Kubelka-Munk theory enhances hyperspectral mineral identification for rock heritage.

Accurate identification of minerals is the fundamental task in the protection of rock cultural relics, directly determining the scientific validity and effectiveness of research on weathering mechanisms, disease prediction, and the development of protective materials. However, remote sensing technology is constrained by spatial and observational conditions, making fine in‑situ surveys difficult to achieve. In contrast, proximal hyperspectral imaging overcomes these limitations, providing a reliable technical foundation and data source for in‑situ mineral detection. In this study, a hyperspectral database of common minerals spanning the 900-1700 nm band was established. The diagnostic spectral features of minerals were enhanced through the integration of Kubelka-Munk theory, while efficient mineral identification was achieved by applying machine learning models. Experimental results indicate that diagnostic spectral characteristics are most pronounced within the 1360-1490 nm band. Performance evaluation shows that average efficiency increased by 12.55% and 28.08%, mean F1 scores improved by 12% and 6.14%, mean Kappa values increased by 10.86% and 4.43%, and the F1score of SVM improved by 30%. In addition, spectral detection was conducted on sandstone samples from the Dazu stone carvings, and the experimental results confirmed the accuracy and reliability of the proposed method.