Machine learning tools for the classification of visible reflectance spectra of inorganic pigments historically used in wall paintings

Three different classification methods applied to the field of Cultural Heritage are compared. Focusing on the non-invasive and in situ identification of historical inorganic pigments using visible reflectance spectroscopy, linear and non-linear classifiers are trained and evaluated (Partial Least Squares - Discriminant Analysis (PLS-DA), Support Vector Machine (SVM-DA), and Artificial Neural Networks (ANN-DA)). Three hues (yellow, blue and green) are considered with three pigments per hue: lead-tin yellow, yellow ochre, orpiment (yellow); celadonite, green malachite, Verona green (green); and Egyptian blue, azurite, and lapis lazuli (blue). All three classification methods yielded high prediction capabilities (above 94.2 % in the external validation sets), with high sensitivity and specificity (above 88 %). In cases where spectral features are more similar, PLS-DA performed slightly worse. However, SVM-DA and ANN-DA models were comparable in all instances, with global accuracies above 98.8 %, sensitivity above 88 %, and specificity above 98 %, respectively.