The utilization of wood samples from xylarium in historical wooden statues: improving the separation accuracy non-destructive measurement for using several algorithms

Abstract There are numerous wooden historical artifacts in Kyoto and other parts of Japan, including Buddhist statues or Shinto deities. The identification of wood species in these historical artifact is desirable for both repair and maintenance purposes. The most common method of identifying wood species involves examining samples taken from the artifacts. However, intentional sampling from old cultural artifacts is prohibited in Japan. As a result, we attempted to determine the wood species of old statues non-destructively using near-infrared spectroscopy (NIRS). In this article, we developed the softwood and hardwood separation model using NIRS to compare the prediction accuracy for few algorithms. The model was created based on wood samples stored in the xylarium of the Forestry and Forest Products Research Institute (TWTw). We then applied this model to old Buddhist statues in order to classify them as either softwood or hardwood. These Buddhist statues were housed in Nazenji temple and are believed to have been carved during the Heian period (8th–12th century). For the near-infrared (NIR) measurements, we collected diffuse reflectance spectra from TWTw sample and Buddhist statues using same spectrometer. Initially, we used the soft independent modeling of class analogy method (SIMCA), partial least squares discriminant analysis (PLS_DA), and support vector machine to analyze the NIR spectra obtained from the TWTw wood samples. Subsequently, we applied the NIR spectra obtained from several Buddhist statues in Nazenji temple to the aforementioned separation model and determined whether spectra data were classified as the softwood or hardwood. Finally, wood specimens detached naturally from the Buddhist statues over time were observed under microscopic analysis to identify the wood species. As comparing the prediction accuracy of few algorithms, SIMCA had a poor result, but PLS_DA had a good result. PLS_DA had better discrimination because it performed calculations to improve regression from both explanatory variables and objective variables.