Detection of moisture content in logs using multi-parameter GPR signal analysis and neural network models

Abstract To address the low accuracy of non-destructive detection of moisture content (MC) of logs (especially in small diameters) by ground penetrating radar (GPR) signals, the MC of 10–15 cm diameter spruce, Manchurian ash, and white birch logs were predicted using the time-frequency parameters of the GPR signals and a back-propagation neural network (BPNN) model. B-scan signals were obtained using tree radar on the barks of discs selected from fresh green logs. Then, 31 time-frequency parameters from the B-scan signals were optimised using the least absolute shrinkage and selection operator (Lasso) and principal component analysis (PCA). Finally, the log MCs of the single and hybrid models was predicted using the BPNN. The accuracy of the least absolute shrinkage and selection operator and back-propagation neural network (Lasso-BP) were higher than those of the principal component analysis and back-propagation neural network (PCA-BP), and the BPNN. The individual species and hybrid models both have good predictive capability; when the log MC is below 20%, the maximum residual errors are relatively small, almost within 6% and 10%, respectively. These models significantly improve the accuracy of non-destructive detection of log MC and are beneficial for efficient wood processing.