Lifting wavelet transform for Vis-NIR spectral data optimization to predict wood density

Visible and near infrared (Vis-NIR) spectroscopy is a mature analytical tool for qualitative and quantitative analysis in various sectors. However, in the face of "curse of dimensionality" due to thousands of wavelengths for a Vis-NIR spectrum of a sample, the complexity of computation and memory will be increased. Additionally, variable optimization technique can be used to improve prediction accuracy through removing some irrelevant information or noise. Wood density is a critical parameter of wood quality because it determines other important traits. Accurate estimation of wood density is becoming increasingly important for forest management and end uses of wood. In this study, the performance of two-dimensional (2D) correlation spectroscopy between wavelengths of various spectral transformations, i.e., reflectance spectra (R), reciprocal (1/R), and logarithm spectra (log (1/R)), were analyzed before optimizing spectral variable. The spectra of optimal transformation were decomposed using biorthogonal wavelet family from 3rd to 8th decomposition level based on lifting wavelet transform (LWT). The optimal wavelet coefficients of LWT were selected based on the performance of calibration set using partial least squares (PLS). Two frequent variable selection methods including uninformative variable elimination (UVE) and competitive adaptive reweighted sampling (CARS) were also compared. The results showed that the dimensionality of spectral matrix was reduced from 2048 to 16 and the best density prediction results of Siberian elm (Ulmus pumila L.) were obtained (R_p²R = 0.899, RMSEP = 0.016) based on LWT.