Identification of softwood species using convolutional neural networks and raw near-infrared spectroscopy

Abstract

ABSTRACT Previous reports have shown that wood species identification result based on near-infrared (NIR) spectroscopy was intimately entwined with spectra preprocessing. However, there is no universal recipes for a suitable preprocessing method, and misuse of preprocessing may bring on worse model performance for new species identification. Therefore, a convolutional neural network (CNN) model incorporating a residual connection structure is created aiming at replacing the preprocessing and identifying 21 Pinaceae species at the species level. The model is compared to the other two CNN models on different wavelength range raw transverse section NIR spectra. 12 preprocessing methods are carried out for 780–2440 nm spectra to evaluate the influence of spectra preprocessing on the model. The model outperforms the other two CNN models on raw and preprocessed spectra and provides the highest macro F1 of 0.9787 and 0.9792 for raw and preprocessed spectra at the wavelength range of 780–2440 nm. The model is further compared to three conventional methods. The results indicated that created model is capable to replace the spectra preprocessing and identify 21 wood species at the species level. It is indicated that a suitable CNN structure can replace the multifarious data preprocessing in traditional methods. It potentially provides a generic raw NIR spectra discrimination method for wood species identification.