Abstract
Murals constitute an invaluable component of cultural heritage, encapsulating profound artistic and historical significance. Hyperspectral remote sensing, as a non-destructive testing technique, offers an effective means for analysing and identifying mural pigments. We proposed a spectral unmixing method based on the negative logarithm of spectral reflectance, aimed at improving the accuracy of quantitative analysis for composite pigments in murals. We obtained three sets of mixed spectra: cinnabar-orpiment, azurite-orpiment, and azurite-malachite. The original reflectance was converted into negative logarithmic values, and linear models were applied for unmixing. Supervised unmixing was conducted in the experiment to evaluate the model’s precision, with the unmixing process carried out using known endmembers. The unmixing accuracy was quantified using the Root Mean Square Error, which compares the estimated abundances of all endmembers with actual values. Additionally, a comparison was made between the negative logarithmic transformation and the commonly used transformations of original reflectance, first derivative, and second derivative. The result show that the logarithmic hyperspace linear model simplifies spectral unmixing complexity, enhancing the accuracy of linear unmixing models in addressing mixed pigment unmixing problems. To validate the effectiveness of the proposed method for real murals, the improved model was applied to the hyperspectral data unmixing analysis of murals collected by the Yungang Academy, located in Shanxi Province, China. Experimental findings demonstrate that, compared to traditional methods, the improved unmixing approach more precisely represents the pigment distribution in the murals.
Introduction
Ancient murals are complex layers of history, but they are physically fragile. To restore them, conservators need to know exactly which pigments were used and in what proportions. However, painters rarely used just pure colors; they mixed mineral pigments to create new hues. While hyperspectral imaging (HSI) allows for non-destructive analysis, "unmixing" these colors—determining the percentage of each pure pigment in a single pixel—is difficult. Most standard tools use a Linear Mixture Model (LMM), which assumes pigments sit side-by-side. In reality, pigments often mix "subtractively," creating non-linear spectral signatures that confuse standard linear models. This research seeks a mathematical bridge to make these complex mixtures compatible with efficient linear unmixing tools.
Materials and Methods
The study utilized two primary data sources for model development and validation. First, a laboratory-controlled dataset was created using 21 mixed pigment samples across three pairs: Cinnabar-Orpiment, Azurite-Orpiment, and Azurite-Malachite. These were mixed in seven specific mass ratios ranging from 20:80 to 80:20 and measured with an ASD-FieldSpec4HI-RES spectrometer (350–2500nm). Second, a real-world validation dataset was obtained from hyperspectral images of a Qing Dynasty mural (1636–1912 AD) at the Yungang Academy in Shanxi, China.
The proposed method centers on a Negative Logarithmic (NL) transformation combined with Fully Constrained Least Squares (FCLS) unmixing. Mathematically, taking the negative log of a product (representing subtractive mixing) converts it into a sum (additive mixing), which aligns with the Linear Mixture Model. The team compared four data processing states: Original Reflectance, First-Order Derivative (FOD), Second-Order Derivative (SOD), and the NL transformation.
Results and Discussion
The accuracy of the models was measured using Root Mean Square Error (RMSE), where a lower value indicates higher precision. The NL-FCLS model demonstrated superior performance for most mixtures:
- Azurite and Malachite: The NL-FCLS model achieved an RMSE of 0.0650, significantly outperforming the standard FCLS (0.1103).
- Azurite and Orpiment: The NL-FCLS model led with an RMSE of 0.1369.
- Cinnabar and Orpiment: Interestingly, the FOD-FCLS model performed best for this specific pair with an RMSE of 0.0976, compared to 0.4112 for the standard model. This suggests that for certain pigments, the rate of change in reflectance (gradient) is a stronger identifier than absolute intensity.
When applied to the Qing Dynasty mural, the NL-FCLS model produced highly detailed abundance maps, successfully distinguishing between different blue and green mineral applications that appeared uniform in standard photography. This "linearization" of the physics of light reduces computational complexity without requiring heavy non-linear machine learning models.
Conclusion
The negative logarithmic hyperspace linear model offers a significant advancement in the digital preservation and restoration of mural composite pigments. While the linear-logarithmic approach is a major step forward, future work will address limitations such as the influence of complex binders (e.g., animal glue) and environmental contaminants. Researchers plan to further explore non-linear mixing mechanisms and algorithm fusion to create more robust, automated identification systems for cultural heritage conservation.
Summary updated