Application of Uniform Manifold Approximation and Projection (UMAP) in spectral imaging of artworks

This study assesses the potential of Uniform Manifold Approximation and Projection (UMAP) as an alternative tool to t-distributed Stochastic Neighbor Embedding (t-SNE) for the reduction and visualization of visible spectral images of works of art. We investigate the influence of UMAP parameters—such as, correlation distance, minimum embedding distance, as well as number of embedding neighbors— on the reduction and visualization of spectral images collected from Poèmes Barbares (1896), a major work by the French artist Paul Gauguin in the collection of the Harvard Art Museums. The use of a cosine distance metric and number of neighbors equal to 10 preserves both the local and global structure of the Gauguin dataset in a reduced two-dimensional embedding space thus yielding simple and clear groupings of the pigments used by the artist. The centroids of these groups were identified by locating the densest regions within the UMAP embedding through a 2D histogram peak finding algorithm. These centroids were subsequently fit to the dataset by non-negative least square thus forming maps of pigments distributed across the work of art studied. All findings were correlated to macro XRF imaging analyses carried out on the same painting. The described procedure for reduction and visualization of spectral images of a work of art is quick, easy to implement, and the software is opensource thus promising an improved strategy for interrogating reflectance images from complex works of art.