【作者】 刘丽娟；

【导师】 范文义； 庞勇；

【作者基本信息】 东北林业大学 ， 森林经理学， 2011， 博士

【摘要】 森林资源是进行物质循环和能量交换的枢纽,具有调节气候、涵养水源、防风固沙、减少污染、保持生物多样性等多种功能,在维持生态平衡、人类生存与发展和社会进步等方面有着极为重要的作用。但是,由于人类长期过度采伐利用和破坏森林资源,致使地球生态环境遭受严重破坏。保护和发展森林资源使之可以永续利用已经得到世界上越来越多国家的广泛关注。先进的遥感技术已经逐步替代传统费时费力的地面调查工作,利用地物特有的光谱特性在遥感影像上的反映,对森林实行大面积的资源调查与监测。但是由于各方面条件的限制,目前研究还多限于星载较低空间分辨率且单一传感器的遥感数据,而对于水平和垂直结构都很复杂的北方森林而言,调查和监测工作很难做到细化。本文的研究工作正是基于机载高空间、高光谱分辨率数据和LiDAR数据相结合展开的,结合两种数据各自的特点,进行了复杂森林的树种识别、叶面积指数(LAI)及冠层叶绿素含量的森林参数反演研究。主要的内容、成果和结论为：1)基于小脚印LiDAR获得的高密度点云数据,分离地面点与非地面点,得到表征冠层高度的冠层高度模型CHM,并结合样地实测树种的树高统计,对林间空隙掩膜,去除非林地区域,减小了非林地对树种光谱的干扰,提高了影像上树种光谱与参考光谱的识别与匹配,为分类前训练样本提取做好数据准备。2)为减少噪声对光谱的影响,利用光谱微分技术将影像光谱和参考光谱进行一阶微分变换,选取代表地物特征的区间,计算两光谱的相关系数,提取相关系数高的像元光谱作为参考样本同类别的训练样本,实现了训练样本的自动提取。3)对于高空间分辨率影像中的阴影像元,考虑到传统阴影像元信息补偿方法中存在的一些问题,提出通过计算太阳入射辐射方位来确定遮挡方向,进而对分类结果中阴影像元采取邻近填充的方法,既有科学依据,又简单易实现。4)比较了仅高光谱数据和融合数据分别利用SAM和SVM分类器进行分类的结果,并得出LiDAR与CASI融合应用SVM分类器,并对分类后结果进行阴影填充的树种分类结果总体精度最高,达到86.68%,说明本文提出的对LiDAR与CASI融合数据的树种分类方法是可行的。5)基于统计模型反演LAI。由于仅利用植被指数建立与LAI的相关关系会有一定的片面性,本文除提取高光谱植被指数外,还基于LiDAR的回波数量、回波强度等信息,提取了表征森林垂直结构信息的参数,共同作为统计模型的输入自变量；根据不同的森林类型,将实测的有效LAI转为真实的LAI作为因变量,通过逐步回归进行变量筛选,建模反演LAI并进行精度验证。LAI模拟值与真实值之间的R2为0.85, RMSE为0.456,表明两者相关性较高,用该模型进行的LAI反演精度是可靠的。6)物理模型反演森林冠层生化参数涉及模型的尺度转换问题。对于第三章分类结果掩膜后的林地区域,根据树种类型,叶片模型分别选择阔叶PROSPECT和针叶LIBERTY、冠层选择SAIL辐射传输模型,并通过敏感性分析确定叶片模型输入参数的变动范围,最后模拟得到冠层的反射率。7)建立输入参数与输出反射率之间的查找表,通过影像与模拟冠层反射率的匹配查找,得到相应的叶片尺度叶绿素含量,经验证,反演叶绿素含量与实测数据R2为0.8379,满足精度要求；再利用第四章反演得到的结构参数LAI,将叶片尺度叶绿素上推至冠层,实现了森林冠层叶绿素含量的反演。更多还原

【Abstract】 Forest is key resource for material circulation and energy exchange. It has functions of climate regulation, water conservation, windbreak and sand-fixation, pollution reduction and biodiversity conservation. It also plays a very important role in maintaining ecological balance, human survival, economic development and social progress. However, the long-term over-harvesting and destruction on forest resources have resulted in serious damage to global ecological environment. Protect and develop forest resources to sustainable use have been concerned by a number of countries in world wide. The traditional time-consuming work of large-scale ground surveys has been gradually replace by remote sensing technology which using the special spectral characteristics of ground objects reflected in remote sensing images for forest resource survey and monitoring. However, due to various constraints, the present study mainly focus on the application of low spatial resolution satellite remote sensing data which with single sensor. It is a hard work to survey and monitoring the Boreal forest for its complex horizontal and vertical structure. This study is based on airborne high spatial resolution data, high spectral resolution data and LiDAR data, combined with the characteristics of each of the two data to identify the forest tree species, and inversed the leaf area index (LAI) and canopy chlorophyll content. The main contents and results include:1) Separated the ground points and non-ground points from small footprint LiDAR with high-density point data, and created a canopy height model (CHM). Then removed the gap between trees by combined the statistics of tree height measured in field plots, it reduced the interference of non-forest spectra and improved matching of the image spectra and reference spectra of tree species, took preparation for the training samples extraction of classification.2) To reduce the impact of noise on the spectrum, using spectral derivative technique to processed both imaging spectra and reference spectra by first spectral derivative transform, then selected the range of features on behalf of characteristics and calculated the correlation coefficient between two spectra, extracted the high correlation spectroscopy pixels as reference samples to achieve the automatic extraction of training samples.3) For the pixels of shadow of high spatial resolution images, the traditional methods of shadow information compensation still have some problems. In this study we identified the block direction by calculating the solar radiation orientation and then filled the pixel of shadow from its neighboring pixel. This method is both scientific and simple in the work.4) Compared the SAM and SVM classification accuracies using hyperspectral data only and the integration data, and results showed that using SVM to classify the integration data of LiDAR and CASI, and filled the shadow pixels after classification had the highest overall accuracy of classification which reached 86.68%. It indicated that the tree species classification method in this study is feasible for the integration data of LiDAR and CASI.5) The inversion of LAI based on statistical model. Since established the correlation between vegetation index and LAI still have onesidedness. In this study, the vegetation index and the vertical structure parameters were both extracted based on the echo numbers and intensities of LiDAR. These parameters were input as statistical model variables. Then the measured effective LAIs were converted into real LAIs as the dependent variable according to the different forest types and stepwise regression was used for variables selection, and the inversion model was created and validated simultaneously.6) The physical model of forest canopy biochemical parameters inversion related the problem of scaling. For the masked forest area with the classified result of Chapter III, we choose PROSPECT, LIBERTY model as broadleaf and coniferous radiative transfer model, respectively. The SAIL model was used as canopy radiative transfer model. The scope of changes was determined by analysing the sensitivity, then, output the simulated canopy reflectance.7) The lookup table was established between input parameters and output reflectances. The leaf chlorophyll contents were derived by matching the coincident images and simulated canopy reflectance. The R2 between retrieved chlorophyll and measured data was 0.8379, which satisfied the requirements accuracy. And then the inversion of forest canopy chlorophyll contents were achieved by scaled the leaf chlorophyll contents up to canopy.更多还原