Visible and near-infrared (Vis-NIR) spectroscopy, which contains the visible and NIR regions, has been widely applied in agriculture, petroleum, pharmaceuticals, and life sciences, such as soil particle size determination (Gozukara et al., 2022), propane content prediction of liquefied petroleum gas (Dantas et al., 2013), polymorphic forms of fluconazole identification (Mansouri et al., 2021), plant stress detection (Liang et al., 2018), and examination of Zika virus (Fernandes et al., 2018). The visible region (380-750 nm) contains the information of the pigments such as anthocyanin and chlorophyll based on their specific absorption bands (Zahir et al., 2022). Meanwhile, the near-infrared light records the vibration of the hydrogen bonds, for instance, C-H, N-H, and O-H, which are the main components of samples. Therefore, when the Vis-NIR light strikes samples, the light absorbed by samples includes the information of pigments and hydrogen bonds, which can be used to predict samples’ components.

Wood density is an essential indicator for the assessment of wood qualities due to the relationship between the mechanical, optical, and chemical properties (Resquin et al., 2019). The traditional measurement of wood density is laboratory test based on the density formula (ρ=m/v, where ρ is wood density, m and v are the mass and volume of wood samples, respectively), such as China National Standards (GB/T 1933-2009), which is challenging because they require burdensome sample processings, meaning that it is a destructive behavior and difficult to measure numerous samples in a short time. In addition, wood properties are influenced by the tree species and geographical origins (climate, moisture, soil, etc.). Even within the same tree, there exist differences between juvenile and mature wood for density (Krajnc et al., 2021). These differences in wood properties have effect on end use and economic benefits. For example, the yellow rosewood (the raw material of classical Chinese furniture) grown in Hainan province is expensive than other locations due to its great quality (Huang et al., 2018). Therefore, it is necessary to analyze wood properties among different geographical locations and tree species, especially for native tree species and the Convention on International Trade in Endangered Species (CITES) listed species.

Many studies (Zhao et al., 2010; Zhao et al., 2012; Tigabu et al., 2020; Toscano et al., 2022) have demonstrated that Vis-NIR or NIR spectroscopy can be used to determine physical and chemical compositions, mechanical properties, wood microstructure, and seed quality over the years with the advantages of rapid, simple, and non-destructive detection for numerous samples. For the prediction of wood density using spectroscopy, various wood samples and chemometric techniques or the combination of these two sections are the main research directions. In terms of wood science, Schimleck et al. (2003) estimated air-dry density of green Pinus taeda radial samples with NIR spectroscopy, the coefficient of determination (R²) are 0.85 and 0.87 for green and dry wood, respectively. Additionally, in another study, Schimleck et al. (2018) found that air-dry density of Pinus taeda L increased from pith to bark at all heights based on NIR spectroscopy technology. As for chemometric techniques, Zhang et al. (2022) proposed a deep transfer learning hybrid method with automatic calibration capability (Resnet1D-SVR-TrAdaBoost.R2) to predict larch wood density in different moisture contents. Fernandes et al. (2013) compared the effect of two calibration methods (neural networks and partial least squares) on Pinus pinea density, the results demonstrated that neural networks was better than PLS technique. In addition, considering spectra quality and model accuracy, Li et al. (2022) analyzed various spectral pre-processing and multivariate calibration methods in the prediction of Chinese White Poplar density, the results showed that the best prediction was obtained by GRNN models combined with LWT and CARS method. These studies displayed that chemometric techniques are essential for NIR spectra analysis to better explore the relationship between spectra and properties.

The original spectra contain irrelevant information due to the interference of background and environment, therefore, chemometric methods are needed for Vis-NIR spectral analysis. The essence of Vis-NIR spectral data analysis is to extract useful information of components using the appropriate chemometric methods, which include pre-processing, feature variable selection, and multivariate calibration models. Pre-processing is an essential step for improving model prediction accuracy through converting raw spectral data into a new data set without interferences (Bian et al., 2020). The common used pre-processing techniques are multiplicative scatter correction (MSC), the first derivative, Savitzky-Golay (SG) filtering, detrending (DT), wavelet transform (WT) and standard normal variate (SNV) (Dotto et al., 2018; Li et al., 2020; Bian et al., 2022; Carvalho et al., 2022; Ling et al., 2022). Different results will be obtained using various pre-processing methods or their combinations due to the different mechanism and functions, thus, it is important to select the most useful method and prevent the phenomena of over-fitting.

A Vis-NIR spectrum of a sample with the region from 350 to 2500 nm includes 2151 spectral variables, the high-dimensional spectral data results in the “curse of dimensionality”. Therefore, feature wavelengths selection techniques should be employed to address the problem. Uninformative variable elimination (UVE), competitive adaptive reweighted sampling (CARS), genetic algorithm (GA), iteratively retains informative variables (IRIV), and successive projections algorithm (SPA) are among the feature variable selection methods that have been analyzed in recent studies (Centner et al., 1996; Araújo et al., 2001; Ielpo et al., 2017). In addition to these mentioned techniques, a hybrid method that combing two or more methods were made to simplify high-dimensional spectral data. For example, Yun et al. proposed a hybrid strategy based on variable combination population analysis (VCPA), IRIV, and GA to optimize key variables (Yun et al., 2015; Yun et al., 2019).

Multivariate calibration models are employed to analyze the relationship between the selected feature wavelengths and targeted properties. Generally, modeling approaches are divided into linear and non-linear models such as partial least-squares (PLS), principal component regression (PCR), artificial neural network (ANN), and support vector machine (SVM). Regarding the performance of linear and non-linear models, there may exist differences in prediction accuracy for the same sample because of the different strategies. According to the Beer-Lambert’s Law (A = ϵ× l × c, where A is absorbance,ϵ, l, and c are the molar absorption coefficient, path length, and concentration, respectively), there is a linear relationship between absorbance spectra and concentrations (Wang et al., 2011). However, the relationship between spectra and properties are complex or non-linearity due to the interference of environment, such as light scattering (Geladi et al., 1985). As for NIR spectra analysis, a linear method of PLS is usually used to model. If the residuals of the predicted model are normally distributed around zero then the PLS model is accurate. If the residuals will not be equally distributed around zero but will follow perhaps a”banana shaped” curve around zero or some unbalanced pattern, another calibration method or non-linear model will be employed to analyze the relationship between spectral data and properties. It is worth mentioning that non-linear calibration models have ability to perform linear analysis. What’s more, non-linear relationships between spectra and concentrations can be handled by the linear calibration models, but at the cost of the increasing of the multivariate complexity. However, the linear models are not always effective in the spectral data analysis when the noises are multiplicative. Additionally, spectral differences as different unknown samples or growing environment are complex. In this case, it is not enough to deliver the optimal solution only using linear techniques for such complex problem.

Recently, some new algorithms, such as random forest (RF) (Marta et al., 2022) and fruit fly optimization algorithm (FOA) (Li et al., 2019), were applied in the analysis of spectral data. Modeling methods are the same as pre-processing and feature variables selection techniques, they can be combined by two or three approaches to obtain the most suitable model. The limitations of modeling methods (linear and non-linear models) are the optimization of modeling parameters. For instance, the principal component number (PC_S) and the selection of radial basis function (RBF) are the key step of PLS and SVM models, respectively.

In this study, various chemometric methods including a new spectral data de-noising (Lifting wavelet transform, LWT), four feature variables selection techniques (SPA, UVE, IRIV, and CARS), and two hybrid multivariate calibration models, i.e., generalized regression neural network (GRNN) optimized by FOA (GRNN-FOA), and PSO-SVM model optimized by response surface methodology (RSM-PSO-SVM), were compared simultaneously to obtain the most suitable chemometric technique for wood density prediction among different tree species. In addition to, the effect of geographical location on wood Vis-NIR spectra was investigated to predict wood density based on Vis-NIR spectroscopy.

Researchers pay more attention to spectral de-noising or feature variable selection to reduce the influence of irrelevant or interference information in the spectral analysis. A comparison of various chemometric methods including spectral pre-processing, feature variable selection, and the optimization of non-linear calibration models in the wood density prediction among different tree species and geographical origins simultaneously were first analyzed using Vis-NIR spectroscopy in this study. The results demonstrated that LWT outperform WT, MSC, SNV, and raw spectra among these tree species for wood spectra optimization. There are few studies on LWT and WT de-noising in NIR spectra analysis. Abasi et al. (2019) employed WT to optimize Gala apple Vis-NIR spectra in the determination of quality parameters. The R² values were higher than 0.85 for soluble solids content, moisture content, and pH. In the forestry field, WT (Daubechies-5, db5) and LWT (db2) were used to optimize Populus davidiana and larch spectra, respectively in our previous studies (Li et al., 2018a; Li et al., 2018b). In this study, LWT with four wavelet functions including Haar, sym3, db3, and bior1.3 were compared simultaneously among various tree species from two locations. The results demonstrated that there has differences in the optimal de-noising parameters of LWT among these tree species. Therefore, an appropriate pre-processing technique should be selected before building models.

As for feature variable selection. compared to full spectra de-noised by LWT, CARS and IRIV achieved the best results, and the spectral dimensionality was reduced by 51.56%, 99.02%, 98.29%, 98.78%, 98.88%, and 99.27% for Tilia tuan Szyszyl., Acer mono Maxim, Chinese white poplar (Jilin), Japanese elm, Chinese white poplar (Heilongjiang), and Dahurian larch, respectively. In terms of non-linear models, RSM-PSO-SVM delivers the best performance than corresponding FOA-GRNN models, except for Chinese white poplar harvest from Heilongjiang province. In order to better analyze the performance of these chemometric methods, two traditional linear models, i.e., PLS and PCR, were employed to build signal and combined models using raw spectra, respectively. Signal model is a model that includes one tree species. And combined model is a model that includes all tree species and geographical locations simultaneously. The results of these two kinds of models are shown in Table 5 , respectively.

Comparison of the accuracy of PLS and PCR method based on raw spectra ( Table 5 ), the calibration set outperforms the cross-validation set with a higher R², RPD and smaller RMSE, RSD value, regardless of signal and combined models. In addition, PLS model provides high accuracy on various wood density prediction than that of PCR model. However, the performance of combined models was worse than signal model, especially for PCR method.

Table 6 , 7 show the prediction accuracy of signal models and combined models, respectively. Similar to the results of calibration set, the PLS method achieved a higher prediction accuracy than PCR model. Additionally, the combined model delivers the worst performance on prediction dataset. Compared to the optimal model of RSM-PSO-SVM, the PLS approach is relatively bad except the Chinese white poplar from Jilin province.

This study demonstrates the feasibility of using Vis-NIR spectra combined with various chemometric methods including spectral de-noising, feature variables selection, and the optimization of modeling parameters, to predict wood density. LWT is excellent for spectral de-noising when comparing traditional methods. CARS and IRIV outperforms in the feature variables selection among these tree species. In terms of linear and non-linear calibration models, RSM-PSO-SVM delivers the best performance except for Chinese white poplar. The optimal model of Chinese white poplar from Jilin and Heilongjiang province are PLS and FOA-GRNN model, respectively. These results indicate that the geographical location has an effect on the selection of chemometric methods in wood density determination. Additionally, in order to overcome the difference of the optimal model, model transfer will be used to predict wood properties of different locations in future studies.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

YiL analyzed the data and wrote the original draft. BV contributed to editing. FH performed formal analysis. ZP performed supervision. All authors contributed to the article and approved the submitted version.