Transfer learning for predicting wood density of different tree species: calibration transfer from portable NIR spectrometer to hyperspectral imaging

Abstract

Wood density is a crucial property indicator for construction material selection, quality assessment, and modification. Spectral analysis techniques and chemometric models offer potential solutions for the rapid and non-destructive assessment of wood density. However, probe-contact spectroscopy has low efficiency in spectrum collection, and spectral models are highly specific to variations in instruments and samples. Traditional calibration transfer methods are diverse and struggle to adapt to domains with significant distributional differences. By simulating operations under natural light, this work aimed at exploring a deep transfer-learning strategy, facilitating the transfer of wood density prediction models between different instruments [from portable near-infrared (NIR) spectrometers to hyperspectral-imaging (HSI) imagers] and among tree species (two softwood and two hardwood species). A bidirectional gated recurrent unit plus attention layer (BiGRUattention) was employed as the basic topology for the deep network. The results indicated that the generalization ability and robustness of HSI model transferred by deep adversarial transfer-learning strategy, including domain-adversarial-neural Network (DANN) and dynamic-adversarial- adaptation network (DAAN), surpassed traditional calibration transfer and deep transfer-learning methods, achieving a level comparable to NIR-calibrated models. DAAN based on Wasserstein distance with gradient penalty (WgpDAAN) optimized model accuracy, convergence speed, and stability. The deep adversarial transfer-learning model could be adapted to wood spectral data from different instruments and tree species, where WgpDAAN significantly reduced modeling costs and enhanced productivity, and could be extended to detecting and characterizing other wood properties.