Development of a discriminant model based on visible – near-infrared hyperspectral imaging data for evaluating the distribution of fracture modes on the surfaces of wooden block-shear specimens

In this study, we developed an objective and automated discriminant model based on visible – near-infrared hyperspectral imaging data for evaluating the fracture modes on the surfaces of wooden block-shear specimens bonded by a polyvinyl acetate emulsion adhesive. The hyperspectral reflectance data of the wood and adhesive-applied surfaces were used as the training data. The wood and adhesive discriminant model was constructed using partial least squares discriminant analysis, and its overall accuracy score was shown to be over 0.92. In the discriminant images of the fracture surface, the distribution of adhesive was consistent with that observed by visual inspection when the fracture surface was flat. However, large wood failures with surface concavity tended to be misassigned as adhesive surfaces. In the paired fracture surface images, three fracture modes (wood failure, adhesive failure, and cohesive failure) were visualized. The wood failure percentages calculated from the images were lower than those determined by visual inspection. In conclusion, although the developed model is not capable of assessing large wood failures, it may nonetheless be suitable for assessing fracture modes at a high resolution that cannot be detected by visual inspection.