High-throughput discrimination of ten poplar (Populus spp.) wood cultivars using near-infrared spectroscopy coupled with convolutional neural networks

Poplar ( Populus spp.) is an important industrial crop, providing sources of fuelwood, raw material for the pulp and paper industry, and wood-based panels. Advances in genetic breeding have led to a diverse array of poplar cultivars, which exhibit significant differences in their suitability for various industrial and ecological applications. To date, wood identification has typically been limited to the taxonomic level of "family" or "genera", with limited literature on precise discrimination of poplar cultivars. This limitation is primarily due to low genetic and morphological variability among poplar species, which poses challenges for accurate classification. This study presents a high-throughput and non-destructive technique for discriminating poplar cultivars by integrating convolutional neural networks (CNNs) with near-infrared (NIR) spectroscopy. Different cultivars of poplar wood were used to gather NIR spectroscopy data in the 350–2500 nm region, and several preprocessing techniques were evaluated methodically. Based on the preprocessed spectra, three convolutional neural network models, NIR-Net, NIR-ResNet, and NIR-Inception, were constructed and trained for variety classification, and the performance of the models was evaluated on an independent validation set. The results showed that the combination of Savitzky-Golay derivatives (SGD) and multivariate scatter correction (MSC) significantly improves the classification accuracy of the NIR-ResNet model: achieving an accuracy of 97.52 % for distinguishing six conventional hybrids cultivars and 98.18 % for discriminating five closely related 'Lulin' hybrids cultivars. This method not only significantly improves classification accuracy of poplar cultivars but also lays a foundation for the industrial and ecological application of specific poplar cultivars. • Accurate discrimination of ten poplar ( Populus spp.) wood cultivars by near infrared spectroscopy. • Convolutional neural networks were employed to significantly enhance the prediction accuracy of poplar cultivars. • Develop high-throughput identification techniques for poplar cultivars suitable for industrial applications.