【作者】 岳彩荣；

【导师】 赵宪文；

【作者基本信息】 北京林业大学 ， 森林经理学， 2012， 博士

【摘要】 近年来,在全球气候变暖的背景下,森林碳储量研究成为了人们关注的焦点。而森林生物量的估算是进行陆地生态系统碳循环和碳储量及变化分析的基础,也是生态学和全球变化研究的重要内容之一。香格里拉县地处横断山脉,位于世界面积最大的滇西北“三江并流”世界自然遗产保护区腹地。研究这一区域森林植被的生物量/碳储量,对进一步开展高原陆地生态系统碳循环研究,科学评价自然遗产保护区的生态服务功能及其在全球气候变化中的作用,提高全社会对保护区生态价值的认识都具有重要意义。随着3S技术的出现和发展,为快速、无破坏地进行较大尺度森林生物量和碳储量的研究提供了可能。本文以香格里拉县为研究区,以2008、2009年陆地卫星数据TM为信息源,结合2008的森林资源二类调查数据和本文研究中调查的186个样地数据,在国家基金项目“滇西北三江并流区森林生态系统碳储量遥感定量研究”(No：40861009)支持下开展了香格里拉森林生物量和碳储量的遥感估测研究,建立了香格里拉县4种主要森林类型的遥感生物量估测模型。利用样地调查方法按4个森林类型估算了林下灌木、草本和枯枝落叶层的生物量,与遥感模型估测的乔木层生物量汇总,完成了香格里拉县主要森林生态系统的森林生物量和碳储量的估测。由于研究地区地处高山峡谷区,地形深切割,高差大(从1503米-5545米),导致遥感数据的预处理和标准化难度大。针对此情况,论文重点在遥感数据的地形阴影去除、森林生物量遥感特征提取、森林生物量遥感建模方法等方面重点开展研究。主要研究结论如下：(1)高山峡谷区遥感数据受地形影响较大,而本文采用的坡度匹配地形辐射校正方法较之常用的方法能更好地减小地形对遥感数据的影响,改善了遥感数据与森林蓄积量之间的相关性。研究表明：仅采用辐射定标处理的各波段数据与森林蓄积量的相关性都很低,TM的6个波段与蓄积量相关都未通过显著性检验；经过大气校正后相关性有了部分改善,但仅有TM4的相关达到p=0.05显著水平；在大气校正的基础上进行的地形校正明显地改善了各波段与样地生物量之间的相关关系。通过0.05显著性检验的波段达到了3个波段,其中TM4的相关系数通过了0.01水平的显著性检验。(2)与传统的遥感特征因子相比,本文提出的通过像元分解获取的森林阴影丰度特征与森林生物量的具有较好的相关性,并在3个树种的逐步回归建模中入选回归方程。研究表明,像元分解后的阴影丰度图像与森林生物量的相关性较之单波段的相关性都有不同程度的提高,除云冷杉外,相关性都通过了0.01水平的显著性检验。表明像元分解得出的阴影丰度特征可以作为森林生物量遥感估测的重要因子。(3)三种建模方法的结果表明,线性回归模型精度最低,神经网络模型精度次之,支持向量机回归模型(SVM)精度最高。利用相同的的建模样本数据和检验数据,采用多元线性回归、神经网络和支持向量机方法进行了森林生物量遥感估测模型的建模实验。检验结果表明,多元线性回归的检验精度最低,神经网络次之,支持向量机精度最高,除云冷杉外,模型预测值的平均相对误差小于20%。多元回归由于要求数据分布满足一些假设条件(如线性、正态性、等方差、独立性等),使模型中得以保留下的变量很少(2-3个),造成了遥感信息利用不充分,导致模型的估测效果较差。神经网络方法可以得到小的训练误差,但对于建模外的新数据,其泛化能力较差,存在过学习问题,从而导致模型的估测精度不高。支持向量机不同于神经网络和线性回归等传统方法以训练误差最小化作为优化目标,而是以训练误差作为优化问题的约束条件,以置信范围值最小化作为优化目标。因此,SVM的泛化能力要明显优越于神经网络等传统学习方法。另外,SVM的求解最后转化成二次规划问题的求解,因此,SVM的解是唯一的、也是全局最优的。本文的研究结果中支持向量机模型精度最高也充分证明了SVM理论的正确和有效性。(4)研究区主要森林生态系统的生物量和碳储量利用支持向量机回归模型估算了香格里拉县主要森林生态系统的乔木层生物量,利用样地调查法估算了林下层生物量和土壤碳,得到香格里拉县主要森林生态系统的总碳储量为302.984 TgC,其中,乔木层、灌木层、草本层、枯落物层和土壤层的碳储量分别为60.196 TgC、5.433 TgC、1.080 TgC、3.582 TgC、232.692 TgC,分别占总碳储量的19.87%、1.79%、0.36%、1.18%和76.80%,各层按照碳储量大小排序为：土壤层>乔木层>灌木层>枯落物层>草本层。不同的森林类型生态系统的碳储量差异较大。研究区主要的森林生态系统按照碳储量大小排序为：云冷杉>栎类>高山松>云南松。(5)研究区主要森林生态系统碳密度香格里拉县主要森林生态系统的平均碳密度为403.480 t/hm2。其中,云冷杉林的碳密度最大,为576.889 t/hm2,其次是栎类,为326.947 t/hm2,高山松和云南松分别为279.993 t/hm2、255.792 t/hm2。香格里拉县的平均碳密度403.480 t/hm2要高于临近地区估算的四川森林生态系统平均碳密度为232.81 t/hm2。这说明,位于世界最大的自然遗产保护区的香格里拉除了具有重要的自然、人文和生物多样性的保护价值外,还是该地区的一个重要碳库,对整个区域生态系统的稳定和平衡发挥着重要的作用。更多还原

【Abstract】 Recently, the research on forest carbon storage attracted more and more attention against the global warming. As the basis of analyzing terrestrial carbon cycle and storage and its dynamic change, forest biomass estimation has become one of the important contents of ecology and global change study. Shangri-La county, located in the alpine and gorge region in the northwest of Yunnan province, is laid at the hinterland of the largest world natural heritage site "Three Parallel Rivers". The forest biomass and carbon storage research at this region plays a significant role for developing plateau terrestrial carbon cycle study, evaluating the ecological services of natural heritage and its role in global climate change scientifically and improving the whole society awareness of the ecological value of protected areas. As the development of GIS, RS and GPS techniques, a rapid and effective research can be provided for studying forest biomass and carbon storage in large scale.In this study, Shangri-La was chosen as study area. The research on remote sensing estimation of forest biomass and carbon storage was carried out, which was supported by National Natural Science Foundation project (No:40861009). The Landsat data TM of 2008 and 2009, the forest management inventory of 2008 and 186 supplemental field samples were integrated. The study area is located in the high-mountain gorge area, the complex terrain (elevation from 1503m to 5545m) resulted in remote sensing data pre-processing and standardizing very difficult. For this case, the RS data terrain shadow elimination, forest biomass RS feature extraction and modeling means was focused on in this study. A forest biomass RS estimation model was built. And the forest biomass and carbon storage estimation for major forest ecosystems was completed in Shangri-La. The main conclusions as below:(1) The RS data, located in the high-mountain gorge area was influenced seriously by the terrain. The correlation between RS data and forest volume was improved in this study by employing the terrain radiation correction.The study result indicated that the correlation between each band and forest volume was very low by using Radiometric Calibration only. The six bands of TM data and forest volume were failed to significance test. After atmospheric correction, the correlation was improved in some degree. But there were also only the TM4 correlation was significant (p=0.05). By using terrain correction, the correlation between the bands and sample biomass was improved obviously based on atmosphere correction. There were three bands correlation had passed the significance test. What is more, the correlation coefficient of TM4 passed the significant test (p=0.01).(2) Compared with numerous traditional RS features, the shadow percent of forest obtained by pixel unmixing had represented much better correlation with forest biomass than the most other ones among the RS features in the study.The study result indicated that the correlation between shadow richness image and forest biomass after pixel unmixing was better in some degree than the correlation of single band. All correlation had passed the significance test (p=0.01) except Spruce-fir. It showed that the shadow feature obtained by pixel unmixing could be a significant factor in the forest biomass estimation based on RS.(3) The result of three model means showed that the accuracy of the linear regression model was the worst, the neural network was followed, and the support vector machine regression model was the highest. With the same sample data set, biomass estimation model was performed by using multiple linear regression, neural networks and support vector machine approach for different forest types, respectively. The results showed that multiple linear regressions leaded to estimates of the model less effective. It was because that the data distribution was required to meet some hypothesis (e.g. linearity, normality, equal variance, independence, etc.), which caused only a few variables (2-3) left in the model equation, and then it resulted in the insufficient use of remote sensing information.The neural network can obtain a smaller training error, for the new and untrained data, however, its generalization ability was poor and there were some learning problems. Consequently, its model estimation accuracy was not high.Unlike traditional means such as neural networks to minimize the training error, the Support Vector Machine made the training error as an optimization constraint, and minimized the value of the confidence range as its optimization objective. The generalization ability of SVM was superior to the traditional study means. Moreover, the SVM solving was transform into quadratic programming solving. Therefore, the SVM solution was unique and global optimal one. In this study results, the correctness and effectiveness of SVM theory was fully demonstrated by that Support Vector Machine model had the highest accuracy.(4) Dominant forest ecosystems carbon storage of the study areaWith the help of Support Vector Machine Model, carbon storage of Shangri-La can be estimated, the total carbon storage of Shangri-La is 302.984 TgC. The carbon storage of arbor Layer, Shrub Layer, Herb Layer, Litter Layer and Soil Layer is 60.196 TgC、5.433 TgC、1.080 TgC、3.582 TgC and 232.692 TgC respectively, and it accounts for 19.87%, 1.79%,0.36%,1.18% and 76.80% of total carbon storage respectively. According to the amount of their carbon storage, the sequence was Soil Layer>Tree Layer>Shrub Layer> Litter Layer>Herb Layer.(5) Dominant forest ecosystem carbon density of the study areaThe average carbon density of Shangri-La dominant forest ecosystem is 403.480 t/hm2. Spruce-fir has the greatest carbon density (576.889 t/hm2). The follow are Oak with 326.947 t/hm2. Alpine larch with 279.993 t/hm2 and Yunnan Pine 255.792 t/hm2 respectively. The average carbon density of Shangri-La was higher than that of its neighboring Sichuan area, which is estimated to be 232.81 t/hm2. It is demonstrated that Shangri-La, locating in the greatest world natural heritage site, was not only significant in its nature, humanities and biological value, but also a vital carbon pool, which plays an important role in maintaining stability and balance for the whole ecosystem of Shangri-La county.更多还原