Utilization of Synthetic Near-Infrared Spectra via Generative Adversarial Network to Improve Wood Stiffness...

Area of Science:

• Materials Science
• Spectroscopy
• Machine Learning

Background:

• Near-infrared (NIR) spectroscopy is a key nondestructive evaluation (NDE) tool for predicting wood properties.
• Developing accurate NIR models is challenged by the need for large, representative training datasets, which are often costly to acquire.
• Machine learning (ML) and deep learning (DL) models, particularly for NIR data, require substantial sample sizes for effective training.

Purpose of the Study:

• To investigate the use of generative adversarial networks (GANs) for augmenting limited near-infrared (NIR) spectral data.
• To evaluate the impact of GAN-generated synthetic spectra on the predictive performance of artificial neural networks (ANNs), convolutional neural networks (CNNs), and light gradient boosting machines (LGBMs) for wood modulus of elasticity (MOE).
• To determine the optimal number of synthetic spectra for improving MOE prediction accuracy.

Main Methods:

• Collected NIR spectra from southern pine lumber (training set: 573 samples, testing set: 145 samples).
• Employed a GAN to generate synthetic NIR spectra (313, 573, and 1000 additional spectra).
• Trained ANNs, CNNs, and LGBMs using original and augmented datasets to predict MOE.

Main Results:

• Data augmentation with GANs improved the coefficient of determination (R²) by up to 7.02% and reduced prediction error by up to 4.29%.
• ANNs and CNNs showed greater benefits from synthetic spectra compared to LGBMs.
• Optimal performance was achieved with the addition of 313 synthetic spectra; larger additions did not further improve results due to declining synthetic data quality.

Conclusions:

• GAN-based data augmentation positively impacts the predictive performance of NIR spectroscopy models for wood mechanical properties.
• LGBMs demonstrated superior performance over ANNs and CNNs, highlighting the importance of model selection in NIR spectral data analysis.
• Further research is needed on the influence of initial training data size and the optimal quantity of synthetic spectra for GAN-based augmentation.