| As one of the detection methods for wood moisture content,near-infrared(NIR)spectroscopy has become a powerful tool for wood moisture content detection.In the northeastern forest area of my country,the low temperature below-20℃ and the temperature difference of about 60℃ in different seasons are relatively rare in the application of NIR spectroscopy of other samples.This discrepancy between experimental research and actual production seriously affects the widespread application of NIR spectroscopy in wood production.Therefore,this paper explores the effect of temperature on wood moisture content detection by NIR and try to improve the prediction accuracy of moisture content at different temperatures from the perspective of spectral preprocessing and modeling.The main research contents are as follows:(1)The changes of the near-infrared spectrum of wood at different temperatures and the effect of temperature changes on the near-infrared prediction of wood moisture content were explored.The near-infrared spectra were collected from 75 log samples of Pinus sylvestris,Fraxinus mandshurica,Populus chinensis and Korean pine totaling 300 samples under the conditions of different temperatures and moisture contents.In order to explore the effect of temperature changes on the prediction accuracy of the wood moisture content model,the calibration set at a single temperature and the validation set at each temperature were used to establish a partial least squares moisture content prediction model.In order to explore the change law of wood near-infrared spectrum when temperature changes,infrared spectral data were collected at different temperatures under the same moisture content,and then they were performed spectral averaging,differential observation,principal component analysis and partial least squares discriminant analysis.The results showed that temperature had a significant effect on the spectrum of wood samples.Principal component analysis and discriminant analysis showed that the samples at different temperatures had obvious clustering trends,and the temperature discrimination accuracy was 96.1%.Temperature will affect the position and absorbance of the near-infrared spectrum of wood at specific wavelengths.Under the same moisture content,with the increase of temperature,the absorption peaks at specific positions tend to shift to high-frequency bands gradually.The shift of the wave crest changes is more obvious at low temperatures below zero.The PLS moisture content prediction models at different temperatures have different adaptability to temperature changes,and the wood moisture content prediction models are more suitable for detecting samples at the same temperature as the modeling samples.Therefore,temperature change is a disturbing factor that cannot be ignored in the process of detecting wood moisture content by NIR.(2)In order to apply near-infrared spectroscopy technology to achieve non-destructive testing of the moisture content of logs at low temperature,and improve the prediction accuracy of it.At the outdoor temperature of-20°C in winter,220 ash log blocks with different moisture contents were obtained from the forest farm.And the near-infrared spectra were collected from them under normal temperature(20°C)and low temperature(-20°C).Principal component analysis was used to compare the differences between the two states under the same moisture content.9 spectral preprocessing algorithms from four categories of baseline correction,scattering correction,smoothing and scaling were used to optimize the spectrum by single method and combined optimization respectively.The average spectra of the low temperature and normal temperature states are different in the intensity and position of multiple absorption peaks.The samples of the two states on the first three principal component score graphs have obvious separation.The influence of low temperature cannot be ignored.Under the optimization of a single preprocessing algorithm,both the moving average smoothing method and the SG-smoothing method have good denoising effects on the spectrum,and the RMSEP is0.4562 and 0.4457,respectively.Among the three scaling algorithms,centralization and normalization are significantly worse than the optimization effect of normalization,with RMSEP of 0.5329,0.5202 and 0.3994,respectively.The first-order and second-order derivatives have obvious effects on the baseline correction of the spectrum,but they also amplify the spectral noise,with RMSEP of 0.4126 and 0.4895,respectively.MSC and SNV handled most of the spectral scattering,with SNV performing the best among the single preprocessing algorithm with Rp of 0.8041 and RMSEP of 0.3841.After combining the selected single preprocessing algorithms,the combined algorithms are generally better than the single algorithm.The combination of SG smoothing and the first derivative solves the problem of noise amplification,reducing the RMSEP to 0.2331.Among the combined preprocessing algorithms,the combination of SG smoothing,SNV,and first derivative performed the best,with a validation set correlation coefficient Rp of 0.9128,an RMSEP of 0.1774,and a 69.85%increase in validation set prediction accuracy.Near-infrared spectroscopy can achieve nondestructive detection of low-temperature wood moisture content,and spectral optimization can significantly improve the accuracy of low-temperature moisture content detection models by comparing and screening different pretreatments.(3)In order to further improve the detection accuracy of wood moisture content under temperature variation by near infrared spectroscopy,the pre-model was optimized from the perspective of modeling.The global calibration model was theoretically deduced,and the validity of PLS global calibration model was verified theoretically from the perspective of prediction variance,and the accuracy of different models was analyzed and compared.The results show that the accuracy of the global calibration model is lower than that of the constant temperature model,but the variance is smaller than that without temperature correction.Combined with SG smoothing + 1st +SNV pretreatment,the spectra at different temperatures were integrated to form a global temperature correction set,and a global calibration model for water content was established.Finally,independent samples were selected for repeatability test and accuracy test.The results show that compared with single temperature model,PLS global calibration model has better prediction effect,adaptability and application potential for temperature change,and RMSEP is lower than most single temperature models when detecting wood moisture content. |
