CNN-CBAM Prediction Model for Moisture Content of Raw Materials for Wood-Based Panels Based on Near-Infrared Spectroscopy

Abstract

Wood moisture content critically influences the performance and processing quality of wood-based panels. Therefore, precise and rapid prediction of wood moisture content is essential for enhancing production efficiency and ensuring product quality. In this study, three representative wood raw materials for wood-based panels (Tiia, Larix, and Picea) were investigated. Their near-infrared spectral (NIRS) data were collected, and anomalous samples were identified and removed. The spectral data were then pre-processed, and the Competitive Adaptive Reweighted Sampling (CARS) algorithm was applied to extract features highly correlated with wood moisture content. Subsequently, Partial Least Squares (PLS) regression, Convolutional Neural Network (CNN), CNN combined with channel attention mechanism (CNN-CAM), CNN combined with spatial attention mechanism (CNN-SAM), and CNN combined with Convolutional Block Attention Mechanism (CNN-CBAM) models were developed and evaluated for their suitability and stability in predicting wood moisture content. A total of 305 wood-based panel materials samples were randomly divided into calibration and prediction sets at a 3:1 ratio. The samples underwent outlier removal, spectral preprocessing, feature extraction, and modeling. Among all methods tested, the CNN-CBAM model combined with CARS-based feature selection achieved the highest accuracy, yielding a prediction set coefficient of determination (R²) of 0.9730 and a root-mean-square error of prediction (RMSEP) of 0.1840. The CNN-CBAM model demonstrates excellent performance in predicting the moisture content of Tiia, Larix, and Picea used in wood-based panels. By incorporating the convolutional block attention mechanism, the model effectively captures key spectral features, accounting for both global inter-channel correlations and local spatial distinctions. This significantly enhances prediction accuracy and generalization, providing an efficient tool for online monitoring and quality control in the production of wood-based panels.