Evaluation of test specimen surface preparation on macroscopic computer vision wood identification

Previous studies on computer vision wood identification (CVWID) have assumed or implied that the quality of sanding or knifing preparation of the transverse surface of wood specimens could influence model performance, but its impact is unknown and largely unexplored. This study investigates how variations in surface preparation quality of test specimens could affect the predictive accuracy of a previously published 24-class XyloTron CVWID model for Peruvian timbers. The model was trained on images of Peruvian wood specimens prepared at 1500 sanding grit and tested on images of independent specimens (not used in training) prepared across a series of progressively coarser sanding grits (1500, 800, 600, 400, 240, 180, and 80) and high-quality knife cuts. The results show that while there was a drop in performance at the lowest sanding grit of 80, most of the higher grits and knife cuts did not exhibit statistically significant differences in predictive accuracy. These results lay the groundwork for a future larger-scale investigation into how the quality of surface preparation in both training and testing data will impact CVWID model accuracy.