CT-referenced end-to-end hyperspectral imaging for spatially resolved moisture content mapping in wood

Abstract

Abstract Spatially resolved moisture content (MC) measurement is important for understanding moisture migration and improving drying control in wood. This study developed a CT-referenced end-to-end near-infrared hyperspectral imaging (NIR-HSI) framework for moisture content mapping in Japanese cypress ( Chamaecyparis obtusa ) during drying. X-ray CT images acquired at different moisture states were calibrated to density values and registered to corresponding oven-dry CT images. A latewood mask-based deformable registration method was introduced to compensate for drying-induced deformation, enabling voxel-wise MC calculation. The resulting CT-derived moisture maps were averaged along the thickness direction and screened for reliability, yielding 163 valid paired datasets for partial least squares (PLS) baseline and U-Net model development and evaluation. A conventional PLS regression model was established as a baseline, and a U-Net model was trained to directly predict MC maps from NIR-HSI data. The proposed CT registration method substantially improved image correspondence, increasing the mean Dice similarity coefficient of latewood masks from 0.536 to 0.964. The CT-derived mean MC showed excellent agreement with gravimetric measurements, with an R 2 of 0.9958 and a root mean square error (RMSE) of 3.8729%. Compared with PLS, the end-to-end U-Net model produced more accurate and spatially consistent moisture maps. On the test set, the U-Net reduced the pixel-wise RMSE from 57.71 to 20.87% and increased R 2 , the structural similarity index measure (SSIM), and Pearson’s correlation coefficient from 0.411, 0.6318, and 0.670 to 0.923, 0.9067, and 0.962, respectively. After averaging the predicted maps and comparing them with gravimetric MC, the U-Net also achieved accurate image-level mean-MC prediction on the independent test set, with R 2 = 0.9755, RMSE = 8.5629%, and a ratio of performance to deviation (RPD) of 6.56. Wavelength importance analysis indicated that both PLS and U-Net relied strongly on the water-related absorption region around 960–970 nm. These results demonstrate that combining CT-derived spatial reference data with end-to-end NIR-HSI modeling enables accurate, non-destructive, and spatially resolved moisture content estimation in wood.