【作者】 佟达；

【导师】 宋魁彦；

【作者基本信息】 东北林业大学 ， 木材科学与技术， 2013， 博士

【摘要】 近年来,我国森林资源总量虽然逐年增加,但大径级可采优质木材却在不断减少,天然林保护工程等重点生态工程的实施,使得国内木材资源的供给压力集中到了人工林。人工林存在单产低、质量差、林龄结构不合理等问题,因此培育人工林获得优质木材已受到世界各国森林培育和木材科学研究者的普遍重视。人工林木材材质材性预测模型的研究,有助于合理确定人工林轮伐期和天然林的更新选择,也有助于木材节约、高效、合理利用。本文以东北主要人工林树种核桃楸和水曲柳的生长轮材质材性特征和木材物理力学特征为研究对象,主要进行三个方面研究：(1)根据生长轮材质材性的径向变异规律,采用有序聚类最优分割法、主成分聚类法、BP神经网络法和支持向量机法等分类方法,分别界定树木幼龄材与成熟材的分界点,对分类结果进行比较分析,明确每种分类方法特点和分类准确度；(2)在得到幼龄材与成熟材分界点的情况下,采用回归方程法、时间序列法、BP神经网络法和支持向量机法等预测方法,根据幼龄材材质材性预测成熟材材质材性,从成熟预测相对误差和标准差与整体预测相对误差和标准差四个方面进行比较分析,明确每种预测方法的特点和预测精准度；(3)在明确支持向量机法具有良好的回归拟合能力和泛化能力的基础上,首先建立生长轮材质材性特征因子间的关系模型,其次建立木材物理力学特征因子间的关系模型,最后建立生长轮材质材性与木材物理力学特征因子间的关系模型,以相关系数R大于0.83的特征因子为核心,最终建立木材材质材性关系模型。通过研究,本文主要得出如下结论：(1)以生长轮材质材性综合指标为研究对象,采用支持向量机法界定核桃楸幼龄材与成熟材的分界点为树木生长的第18年,材质材性训练集的选择以树木生长前6-10年与后2-6年组合为主；界定水曲柳幼龄材与成熟材的分界点为树木生长的第23年,材质材性训练集的选择以树木生长前10-14年与后2-10年组合为主。(2)以生长轮材质材性综合指标和单项指标为研究对象,界定核桃楸和水曲柳幼龄材与成熟材,采用支持向量机法得到的分类结果与主成分聚类法和BP神经网络法基本相同；综合指标与有序聚类最优分割法得到的分类结果差别年限较大,单项指标与有序聚类最优分割法得到的分类结果基本相同。(3)在界定树木幼龄材与成熟材的过程中,有序聚类最优分割法以单项指标为研究对象得到分类结果的准确性优于综合指标；BP神经网络法和支持向量机法以综合指标为研究对象得到分类结果的准确性优于单项指标；主成分聚类法以生长轮材质材性综合指标为研究对象,能够得到单项材质材性指标的贡献率,采用2个主成分能够概括生长轮材质材性,能够用图解法直观给出聚类结果。(4)在成熟材材质材性的预测过程中,采用回归方程法得到的预测值对部分实测值的离散点拟合不好,能够体现幼龄材材质材性变化趋势,但不能表现成熟材材质材性变化趋势；采用时间序列法得到的预测值能够拟合幼龄材实测值的离散点,对成熟材部分实测值的离散点拟合不好,预测曲线能够体现幼龄材材质材性变化趋势,但对成熟材材质材性变化趋势表现不足；采用BP神经网络法得到的预测值与实测值偏差小、但对部分实测值的离散点拟合不好,预测曲线对成熟材材质材性变化趋势表现不足；采用支持向量机法得到的预测值能够拟合实测值的离散点,对成熟材部分实测值的离散点拟合不好,预测曲线能够体现生长轮材质材性整体变化趋势,但对成熟材材质材性局部上下波动变化表现不足。(5)根据幼龄材材质材性预测成熟材材质材性,回归方程法操作简便、预测精准度属中下等、回归拟合结果不够理想；时间序列法步骤多、操作复杂,预测精准度属中等、回归拟合结果比较好；BP神经网络法操作简便、预测精准度属中上等,只能得到成熟材预测趋势,不能得到整体预测趋势；支持向量机法操作简便、预测精准度属中上等,预测泛化能力强,回归拟合能力强,对变异规律性不强的材质材性指标进行预测时,也能得到较低的预测相对误差和标准差。(6)核桃楸木纤维长度、生长轮密度、木材基本密度、抗弯强度和顺纹抗压强度间存在很高的相关性,相关系数R大于0.9310；水曲柳木纤维长度、木纤维胞腔直径、胞壁率、生长速率、木材基本密度和抗弯强度间存在很高的相关性,相关系数R大于0.8674。核桃楸解剖特征因子间的径向变异规律主要以第7和14年为界,分两部分变化,物理力学特征因子间的径向变异规律不显著,大致以距髓心处的第4块试材为界,分两部分变化；水曲柳解剖特征因子间的径向变异规律主要以第11和20年为界,分两部分变化,物理力学特征因子间的径向变异规律不显著,大致以距髓心处的第4-5块试材为界,分两部分变化。更多还原

【Abstract】 In recent years, in spite of the increasing amount of forest resources, the high quality wood with large diameter grade is decreased. Implementation of natural forest protection project has concentrated the wood resources supply pressure on planted forest. However, the planted forest was puzzled by the low production rate, poor quality and unreasonable forest age distribution. Therefore, forest cultivation and wood science researchers from the whole world have put their eyes on high quality planted forest breeding. Researches on predictive models of wood characteristics of planted forest will contribute to the reasonable time to fell planted forest, the update selection of natural forest, and the economical, efficient and reasonable utilization.The paper had focused on wood properties and physical and mechanical characteristics between growth rings of walnut (Juglans mandshurica Max.) and ash (Fraxinus mandshurica Rupr.) plantation. There are three main aspects:(1) On the basis of radial variation rules on wood properties between growth rings, sequential clustering optimal segmentation, principal component clustering, BP neural network and support vector machine(SVM) methods were used respectively for the demarcation of juvenile and mature wood. A comparison of the results, including characteristic and accuracy, were analyzed and confirmed.(2) After demarcated of the juvenile and mature wood, prediction methods of the regression equation, time series, BP neural network and support vector machine(SVM) were compared on the relative deviation and standard deviation of mature period prediction and the whole period prediction respectively based on the prediction of mature wood properties from the juvenile wood properties. The characteristic and accuracy of each predicted methods were analyzed.(3) After confirm that SVM method has the best regression and fitting capacity and generalization capability, firstly, relational models among wood properties’ characteristic factors were established; secondly, relational models among wood physical and mechanics characteristic factors were established; finally, relational models between wood properties and wood physical and mechanics characteristic factors were established. In core of the characteristic factors which with the correlation coefficient R more than0.83, wood characteristic models were established in the end.The main conclusions of the paper were listed below:(1) Comprehensive indexes of wood properties among the growth rings were acted as the research object. The demarcation of juvenile and mature walnut wood determined by SVM method was the18th-year. The training sets were selected mainly on the group of earlier6to 10years and the later2to6years. The demarcation of juvenile and mature ash wood was the23th-year. The training sets were selected mainly on the group of earlier10to14years and the later2to10years.(2) Walnut and ash juvenile and mature wood were demarcated with the research objuect of comprehensive indexes and single index of wood properties among the growth rings. The classification result of SVM method was primarily the same to the result of the principal component clustering and BP nueral network methods; while it was obviously different to the result of sequential clustering optimal segmentation method based on comprehensive indexes; it was primarily the same to the result of the sequential clustering optimal segmentation method based on single index.(3) In the process of juvenile and mature wood demarcation, the sequential clustering optimal segmentation method had better classification result with single index as research object than that with comprehensive indexes; BP neural network and SVM methods had better classification result with comprehensive indexes as research object than that with single index; with the comprehensive indexes of wood properties among growth rings as the research object, principal component clustering method can get the contribution from the single index of wood properties. Two pingcipal components can be used to summarize the wood properties among growth rings and the classification result can be directly showed by the graphical method.(4) In the process of mature wood properties prediction, the predicted value using the regression equation method had poor fitting effect to part of the discrete point of measured value; and the predicted curve could reflect the variation trend of juvenile wood but mature wood. The predicted value using the time series method could fit the juvenile wood discrete point of measured value, but part of the mature wood discrete point of measured value; and the predicted curve could reflect the variation trend of juvenile wood but mature wood. The deviation between predicted value and measured value of mature wood was small using the BP neural network method; but the fitting of discrete points was bad in part of the measured value. The prediction curve has poor reflection on the overall variation trend of mature wood properties. The predicted value using SVM method could fit the discrete point of measured value, but part of the mature wood discrete point of measured value; and the predicted curve could reflect the overall variation trend, while it has poor reflection on partly ups and downs of mature wood.(5) To predict the mature wood properties from juvenile wood properties, the regression equation method is easy handling, low or middle level prediction accuracy, and low fitting effect; the time series method is multi-steps, complicated handling, middle or high level prediction accuracy, and good fitting effect; the BP neural network method is easy handling, middle or high level prediction accuracy, and capacity to get mature wood predicted trend insteading of the overall trend prediction; SVM method is easy handling, middle or high level prediction accuracy, high predicted generalization capability, high fitting effect, and low predicted relative error and standard deviation when predicting from the indexes with insignificant variation relationship.(6) There was high relevance with correlation coefficient R more than0.9310among wood fiber length, density among growth rings, wood basic density, bending strength, and compressive strength parallel to grain of walnut. There was high relevance with correlation coefficient R more than0.8674among wood fiber length, wood fiber lumen diameter, wood cell walls percentage, growth rate, wood basic density and bending strength of ash wood. The radial variation rule among anatomy characteristics factors of walnut wood was separated into two parts mainly in the7th-year or the14th-year. There was insignificant radial variation rule among physical and mechanics characteristic factors, which was separated into two parts approximately in the4th specimen from the pith. The radial variation rule among anatomy characteristics factors of ash wood was separated into two parts mainly in the11th-year or the20th-year. There was insignificant radial variation rule among physical and mechanics characteristic factors, which was separated into two parts approximately in the4th to5th specimen from the pith.更多还原