Visible-Near Infrared Spectroscopy For Wood Species Identification And Density Prediction

The development of forestry in China is currently in an important stage from digital forestry to smart forestry.However,the testing of wood properties lags behind production due to the long-term crude wood production and management,time-consuming,laborious,and destructive traditional testing methods.Additionally,it is difficult to rapidly test many samples,which seriously hinder the development of forestry intelligence.Visible and near infrared(Vis-NIR)spectroscopy,as a fast,green high-tech detection technology,has been widely used in agriculture,petrochemical,pharmaceutical and other fields due to the advantages of without sample pretreatment,high repeatability,low cost of analysis,and rapid detection of large batches of samples.In the field of forestry,it is still in the stage of exploration and has a good prospect.However,there are still some problems that hinder the in-depth application of visible and near-infrared spectroscopy in the forestry due to the heterogeneity,diversity,and complexity of wood.For example,the different prediction accuracy will be obtained with different spectral transformations;non-correlated component information such as noise of spectra results in low prediction accuracy;longer computation time and a large amount of storage space are required due to the high dimensional matrices generated by thousands of dimensional spectral data.In addition to,it is difficult to obtain satisfactory modeling results with the default modeling parameters.Therefore,in view of the above problems,the prediction of wood properties with Vis-NIR spectroscopy combined with the emerging chemometrics were discussed in this study.The aim of this study is to develop the rapid and efficient qualitative models of trees/origin identification and wood density quantitative models in the beginning of the timber production chain,namely the timber harvesting stage,which will provide theoretical basis and technical support for the end use of wood.The main research contents of this paper are as follows:(1)Construction of qualitative models based on different spectral transformations.The spectral transformation technology was introduced into the qualitative analysis model of wood species and origin identification.The reflectance spectra(R)of six kinds of trees from Jilin Province and Heilongjiang Province were subjected to the reciprocal(1/R)and logarithm reflectance(log(1/R)),support vector machine(SVM)was used to identify origins and species.Generic algorithm(GA),grid search(GS),and particle swarm optimization(PSO)were employed to optimize the parameters of SVM model.In addition to,the same species of Populus vidiana was collected from different geographical origins to analyze the influence of producing area factors on tree species identification.The spectral transformations and SVM were implemented by using Matlab R2010a.(2)Comparison with single and combined linear density Vis-NIR models.Partial least squares(PLS)and principal component analysis(PCA),the classical method of linear modeling,were employed to establish the single model and three different types of joint models based on the optimal spectra of wood.Additionally,the prediction ability of each model for different tree species were analyzed,which laid the foundation for further optimization of the subsequent wood density quantitative model.(3)Optimization of quantitative models based on different algorithms.Emerging chemometrics were employed to remove the noise and irrelevant information of Vis-NIR spectral data.Firstly,the de-noising effect of the second-generation wavelet,lifting wavelet transform(LWT),under different de-noising parameters was analyzed,and the optimal de-noising parameters of LWT,namely mother wavelet,order,and decomposition level,were determined.At the same time,the de-noising effect of LWT,wavelet transform(WT),locally weighted scatterplot smoothing(LOWESS),locally estimated scatterplot smoothing(LOESS),multiplication scatter correction(MSC)and standard normalized variate(SNV)were compared to determine the optimal methods for various tree species.Then,different variable optimization algorithms were used to reduce the dimension of the high-dimensional spectral matrix to provide new ideas and methods for the problem of "matrix dimension disaster" for the wood Vis-NIR spectra.Finally,to solve the problem of randomness and complexity of modeling parameter selection,the optimal wavelength variables were input to the non-linear models of generalized regression neural network(GRNN)and support vector machine(SVM).The fruit fly optimization algorithm(FOA)was used to intelligently optimize the Spread parameters of the GRNN model.The particle swarm optimization(PSO)algorithm was used to optimize the penalty parameters C and RBF kernel functions of the SVM model.In addition,the response surface methodology(RSM)was used to optimize the parameters of PSO-SVM model,i.e.number of cross-validation,maximum generation,and population size to determine the important factor of PSO-SVM model,which improves the accuracy of wood density.