Predicting the strength of European beech (Fagus sylvatica L.) boards using image-based local fibre direction data

Abstract

Image-based local fibre direction data, generated based on the analysis of the medullary spindle pattern, was used to improve the prediction of the tensile strength parallel to the grain of European beech (Fagus sylvatica L.) boards. An approach to characterise the local fibre orientations in a board using a single numerical grading parameter was further developed. This parameter was used, in combination with the dynamic modulus of elasticity, to develop a regression model providing very good predictions of the experimentally determined tensile strength parallel to the grain (R 2 = 0.84). Subsequently, machinelearning techniques were used to improve the strength model. Non-destructive and destructive tests were performed on (N =) 47 boards. A data (sub-) set (n = 36) was used to train different machine learning techniques (Support-Vector Machines, Decision Tree, Random Forest, and Artificial Neural Network) using a 6-k cross-validation approach. The generalisation ability of the models was then assessed by a holdout dataset (n = 11). The results showed that all machine-learning models presented good prediction accuracy (R 2 up to 0.88 and MAPE below 8%). The Support-Vector machine and Random Forest methods showed the best performance. The combination of experimental methods with machine learning allows for a more precise strength grading of timber and, thus, can contribute to a more resource-efficient use of wood and may open new and more demanding fields for high-level timber applications in structures.

Probing the complexity of wood with computer vision: from pixels to properties

“…The high accuracy of our models suggests that a recording and subsequent processing of data of surface patterns of the wood surface can sufficiently display the relationship between hierarchical wood structure and selective mechanical properties. This is in line with the emergence property of fibre-based composites discussed by Jeronimidis & Vincent [18] and the results obtained by Ehrhart et al in a similar study [16]. They used image analysis with predefined indicators of wood fibre directions to infer the strength of beech lamellae.…”

“…Therefore, probing the complexity of wood is indispensable for understanding better the relationship between structure and properties. Indeed, a recent study showed that local fibre direction data, generated based on the analysis of the spindle pattern of bigger rays on tangential surfaces, could improve the prediction of the tensile strength parallel to the grain of European beech (Fagus sylvatica) lamellae [16]. In addition, Olsson et al developed a novel strength grading method, approved for the European market, based on fibre orientation and dynamic excitation of Norway spruce lamellae [12,20].…”

Sustainability in Wood Products: A New Perspective for Handling Natural Diversity

“…It is of no surprise that such linear models perform poorly, placing the coefficient of determination somewhere between 0.45 and 0.7 . From a scientific perspective, so far there have been very few attempts to use ML to improve the machine grading of wood. − The major part of research in the field of machine grading is dedicated to finding new measurement techniques and better indicating properties in order to improve the accuracy of the prediction …”

“…The introduction of ML into the manufacturing process of engineered wood products would provide the opportunity to predict in real time the performance characteristics of the final product, to carry out production according to certain specifications, and to reduce part of the raw material consumption and/or increase the overall throughput. ,, While traditional ML methods based on supervised learning (labeled data) have been used with some success to predict the quality of wood products, ,,− the limited availability of labeled data is the major hurdle for further improving ML model performance.…”

Semi-supervised learning for quality control of high-value wood products

“…Traditional ML methods based on supervised learning have been used with some success to predict the quality of wood products (Barnes 2001;Gupta et al 2007;André et al 2008;Esteban et al 2011;Bardak et al 2016a, b;Schubert and Kläusler 2020;van Blokland et al 2021;Ehrhart et al 2022;Rahimi and Avramidis 2022). For example, artificial neural networks (ANN), support vector machines (SVM), and Naive Bayes (NB) models were used to classify the quality of thermally modified wood (Nasir et al 2019).…”