【作者】 王学顺；

【导师】 戚大伟；

【作者基本信息】 东北林业大学 ， 森林工程， 2010， 博士

【摘要】 随着我国经济、社会的快速发展,对木材的需求量日益增加。为了缓解木材供需矛盾紧张的状况,必须加快林木的定向培育、实现木材的高效利用。因此,寻求快速、准确的木材性质检测方法,对于提高我国林木培育质量、木材的遗传改良以及木材的高效利用具有重要意义。近红外光谱分析技术是一种新型的分析技术,能够快速、准确地对固体、液体、粉末状等有机物样品的物理、力学和化学性质等进行无损检测。它综合运用了现代光谱信息技术、计算机信息处理技术以及化学计量学数据分析和多元校正技术最新研究成果,并使之融为一体,以其独有的特点在众多领域得到了广泛应用。近红外光谱信息的特征提取及预测模型的建立是近红外光谱分析的关键技术,如何从复杂、重叠、变动的光谱中提取有效光谱信息是影响近红外光谱技术发展的重要问题。论文在综合分析了近红外光谱信息产生机理的基础上,对木材近红外光谱的信息特征提取及其定量表示进行了研究。以我国人工林杉木和桉树近红外光谱为信息源,对近红外光谱信息的特征提取进行了定量分析,利用偏最小二乘法建立了杉木密度和桉树木质素含量预测模型,对比分析了不同光谱信息提取方法对所建模型的影响。论文主要研究内容包括：(1)以光谱二阶导数数据平方和与均方根误差为标准,比较分析了光谱数据平均平滑法和卷积平滑法在不同窗口下提取光谱信息的效果。平均平滑法当窗口宽度为15、17和19时,提取及保留光谱有效信息效果最好；卷积平滑法当窗口的最佳宽度为13、15和17时,提取及保留光谱有效信息效果最好。光谱数据平均平滑法和卷积平滑法可以去除光谱测量噪声,优化光谱信息。(2)提出了基于移动窗口方差法的木材近红外光谱信息处理方法。该方法采用局部波段上光谱数据的方差来衡量数据的起伏,以识别光谱信息起伏较大的波段,然后对这些波段进行去噪处理。以光谱数据平方和与均方根误差为标准,研究了选取多种窗口、多种阈值条件下该方法对光谱信息的处理效果。结果显示,当窗口大小取4-8、窗口方差取0.87-0.9或0.94-0.95的下侧分位数为阈值时,该方法具有很好的去除光谱噪声的效果。(3)采用小波变换阈值法对木材一阶导数光谱进行去噪研究。以信噪比和均方根误差为标准,对固定阈值规则、无偏似然阈值规则、混合阈值规则和极大极小阈值规则去除导数光谱噪声进行了对比分析。在对导数光谱信号进行小波4尺度分解、选取固定硬阈值规则时,导数光谱信噪比为10.22,均方根误差为0.000307,去噪效果优于其他方法。(4)采用小波变换模极大值进行光谱信息特征提取研究。根据信号和噪声在不同尺度上的极大值的不同传播特性,将木材近红外光谱信号进行8尺度小波分解,在相邻尺度间搜寻信号和噪声的小波模极大点,提取信号的模极大值,消除噪声模极大值,经逆小波变换重构去噪信号,达到提取光谱特征信息、去除噪声的目的。经小波4尺度分解,小波模极大值去噪后的光谱信噪比达到15.14,均方根误差为0.000953。小波变换模极大值可以有效提取光谱特征信息,去除光谱噪声。(5)采用偏最小二乘法建立了杉木密度预测模型。通过对原始光谱进行一阶导数、二阶导数、移动平均平滑法、卷积平滑法、移动窗口方差法、多元散射校正、数据标准化、小波阈值法和小波模极大值进行预处理后的光谱数据所建立的杉木密度预测模型进行综合分析,移动窗口方差法和小波模极大值法所建校正集模型相关系数分别为0.9391和0.9405,预测集相关系数分别为0.8706和0.8756,所建模型预测效果好于其他光谱信息预处理方法。对模型进一步优化,将一阶导数光谱进行25点卷积平滑剔除6个异常样本并选取10个主成分,所建杉木密度校正集模型相关系数为0.9692、预测集相关系数为0.8976。(6)采用偏最小二乘法建立了桉树木质素含量预测模型。通过对预处理后的光谱数据所建立的桉树木质素含量预测模型的综合分析,结果显示,移动窗口方差法所建校正集模型相关系数为0.9011,预测集相关系数为0.8414,预测效果好于其他预处理方法。对模型进一步优化,将原始光谱的一阶导数进行19点移动平均平滑剔除4个异常样本并选取7个主成分,所建校正集模型相关系数为0.9724、预测集相关系数为0.8768。更多还原

【Abstract】 With the rapid development of China’s economy and society, and the growing demand for wood, it is necessary to speed up the directed breeding of wood as well as improve the efficiency in its utilization so as to ease the pressure of its supply. It is, therefore, significant to find fast and accurate methods to detect wood properties with regard to the silviculture improvement, genetic modification and effective utilization for China.As a new analytical technique, the NIR spectroscopy can make a fast and accurate nondestructive examination of the physical, mechanical and chemical properties of organic solid, liquid, and powder samples. It combines and integrates the latest findings in modern spectroscopy, computerized information processing, stechiometric data analysis and multivariate correction, and is applied in many fields with its uniqueness. There are still some technical problems in NIR spectroscopy remaining to be solved. One of the problems is how to extract and enhance the effective information in the complicated, overlapping and changing spectrum, so as to provide excellent spectral information to build an NIR composition concentration prediction model.This dissertation has made a systematic study of how to extract spectral signatures from wood NIR and how give a quantitative expression, based on the overall analysis of generation mechanism of NIR. With the Chinese fir and eucalyptus NIR in planted forest as the information source, it makes a quantitative analysis on the method of NIR information extraction, builds Chinese fir density and eucalyptus lignin content prediction model by using the principle component analysis and partial least square method, and compares the influences by different spectral information extraction methods.The dissertation mainly studies:(1) Taking the quadratic sum of the second derivatives of spectra as well as the root-mean-square error as the standard, it compares the influences of average smoothing and fold smoothing of spectral data on the spectral information. For the average smoothing method, when the window widths are 15,17 and 19, it shows sound results in extraction and maintenance of effective spectral information; while for the fold smoothing method, when the optimum window widths are 13,15,17, it also shows sound results in extraction and maintenance of effective spectral information. Both averaging smoothing and fold smoothing methods used in spectral data can remove the measurement noise, and optimize the spectral information.(2) It provides a method for preprocessing the wood NIR information, based on the moving window variance method which measures the fluctuations of data using the variances of spectral data over a local band, so as to identify the bands with large fluctuations and then de-noise these bands. Taking the quadratic sum and root-mean-square error of spectral data as the standard, it analyzes and discusses the processing effects of this method under different windows and different thresholds. When the window size takes 4 to 8, the lower quartile of 0.87-0.9 or 0.94-0.95 of window variance are thresholds, this method produces sound de-noising effects.(3) It studies the de-noising of the first derivative spectrum from wood by using wavelet transform. Taking the signal-to-noise ratio and root-mean-square error as the standards, it compares the four wavelet threshold rules as fixed sqtwolog, partial likelihood estimation (rigrsure), mixd (heursure) and maximum and minimum (minimaxi), and studies the de-noising of derivative spectral data. When spectral information is decomposed under the scale of 4, and the fixed threshold rule is used, the signal-to-noise ratio (SNR) is 10.22, and root-mean-square error (RMSE) is 0.000307, which shows that the de-noising effect is better than that of other methods.(4) It studies the signature extraction of spectral signals by using the wavelet transform modulus maximum. According to the different propagation properties of noise and signal at the maxima of wavelet transform, it carries out wavelet decomposition under scale 8 on the wood NIR signals, and searches the wavelet modulus maxima of signals and noise in the neighboring scales., extracts signals modulus maximum, eliminate noise modulus maximum. After wavelet inverse transform, it rebuilds de-noised signals, which achieves the purposes of extracting spectral signatures, and removing the noise. After decomposition under the scale 4, and de-noising by wavelet modulus maximum, the signal-to-noise ratio (SNR) reaches 15.14, and the root-mean-square error (RMSE) is 0.000953. Wavelet transform modulus maximum can effectively extract the spectral information and remove the spectral noise.(5) It builds the Chinese fir density prediction model by using the principle component regression method and partial least square method. After the comprehensive analysis and assessment of Chinese fir density prediction model built based on the original spectra and the spectral data preprocessed by the first derivative method, the second derivative method, the moving average smoothing method, fold smoothing method, moving window variance method, multivariate scattering correction, data standardization, wavelet threshold method, and wavelet modulus maximum, the correlation coefficients of correction set models by the moving window variance method and wavelet maximum are 0.9391 and 0.9405, and for prediction set, the correlation coefficients are 0.8706 and 0.8756 respectively, which shows better results than that by other preprocessing methods. It carries out 25 point fold smoothing on the first derivative of original spectra, rejects 6 abnormal samples, and selects 10 principle components to further optimize the models, then the correlation coefficient for the correction set model built is 0.9692, and for prediction set, it is 0.8976.(6) It builds the prediction model of eucalyptus lignin content by the partial least square method. After the whole analysis of model of eucalyptus lignin which after predispose.The results shows that the correlation coefficient of correction set models by the moving window variance method is 0.9011, and for prediction set, the correlation coefficient are 0.8414, which shows a better result than that by other preprocessing methods. It carries out 19-point moving average smoothing on the first derivative of original spectra, rejects 4 abnormal samples, and selects 7 principle components to further optimize the models, the correlation coefficient for the correction set model built is 0.9742, and for prediction set, it is 0.8768.更多还原