Abstract:

Objective: The feasibility of accurately identifying wood species by integrating near infrared spectroscopy (NIR) and digital image feature information, automatically extracted based on convolutional neural networks (CNN), is investigated. Method: 10 wood specimens from the Lauraceae family were used as examples. NIR spectra and digital images from the transverse surfaces of specimens were collected using a handheld NIR spectrometer and portable scanner. The recurrence plot (RP) method was innovatively employed to encode one-dimensional (1D) NIR spectra as two-dimensional (2D) images. This not only enabled the CNN model to extract more robust discriminative features from the handheld NIR spectrometer’s shorter-wavelength NIR spectra, but also facilitated the fusion of NIR spectra and images. A simple two branch CNN (TB-CNN) model was developed to automatically extract and fuse NIR spectral and digital image features for wood species identification. Result: Compared to different modeling methods that utilize 1D NIR spectra directly, integrating 2D RP of NIR spectra with CNN resulted in a performance improvement ranging from 1.79% to 14%. Furthermore, when compared with the results obtained using a single feature from either NIR spectra or digital image, the TB-CNN method showed a significant increase in identification rates of at least 3%. Notably, the accuracy, precision and recall values all surpassed 99%. Conclusion: The conversion of 1D short-wavelength NIR spectra into RP enhances the ability of CNN model to extract more discriminative features from the NIR spectral data, improving model performance in species identification. The TB-CNN model effectively extracts and integrates wood NIR spectra and digital image features, addressing the limitations of using a single feature type for wood species identification and improving overall identification accuracy.

Key words: wood species identification, convolutional neural network (CNN), near-infrared spectroscopy, image, feature extraction and fusion