Wood you believe it-Detecting copper with hyperspectral imaging

Enhancing circularity and promoting the cascade use of wood resources, currently considered non-recyclable despite containing considerable amounts of recyclable material, requires the ability to identify and manage inorganic wood preservatives, with particular focus on copper. In this study, hyperspectral imaging (VNIR and SWIR) combined with artificial intelligence was used to detect and assess copper concentration ranges in treated wood. Therefore, a representative dataset of 1320 samples was constructed using six wood species. This dataset included untreated samples and samples treated with three inorganic copper-based and two organic preservatives, each applied at multiple concentrations to ensure sufficient diversity for robust model development. Three model types were evaluated: a support vector classifier (SVC), partial least squares discriminant analysis (PLS-DA), and a convolutional neural network (ResNet18). While PLS-DA was included as a baseline, its linear nature proved inadequate for capturing the strong non-linearity present in hyperspectral signals, resulting in poor classification performance. In contrast, the SVC and ResNet18 models achieved accuracies above 90%, highlighting the advantage of non-linear approaches for capturing complex spectral patterns. As part of the study's exploratory scope, Out-of-Distribution (OOD) Generalization tests using the ResNet18 model were carried out to explore species- and treatment-specific effects on spectral response and model behavior. In these tests, specific wood species or preservative types were excluded from training and used only for evaluation. The results highlighted that model performance depends strongly on the composition of the training dataset, emphasizing the need for broad coverage across species and preservative types to support robust generalization.