Non-destructive identification of mineral pigments in painted cultural heritage based on hyperspectral imaging...

Accurate identification of mineral pigments is fundamental to the conservation and digital restoration of painted cultural heritage. However, traditional survey methods suffer from limitations such as strong subjectivity, destructive physical sampling, and difficulty in achieving macroscopic global mapping. Taking the reclining Buddha at the Baoding Mountain Rock Carvings in Dazu as a representative case, an intelligent hyperspectral-imaging-based framework is proposed for mineral pigment identification and spatial distribution visualization. The main pigments found are terracotta, ultramarine, paris green, and carbon black. Savitzky-Golay smoothing and Min-Max normalization were applied to the spectral data in the 400-1000 nm range, which effectively eliminated the influence of random noise. Competitive adaptive reweighted sampling and continuous projection algorithms were used to screen feature wavelengths, successfully reducing collinearity redundancy in high-dimensional data. After constructing and evaluating four classification models (RF, SVM, KNN, and LightGBM), among the four classical classifiers evaluated, SVM demonstrated the best noise resistance and robustness, and was able to effectively identify typical pigments such as terracotta, ultramarine, Paris green, and carbon black, while effectively removing interference from complex bedrock backgrounds.Based on the established mapping benchmarks, a large-scale, time-series dynamic monitoring framework suitable for preventive conservation can be further developed. This study provides an efficient and objective technical approach for the accurate identification and data-driven protection of cultural heritage in high-humidity environments.