| To be able to breed for suitable trees, it is essential to be able to screen large numbers of individual trees. Measuring wood properties by traditional chemistry is costly and time-consuming. Near infrared spectroscopy has made rapid progress as analytical techniques. Being capable of making nondestructive, rapid, high efficient and convenient analysis, NIR technique is suitable to analyze wood properties.The calibration models to predict the chemical composition, such as moisture content, lignin content, holocellulose, a-cellulose, hemicelluloses and extractive, basic density and pulp yield in Acalia Spp were developed by applying on Fourier near infrared spectroscopy. The factors that influenced the NIR veracity were studied, including resolution, scan number, test repetition, scan numbers, degree of tightnsss, scanning time, pretreatmeng methods, wavenumbers, samples number, the different cross of wood, and the accuracy of reference data.To provide foundation with optimum test condition when modeling, factors affecting the accuracy of wood near-infrared spectroscopy were studied. The results show that scan number does have effect on NIR spectra and prediction models, the calibration and prediction model was robust with 32 scan number. Resolution does not have significant effect on the NIR models, but the spectrum with high resolution was rougher than that with low resolution. Moreover, the scan speed was lower and more data size was required when high resolution was used. The result of NIR model can be enhanced when scan the sample repetitive. The model based on rough powders can gave more accurate evaluation indexes vs that based on fine powders. The best model is based on the mixture powders. Depressing sample could increase errors brought by sample tightness. The best NIR calibration statistics and the most accurate prediction results were aligned with the most accurate reference data. However, based on statistical analysis of numerous calibration samples, it is possible for NIR calibration models to obtain more accurate prediction results than the laboratory reference data used in the calibration sets. It is better to make less search for high accurate reference data and instead to introduce more calibration samples to improve the ruggedness of the calibration models. In order to search an appropriate pretreatment method to determine hemicellulose in Acalia Spp.. The effect of 11 pretreatment methods to the model based on PLS has been compared. The model developed by pretreated spectra had the highest correlation coefficient, and the lowest relative deviation than the model developed by raw spectra. The best pretreatment method was 1st derivative+straight line subtraction. The different smoothing point were secleted when the spectra were pretreated by the 1st derivative+straight line subtraction,17,23 and 25 point smooth can obtain good results. It shows that 6000 cm-1～5500cm-1 band have significant correlation with hemicelluloses content, and select appropriate wave band is very important for a good calibration model. Selection of representative samples for calibration can directly influence the representative and accurateness of the model. Two methods based on the grads of samples'concent and the Mahalanobis distance of spectral was compared. The later had built better calibration.The study showed that NIR analysis can be reliably used to predict moisture content, lignin content, hemicelluloses, and extractive content in Acacia spp.. NIR calibration predicted values for moisture were close to the laboratory results. The regression results obtained were explained with an R2CV=0.9884 and RMSECV=0.194%, R2val=0.9886, RMSEP=0.216%. The cross validation of lignin content is explained with an R2CV=0.9553 and RMSECV=0.371%, R2val= 0.9425, RMSEP=0.5%. The cross validation of extractive content is explained with an R2CV=0.9345 and RMSECV=0.227%, R2va,= 0.9384, RMSEP=0.267%. The cross validation of hemicellulose content is explained with an R2CV =0.9382 and RMSECV=0.509%, R2val= 0.9137, RMSEP=0.575%. The cross-calibration results of the holocellulose and a-cellulose were not as good as for the hemicellulose and lignin's. The cross-calibtation models for holocellulose resulted in an R2CV of 0.8922 while fora-cellulose the R2CV was 0.858, and the RMSECV is 0.573% and 0.828%. When used the cross-calibration to predict the validation set, the R2val of holocellulose and a-cellulose were 0.858 and 0.7495, respectively. The RMSEP were 0.828% and 1.17%, respectively. The results indicated that the accuracy of models was influenced by the sample surface from which NIR spectra were obtained. The model based on transverse section of wood owned the highest accuracy, followed sequentially by the models respectively based on the radial section tangential section of wood. Good correlations between NIR spectra and basic density were achieved for wood meal and solid wood. The screened pulp yields were predicted by NIR. The regression results obtained were explained with an R2CV=0.7327 and RMSECV=0.739% when the original spectrum was expressed as the Straight line subtraction and the wavelength range between 9770.1cm-1 and 5446.3cm-1.The prediction bias is very large when using the calibration built with samples from Guangxi province to predict Fujian samples, however, addition of three Fujian samples to the Guangxi calibration set was sufficient to greatly reduce predictive errors and that the inclusion of eight Fujian samples in the Guangxi set was sufficient to give relatively stable predictive errors. The RMSEP is very large when using the calibration built with Acacia spp samples to predict Triploid Populus tomentosa samples. Addition of four Triploid Populus tomentosa samples to the Acacia spp samples set was greatly reduce the predictive errors.According to the chemical component, microstructure and pulp yield, the 7 years old A. cincinnata, in Guangxi qinlian farm, the A. melanoxylon and A.mangium×A.auriculiformis in Guangxi gaofeng farm and A. cincinnata in Fujian are fit to pulping.
|
PDF Full Text Request