Larix Gmelinii Wood Properties Prediction Based On NIR And The Thinning Intensity Optimization

Larix gmelinii is an important tree species in northeast forest region and has a great potential for further utilization.Fast and accurate acquisition of wood property information could aids in precise classification,utilization and directional cultivation.Compared with the traditional laboratory based testing methods,near-infrared spectroscopy(NIRS)technology has the advantages of real-time,fast,nondestructive and field operation.NIRS is very sensitive to wood properties,enables it to analyze the composition and structure of samples.As NIRS contain a lot of background and noise,and it is difficult and workload to model and optimize high-dimensional spectral data.Therefore,this study used the combination of NIRS and chemometrics to reduce the difficulty of modeling and improve the prediction accuracy.In this study,the natural secondary forest of Larix gmelinii in the Greater Khingan Mountains was selected as the research object,and the NIRS prediction model was constructed by different modeling methods.The optimal NIRS prediction model was established by selecting appropriate pretreatment methods and band optimization methods,which provided an effective means for the real-time detection and transformation of wood properties in the future.The study on wood properties of different thinning intensities can provide reference for adjusting forest cultivation and forest management measures,and provide theoretical basis and technical guidance for the directional transformation of natural secondary larch forests.The main research results are as follows:(1)The wood density NIRS model of Larix gmelinii was constructed by using Artificial neural network(Res Net),Back propagation(BP)neural network and Partial least squares regression(PLSR)methods.The models established by PLSR had the highest accuracy and the best stability,with R~2 above 0.80 and RMSE below 0.03.The performance of Res Net prediction model was the worst,with R~2 below 0.50 and RMSE above 0.04.The comparison of model accuracy and stability showed that the effects of three modeling were:PLSR model>BP neural network model>Res Net model.(2)In this experiment,five pretreatment methods(SG smoothing,standard normal transformation(SNV),multivariate scattering correction(MSC),first derivative and second derivative)and 28 combined pretreatment methods were used to pretreat wood lignin content,cellulose content and hemicellulose content by NIRS.The results showed that the first derivative was the optimal single pretreatment method,which can effectively improve the model accuracy and stability.The optimal pretreatment methods for lignin content,cellulose content and hemicellulosic content of larch wood were the combination of multiple pretreatment methods,which were the first derivative-SNV-SG,the first derivative-SNV and the first derivative-SNV,respectively.The optimal model had high accuracy and good prediction robustness.In the pretreatment combination,the optimal pretreatment sequence is baseline correction,scattering correction and smoothing.(3)To further reduce the difficulty of modeling,this study compared the PLSR model results of original spectra,CARS,UVE,SPA,Si PLS,UVE-CARS,Si PLS-UVE,Si PLS-CARS and Si PLS-SPA.The results show that the single wavelength variable selection method is CARS.The combination of characteristic variable selection method is better than the NIRS modeling of single wavelength variable selection method.Si PLS-CARS is the optimal characteristic wavelength variable selection method for the compression strength along grain,bending strength and bending modulus of elasticity.The model has high prediction accuracy and good robustness,which can effectively predict the mechanical properties and provide technical support for real-time and rapid determination of wood mechanical properties.(4)In order to realize model transfer between different near-infrared spectrometers and improve the applicability of near-infrared spectral prediction models in field applications,PLS and Di PLS methods were used to study model transfer between different NIR spectrometers.In this experiment,supervised and unsupervised learning of the Di-PLS algorithm was used with an optimal sampling intensity of 40%.Di-PLS can effectively align the distribution of the source and target domains,thus improving the model accuracy.Compared with the PLS model,the accuracy of Di-PLS model was significantly improved.Supervised learning provides superior modelling results compared to unsupervised learning.In the model construction of different NIR spectrometers,the Di-PLS method can effectively improve the model accuracy and model stability,and realize the model transfer between different NIR spectrometers of wood density.(5)In this study,natural secondary forests of Larix gmelinii in the Greater Khingan Mountains with thinning intensities of CK(0%),34.38%,16.75%and 59.92%were selected as the test objects.Near infrared spectrum model was used to predict wood properties of different thinning intensities.It includes wood density,cellulose content,hemicellulose content,lignin content,compressive strength,flexural strength and flexural elastic modulus.Different thinning intensities had significant effects on wood properties,among which 16.75%and 34.38%had significant effects.A combination weight determined by AHP-Entropy was adopted to conduct a comprehensive evaluation of wood quality of different thinning intensities through principal component analysis.The comprehensive effect scores of the 4 tending thinning intensity groups were 34.38%>59.92%>CK>16.75%.In this study,the NIR prediction model of larch wood property was constructed,which provided theoretical basis and technical support for the real-time detection of wood properties,support the fine classification of wood and targeted cultivation of forest trees.Tending thinning intensity has a significant impact on wood properties,and the optimal thinning intensity is determined to be 34.38%,which provides a theoretical basis for the subsequent forest directed management.