Near-infrared (NIR) spectroscopy is widely used as a nondestructive evaluation (NDE) tool for predicting wood properties. When deploying NIR models, one faces challenges in ensuring representative training data, which large datasets can mitigate but often at a significant cost. Machine learning and deep learning NIR models are at an even greater disadvantage because they typically require higher sample sizes for training. In this study, NIR spectra were collected to predict the modulus of elasticity (MOE) of southern pine lumber (training set = 573 samples, testing set = 145 samples). To account for the limited size of the training data, this study employed a generative adversarial network (GAN) to generate synthetic NIR spectra. The training dataset was fed into a GAN to generate 313, 573, and 1000 synthetic spectra. The original and enhanced datasets were used to train artificial neural networks (ANNs), convolutional neural networks (CNNs), and light gradient boosting machines (LGBMs) for MOE prediction. Overall, results showed that data augmentation using GAN improved the coefficient of determination (R²) by up to 7.02% and reduced the error of predictions by up to 4.29%. ANNs and CNNs benefited more from synthetic spectra than LGBMs, which only yielded slight improvement. All models showed optimal performance when 313 synthetic spectra were added to the original training data; further additions did not improve model performance because the quality of the datapoints generated by GAN beyond a certain threshold is poor, and one of the main reasons for this can be the size of the initial training data fed into the GAN. LGBMs showed superior performances than ANNs and CNNs on both the original and enhanced training datasets, which highlights the significance of selecting an appropriate machine learning or deep learning model for NIR spectral-data analysis. The results highlighted the positive impact of GAN on the predictive performance of models utilizing NIR spectroscopy as an NDE technique and monitoring tool for wood mechanical-property evaluation. Further studies should investigate the impact of the initial size of training data, the optimal number of generated synthetic spectra, and machine learning or deep learning models that could benefit more from data augmentation using GANs.

The growing applications of near infrared (NIR) spectroscopy in wood quality control and monitoring necessitates focusing on data-driven methods to develop predictive models. Despite the advancements in analyzing NIR spectral data, literature on wood science and engineering has mainly uti- lizedthe classic model development methods, such as principal component analysis (PCA) regression or partial least squares (PLS) regression, with relatively limited studies conducted on evaluating machine learning (ML) models, and specifically, artificial neural networks (ANNs). This couldpotentially limit the performance of predictive models, specifically for some wood properties, such as tracheid width that are both time-consuming tomeasure and challenging to predict using spectral data. This study aims to enhance the prediction accuracy for tracheid width using deep neural networks and tree-based ensemble learning algorithms on a dataset consisting of 2018 samples and 692 features (NIR spectra wavelengths). Accord- ingly, NIR spectra were fed into multilayer perceptron (MLP), 1 dimensional-convolutional neural net- works (1D-CNNs), random forest, TreeNet gradient-boosting, extreme gradient-boosting (XGBoost), and light gradient-boosting machine (LGBM). It was of interest to study the performance of the models with and without applying PCA to assess how effective they would perform when analyzing NIR spectra with- out employing dimensionality reduction on data. It was shown that gradient-boosting machines outper- formed the ANNs regardless of the number of features (data dimension). Allthe models performed better without PCA. It is concluded that tree-based gradient-boosting machines could be effectively used for wood characterization utilizing a medium-sized NIR spectral dataset.

Efforts to restore longleaf pine across the southeast United States have occurred on two distinct site types: cutover forests and old agricultural fields. We measured wood and bark physical properties of unthinned planted longleaf pine from 16 stands across Georgia, ages 12 to 25, with eight stands sampled from each site type. Three-hundred and twenty trees were felled and 3572 disks collected from within the trees. Wood and bark specific gravity (SG), moisture content (MC), and proportion of bark were measured. Non-linear mixed effects models were developed to predict the variation in wood and bark SG with respect to relative height, age, and site type. Cutover sites had higher whole-tree wood SG (0.504 vs 0.455) and bark SG (0.374 vs 0.347) than old agricultural fields. The models explained 50% and 37% of the variability in wood and bark SG, respectively. Moisture content models were fitted as a function of SG for wood (R² = 0.87) and bark (R² = 0.71). Bark thickness, dry mass, and green volume were higher for cutover forest sites. Trees sampled included both defect-free and defect-containing trees; however, no significant differences in the wood physical properties were found. These results provide important information for the utilization of plantation longleaf pine.

Rubberwood (Hevea brasiliensis Muell. Arg) is underutilized in most developing countries but has potential to be a solution to the shortage of wood for pulp, paper and wood composite products at the end of its production cycle. Determining and understanding its fibre properties (fibre dimensions and morphology) is key in its utilization for pulp and paper as well as composites. This study assessed the effect of age on the fibre properties of rubberwood. Samples of wood at four ages (10, 15, 20 and 25 years) were obtained at both the axial (base and top of merchantable length) and radial direction (innerwood, middlewood and outerwood). Slivers were obtained from the samples and macerated to provide individual fibres for optical measurement. Fibre dimensions at the four ages were measured, and their morphology was determined. Fibre properties and morphology were significantly affected by age at p = 0.05. Average fibre length was 1.47 ± 0.12 mm, with non-constant variation observed across the different ages, and from pith to bark. Average lumen width was 16.87 µm, and fibre diameter ranged from 25.02 µm to 27.23 µm. The fibre properties of rubberwood potentially make it suitable for pulp and paper production as well as wood composite boards.

Near infrared (NIR) spectroscopy combined with partial least squares (PLS) regression has been widely used to predict the physical, mechanical, and chemical properties of wood. Recently, advanced modeling techniques such as deep learning and machine learning models using spectral data have increasingly been used to predict variability in wood physical and mechanical properties, but studies on chemical properties are scarce. Herein, southern pine wood samples were acquired from multiple height levels from 119 trees to yield 1007 pith-to-bark radial strips. The strips were scanned on an NIR hyperspectral imaging system and the spectral data aligned to ring-level measurements. Acetone soluble extractives content was measured on 50-mm radial length samples with 262 data points assigned to training a PLS regression model, an artificial neural network (ANN) model, and a light gradient-boosting machine learning model (LGBM). The models were then evaluated on the test dataset (140 samples). The fit statistics were good for all models, although the LGBM model (RMSE = 2.1%) performed better than the PLS regression (3.3%) or ANN (2.6%) models. The models were subsequently used to predict extractives content for 31,314 rings. The LGBM model provided biologically realistic values for all rings, whereas the PLS regression model predicted 9648 negative values, and the ANN model extrapolated well beyond the measured values (max predicted = 69.4% vs max measured = 53.2%). This study highlights real benefits of using deep learning and machine learning models compared to PLS regression models for NIR hyperspectral imaging datasets.