conditional regression was implemented. On laboratory data, the framework achieved near-perfect classification of pure pigments and reliable mixture estimates (R2 = 0.98). Applied to a hand-painted Thangka, it reached 75% accuracy—comparable to spectral-angle mapping (SAM)—while additionally providing quantitative ratios and better scalability than conventional matching/unmixing. Although validation is limited to one artwork and binary mixtures, this proof-of-concept shows that hyperspectral imaging with machine learning can deliver interpretable, practical tools for pigment analysis in cultural-heritage conservation.