【作者】 雷成亮；

【导师】 邸雪颖；

【作者基本信息】 东北林业大学 ， 生态学， 2012， 博士

【摘要】 作为森林生态系统重要组成部分的森林火灾,以从地表火到林冠火的多种形态影响着森林生态系统的树种组成、年龄结构和空间格局。它既毁灭了大量的林木,同时又对人类的生命财产以及生态环境造成了巨大的危害,影响着全球碳循环,是人类面临的最重要的自然灾害之一。我国森林火灾也非常严重,仅大兴安岭2003-2009年间,就发生森林火灾649次,过火面积75.2万公顷。本论文研究森林火灾火烈度的遥感估测方法,有助于定量评价森林火灾所造成的林木损失和林火对森林生态系统健康状况所造成的影响,从而为灾后是否进行植被恢复和选择何种森林生态系统恢复模式提供科学数据。同时对于正确估算森林火灾所造成的碳排放的多少及其对生态环境恶化的贡献大小提供技术方法与数据支持。论文选择大兴安岭南瓮河林场2006年过火区为研究区域,应用森林资源清查资料、地面实地调查资料和火烧前后两期Landsat5TM遥感数据,研究分析了基于归一化燃烧比(NBR)与燃烧比差值(dNBR)的非线与非线性模型估测火烈度的精度,并用先进的数学建模方法偏最小二乘回归、神经网络、支持向量机等进行多变量森林火烈度估测比较,为了简化模型,也比较了变量投影重要性(VIP)、正交信号修正法(OSC)与平均影响值(MIV)筛选模型的效率问题。取得了以下初步成果：目前遥感数据的应用中,存在3种对数据的处理形式：DN值、辐亮度与大气层表观反射率。利用3种形式对数据处理后提取NBR估测火烈度,得到的3个线性回归模型精度相差不多,R2从0.67-0.69。大气层表观反射率是最接近地物真实反射率的指标,是生物物理参数定量遥感的基础,因此本文的所有分析都是基于大气层表观反射率。利用单变量估测火烈度,以二次多项式模型的精度最高。在国外研究中表现较好的两个非线性模型在本文的研究中没有得到有效结果。进一步用混淆矩阵对结果的评价表明,用3种模型(线性、多项式、指数)进行火烈度估测的总体精度并不高,估测正确率还不到65%,说明还需要探索别的遥感因子和新的方法来估测火烈度。根据相关研究,选择20个遥感因子建立多变量估测火烈度模型：即支持向量分类机、广义回归神经网络和偏最小二乘模型,评价模型估测精度采用4-折交叉验证的方法。支持向量分类机的精度最高,最高估测准确率达到了85%。不同核函数的支持向量分类机的估测精度不同,径向基核函数的支持向量分类机的精度最高。当惩罚参数C取16,核函数参数gama取0.25时,径向基核函数支持向量分类机模型的精度最好。同支持向量分类机相比,广义回归神经网络模型估测火烈度的精度要逊色一点。模型除受神经网络自带的径向基传播速度参数(spread)影响以外,不同的建模数据处理方式对模型的预报精度也有影响,按[-1,1]标准化方式对建模数据进行标准化所建立的模型精度最高,当spread为0.9时模型估测精度达到80%。用偏最小二乘回归方法估测火烈度的精度最低,最高准确预报率仅有65%。使偏最小二乘回归模型预报精度最高的数据标准化方式是[0,1]标准化方式。除了偏最小二乘回归模型,多变量模型的预报精度都要比单变量多项式模型的精度高,说明单变量模型不足以概括地而火烈度的丰富信息。变量过多会导致解释与应用的困难,论文采用和比较3种变量选择方法进行模型简化。利用变量投影重要性(VIP)方法选择出10个对模型有显著影响的变量,而正交信号修正法(OSC)与平均影响值法(MW)各选出了3个,它们互有重叠。两次采用VIP方法后选出的变量所建模型预报平均正确率为61.01%,最高预报正确率为70%,均高于全变量模型。OSC方法选出的3个变量是近红外与中红外原始波段以及它们的组合变量,所建立的PLS模型使得平均与最高正确预报率均有所提高。MIV方法选出的变量所建立的GRNN模型与全变量GRNN的预报精度相当,比另两种方法所选变量建立的PLS模型精度要高,但是MIV所选变量建立的PLS模型精度很底,说明MIV方法对神经网络模型有依赖性。也说明GRNN模型有较好的鲁棒性,能正确挖掘变量间的线性与非线性关系,而PLS回归模型只变量间的线性相关关系敏感。总之,森林火烈度可以通过遥感数据进行估测,多变量模型由于充分利用卫星平台的多光谱信息而优于单变量模型,其中支持向量机、神经网络与偏最小二乘回归模型都是多变量模型中不错的选择,不仅节约灾后实地调查的时间和劳动强度,也将为全面正确评估森林火灾损失提供准确的实时基础数据。更多还原

【Abstract】 Forest fire, as an important part of forest ecosystem, has an critical effect on species composition, age structure and spatial pattern of forest ecosystem with a variety of forms from surface fire to canopy fire. It destroyed a large amount of trees, caused tremendous harm to human life, property and the environment at the same time. Its affecting on the global carbon cycle is one of the most serious natural disasters faced by the mankind. Forest fire is also very severe in China. Merely in Da Hinggan Mountains district,649forest fires occurred from2003to2009, with a burned area of752,000hectares. Remote sensing estimation of burn severity was studied in this paper, which would contribute to the quantitative evaluation of tree losses caused by forest fires and impact of forest fires on forest ecosystem health, so that to provide scientific basis for post-disaster vegetation recovery and selection on proper forest ecosystem recovery mode. At the same time, it would provide technical methods and supportive data for the correct estimation of carbon emissions caused by forest fire and its contribution to ecological deterioration.Burned district of2006in South Urn River forestry center of Da Hinggan Mountains was selected as the study area. With the application of forest resource inventory data, the ground field survey data and Landsat5TM remote sensing data of the area before and after the fire, the accuracy of fire intensity estimation from non-linear and linear model based on the normalized burn ratio (NBR) and delta normalized burn ratio (dNBR) was tested. Multivariate comparison of burn severity was made through advanced mathematical modeling, partial least squares regression, neural networks and vector supporting machines. In order to simplify the model, the efficiency of the screening model was also compared, such as the importance of the variable projection (VIP), and orthogonal signal correction (OSC) and mean impact value (MIV). Preliminary results were as follow:While using a single variable to estimate the burn severity, the accuracy of the quadratic polynomial model was the highest. The two nonlinear models performed better in abroad study did not get effective results. Further evaluation of the results by confusion matrix showed that the three models (linear, polynomial, exponential) didn’t have a high accuracy in estimating the overall accuracy burn severity, with the estimation accuracy rate of less than65%. It reveals that other remote sensing factors and new methods also need to be explored to estimate the burn severity.There are three kinds of data processing forms in applications of remote sensing data:DN value, radiant brightness and atmosphere apparent reflectance., three linear regression model, through using three kinds of data processing then extracting NBR to estimate burn severity, had similar accuracy, R2from0.67to0.69. But in inspection of the regression coefficients and constant term, there was a significant difference between every two of the three models. The atmosphere apparent reflectance is closest to the perpendicular incidence reflectivity of ground objects, and is the basic of quantitative remote sensing of biophysical parameters foundation.Building multivariable model to estimate burn severity was based on20remote sensing factors. A4-fold cross-validation method was used for evaluation model estimation accuracy. Vector supporting machine had the highest accuracy, with the highest estimation accuracy rate of85%. The estimation accuracy of different kernel function vector supporting machine was different, and the accuracy of the RBF kernel function vector supporting machine was the highest. When the penalty parameter C was16and the kernel function parameter gama was0.25, RBF kernel function vector supporting classifier machine had the highest accuracy. The generalized regression neural network model to estimate the accuracy of fire intensity was inferior while comparing with the vector supporting classifier machine. In addition to the effect from neural network that came with radial velocity of propagation parameters(spread), model accuracy was also impacted by different modeling approach. And modeling data standardization by [-1,1] standardized way gave the model the highest accuracy, model estimation accuracy was80%when the spread was0.9. Estimating the accuracy of burn severity with partial least squares regression method had the minimum accuracy, with a maximum accuracy of forecasting is only65%. The standardizing way to make partial least squares regression model predict with the highest accurate data was [0,1] standardizing way.10variables were selected by VIP method, while3were selected by both the OSC and MIV with each having overlaps. Average accuracy rate of61.01percent were worked out from two selected variables with VIP method, and the highest prediction accuracy was70%which was higher than the full-variable model. The three variables selected by the OSC method was the original band of near-infrared and mid-infrared as well as variables of their combination. The established PLS model increased the average and the highest correct prediction rate. GRNN model established with variables selected by the MIV method and all variables GRNN had a same prediction accuracy, which was higher than the PLS model accuracy established by the other two methods selected variables. But PLS model established with variables selected by MIV had a low accuracy, indicating MIV method as quite depending on the model. It also revealed that GRNN model was robust, and could excavate linear and nonlinear relationships between the variables correctly. While at the same time, the PLS regression model was only sensitive to the linear relationship between the variables.In short, forest fire intensity can be estimated by remote sensing data, especially, multi-variable models perform better than single variable model, and vector supporting machines, neural networks and partial least squares regression model are all good choices of multi- variable models. They do not only save investigating time and labor intensity of post-disaster field surveys, but also provide accurate real-time basic data for comprehensive assessment of the correct loss of forest fire.更多还原