Fig. 2

Each artist (artists 1–4) created three paintings, one of which was reserved for testing the trained ML algorithm. These test paintings are shown for all four artists as A high-resolution photographs (all paintings used in the study are presented in Additional file 1: Fig. S1), B height data, shaded in grayscale from low (darker) to high (lighter). C Attribution results of the ensemble ML predictions on height data. Color shadings are overlayed on the grayscale image from row B corresponding to the most likely artist for each patch of side-length 200 pixels (10 mm; 1: red, 2: orange, 3: green, 4: blue). More opaque shades of color indicate greater predictive confidence (larger attribution probabilities). The overall accuracy of the patch attribution is 96.1%