Spectrometric prediction of wood basic density by comparison of different grain angles and variable selection methods

Abstract

Background Wood basic density (WBD) is one of the most crucial wood property in tree and mainly determined the end use of wood for industry. However, the measurement WBD is time- and cost-consuming, which an alternatively fast and no-destructive measurement is needed. In this study, capability of NIR spectroscopy combined with partial least squares regression (PLSR) to quantify the WBD were examined in multiple wood species. To obtain more accurate and robust prediction models, the grain angle (0° (transverse surface), 45°, 90° (radial surface)) influence on the collection of solid wood spectra and a comparison of found variable selection methods for NIR spectral variables optimization were conducted, including significant Multivariate Correlation (sMC), Regularized elimination procedure (Rep), Iterative predictor weighting (Ipw) and Genetic algorithm (Ga). Models made by random calibration data selection were conducted 200 times performance evaluation. Results These results indicate that 90° angle models display relatively highest efficiency than other angle models, mixed angle model yield a satisfied WBD prediction results as well and could reduce the influence of grain angle. Rep method shows a higher efficiency than other methods which could eliminate the uninformative variables and enhance the predictive performance of 90° angle and mix angle models. Conclusions This study is potentially shown that the WBD (g/cm3) on solid wood across grain angles and varies wood species could be measured in a rapid and efficient way using NIR technology. Combined with the PLSR model, our methodology could serve as a tool for wood properties breeding and silviculture study.

Non-destructive Measurements of Toona sinensis Chlorophyll and Nitrogen Content Under Drought Stress Using Near Infrared Spectroscopy

“…At present, there are many kinds of spectral preprocessing methods, which can be divided into baseline correction, scatter correction, smoothing, etc. according to the effect of preprocessing (Li et al, 2019(Li et al, , 2021. Baseline correction includes first-order and second-order derivative, etc.…”

“…The NIRS data were taken from the upside surface of the leaves three times with a handheld fiber optic contact probe from a fieldbased spectrometer (LF-2500, Spectral Evolution, United States) (Li et al, 2021). Each spectrum took on average 20 scans with 8 m/s integration time and a range between 1,100 and 2,500 nm with a 6 nm spectral resolution.…”

“…The PLSR model (Hansen and Schjoerring, 2003) is a statistical linear method for fitting a curve by minimizing the sum of squared deviations, which combines the advantages of multiple linear regression, correlation analysis, and principal components. It is broadly applied in the near-infrared spectroscopy context (Li et al, 2021). Here, the samples were randomly split 100 times into calibration (80%) for model building and validation (20%) for testing.…”

“…Numerous research investigations have used NIRS to predict the ecophysiological variables related to plant drought stress, which constrains relative water and leaf water in disease-resistant trees (Warburton, 2014;Conrad and Bonello, 2016). Partial least squares regression (PLSR) is one of the most frequently used chemometric methods in spectral calibration analysis (Li et al, 2021). It has advantages when large amounts of data with redundancy and high collinearity exist and when the number of variables is greater than the number of samples (Liang et al, 2020).…”

“…It has advantages when large amounts of data with redundancy and high collinearity exist and when the number of variables is greater than the number of samples (Liang et al, 2020). Pre-processing approaches like standard normal variate (SNV), smoothing and derivatives are often required in the process of spectral analysis (Li et al, 2021). In addition, we can also choose a variety of variable selection methods to reduce the impact of irrelevant variables on the accuracy of the model (Shao et al, 2017).…”