Near infrared spectroscopy for wood identification based on wavelet convolutional networks

Deep learning-based methods have demonstrated great potential for wood species identification using near-infrared spectroscopy (NIRS). However, extracting intrinsic and subtle chemical absorption signatures from limited spectral data poses a long-standing challenge for purely data-driven models. This limitation causes these models to overfit to high-frequency instrumental noise, which consequently degrades the model's generalization and anomaly rejection capabilities. To address this problem, we proposed a wavelet-enhanced physics-informed neural network (Wavelet-PINN) for robust and accurate wood species identification. Specifically, the model decomposes 1D spectral signals into multiscale time-frequency components via the discrete wavelet transform (DWT) to isolate core structural trends from background noise. We introduce a physics-aware gating mechanism guided by Sobolev gradient losses to generate physical attention masks, which adaptively highlight informative absorption bands. The learnable physical priors are subsequently embedded into the hierarchical deep features of the backbone network. Additionally, we build an independent autoencoder (AE) in parallel for unsupervised signal reconstruction, which regularizes the solution space via a dual-task optimization objective. Extensive experiments on a comprehensive wood species dataset validate that the Wavelet-PINN achieves superior classification performance to state-of-the-art deep learning methods. Notably, Wavelet-PINN attains a classification accuracy of 98.81%, surpassing both Transformer-based and polynomial network-based baselines. Concurrently, the reconstruction branch offers an effective physical threshold for anomaly detection in open-world settings.