Spectral Band Attribution in Historical Ink and Substrate Recognition

We present an interpretative approach to classifying inks and substrates in historical handwritten documents using hyperspectral imaging and deep learning. We introduce a custom convolutional neural network specifically tailored for high-dimensional spectral data and the low cardinality of the available training set. Additionally, we address typical challenges encountered in hyperspectral analysis, including data imbalance through class balancing techniques, and interpretability via spectral band attribution analysis using Integrated Gradients. Our findings demonstrate improved interpretability and offer practical insights - for instance, the association of Sepia and Pencil with SWIR bands, and the sensitivity of Cotton and Hemp to distinct spectral regions - that can guide optimized imaging protocols and inform preservation strategies. This framework advances the non-invasive analysis of historical documents, supporting both accurate classification and interpretation for heritage conservation.