Abstract The identification and quantitative analysis of pigments in ancient mural paintings are essential for conservation and restoration. Hyperspectral imaging enables non-destructive pigment analysis, but long-term environmental exposure increases spectral variability and challenges conventional workflows. We propose an integrated workflow that combines region-wise pigment identification with quantitative spectral unmixing while preserving spatial continuity. Entropy Rate Superpixel Segmentation (ERS) partitions hyperspectral images into homogeneous regions for pigment identification using continuum removal (CR) and spectral correlation mapper (SCM). Quantitative unmixing is achieved through a pseudo-endmember-constrained convolutional autoencoder for unmixing (PECCAU), which jointly learns spatial–spectral representations to produce coherent abundance maps. Constraints derived from N-FINDR enhance the separation of spectrally similar pigments. The method is validated on hyperspectral images of Cave 171 in the Kizil Grottoes, identifying six principal pigments and revealing localized degradation features. Overall, ERS-guided identification mitigates spectral variability, and constrained autoencoder unmixing yields accurate and spatially consistent abundance maps.
Chen et al. (Wed,) studied this question.