Classification of thermally treated wood using machine learning techniques

Classification of thermally modified wood is a critical assessment and control task that assures the quality of thermally treated wood. Machine learning methods can be used for identifying the optimal feature(s) for wood classification. In this study, the performance of artificial neural networks (ANN), support vector machines (SVM), and naïve Bayes (NB) classifiers for thermowood classification was evaluated and compared. The moisture content, water absorption, swelling coefficient, color, hardness, and dynamic modulus of elasticity of untreated and thermally treated western hemlock wood were measured and analyzed to identify the optimal set(s) of feature(s) for wood classification. The results showed that mechanical attributes such as dynamic modulus of elasticity obtained from the stress wave timer test and wood hardness account for the least suitable features, whereas color measurement provided an accurate classification. Both SVM and naïve Bayes model showed significantly higher performance than ANN because the latter requires a higher number of tuned and optimized parameters. Having only one feature, the accuracy of SVM and naïve Bayes model obtained from the color lightness parameter (L*) was 0.960 and 0.949, respectively. By increasing the dimension of the features, naïve Bayes model outperformed SVM and resulted in a robust classifier with an accuracy of 0.990. A trade-off between increasing the model accuracy and minimizing the number of selected features was observed. The SVM and NB models showed promising performance for the classification of thermally modified wood, which could be implemented for in-line quality control.