Animal Species Identification in Historical Parchments by Continuous Wavelet Transform– Convolutional Neural Network Classifier Applied to Ultraviolet–Visible–Near-Infrared Spectroscopic Data

Identification of animal species in medieval parchment manuscripts is highly relevant in cultural heritage studies. Usually, species identification is performed with slightly invasive methods. In this study, we propose a contactless methodology based on reflectance spectrophotometry (ultraviolet–visible–near-infrared) and a machine learning approach for data analysis. Spectra were recorded from both historical and modern parchments crafted from calf, goat, and sheep skins. First, a continuous wavelet transform was performed on the spectral data as a preprocessing step. Then, a semisupervised neural network with a 2-component architecture was applied to the preprocessed data. The network architecture chosen was CWT-CNN (continuous wavelet transform–convolutional neural network), which, in this case, is composed of a convolutional autoencoder and a single-layer dense network classifier. Species classification on holdout historical parchments was attained with a mean accuracy of 79%. The analysis of Shapley additive explanations values highlighted the main spectral ranges responsible for species discrimination. Our study shows that the animal species signature is encoded in a wide band-convoluted wavelength range rather than in specific narrow bands, implying a complex phenotype expression that influences the light scattering by the material. Indeed, the overall skin composition, in both micro-and macroscopic physicochemical properties, is relevant for animal identification in parchment manuscripts.