Joined Spatial and Spectral Segmentation of Hyperspectral Datasets on Historical Art Objects

In the context of clustering and classification, the choice between spatial and spectral features hinges on data characteristics and analytical goals. Spatial features excel in spatially contextualised data like urban planning and object detection, capturing adjacent pixel relationships. Spectral features, prominent in hyperspectral data, shine in tasks requiring precise material identification. However, managing extensive hyperspectral data poses challenges, prompting hybrid approaches for efficient data segmentation. This study presents an innovative data processing pipeline that combines spatial and spectral clustering techniques for the extraction and mapping of spectral signatures in both cinematic films and pointillism painting. We explore Simple Linear Iterative Clustering (SLIC) Superpixel’s potential in reducing hyperspectral data while maintaining spectral richness. By optimizing superpixel segmentation and extracting central spectra, we reduce data size significantly. This facilitates applying various machine learning algorithms, especially the soft clustering algorithms Fuzzy C-Means clustering (FCM) and Gaussian Mixture Models (GMM), to classify spectra into distinct colourant groups, identifying complex mixtures and areas of degradation. This research highlights the potential of machine learning in aiding artwork diagnostics, conservation, and restoration, with transferable models for similar scenarios.