Near infrared spectroscopy for wood identification based on wavelet convolutional networks

Area of Science:

• Spectroscopy
• Machine Learning
• Materials Science

Background:

• Near-infrared spectroscopy (NIRS) is promising for wood identification.
• Data-driven models struggle with spectral noise and subtle chemical signatures, leading to overfitting and poor generalization.
• Robust wood identification requires methods that can extract meaningful information from noisy spectral data.

Purpose of the Study:

• To develop a robust and accurate wood species identification method using NIRS.
• To overcome limitations of purely data-driven models in extracting subtle spectral features.
• To improve generalization and anomaly rejection capabilities in spectral analysis.

Main Methods:

• Proposed a wavelet-enhanced physics-informed neural network (Wavelet-PINN).
• Utilized discrete wavelet transform (DWT) for multiscale spectral signal decomposition.
• Incorporated a physics-aware gating mechanism with Sobolev gradient losses for attention mask generation.
• Employed an autoencoder (AE) for unsupervised signal reconstruction and regularization.

Main Results:

• Wavelet-PINN achieved superior classification performance compared to state-of-the-art deep learning methods.
• Attained a classification accuracy of 98.81%, outperforming Transformer-based and polynomial network baselines.
• The reconstruction branch provided an effective physical threshold for anomaly detection.

Conclusions:

• Wavelet-PINN offers a robust solution for wood species identification using NIRS.
• The physics-informed approach effectively isolates informative spectral bands and mitigates noise.
• The method demonstrates strong potential for both accurate classification and anomaly detection in real-world applications.