【作者】 孙明；

【导师】 沈渭寿；

【作者基本信息】 南京信息工程大学 ， 自然地理学， 2011， 硕士

【摘要】 遥感科学技术的形成与发展,以及与全球定位系统等地理信息系统科学的融合、渗透和统一,形成了新型的对地观测系统,遥感信息在科学研究和国民经济中的应用越来越受到各行各业的重视。本文研究对象是环保公益专项“青藏高原生态退化及环境管理研究”中一个重要研究区雅鲁藏布江源区,根据野外实地勘察数据,结合可用的遥感影像数据,对源区的草地类型进行遥感识别,并对植被盖度进行遥感的定量反演,为源区生态退化研究服务。本文选取成像质量较好的2009年8月1日左右的Landsat5 TM影像,根据不同草地类型的波段组合特征,结合源区1：100万植被类型图、DEM和NDVI数据,构建草地识别的判别规则,利用决策树分类的方法对雅鲁藏布江源区草地类型进行遥感识别研究；以TM派生数据NDVI、RVI、VI3、PVI、DVI、MSAVI、SAVI、TM4/TM5为主要分析因子,结合野外植被样方调查数据,选取相关性最高的因子与实测植被盖度建立回归模型,利用该模型生成植被盖度分布图。研究结果表明：(1)不同的草地由于其生境不同,受背景土壤类型和土壤水分的干扰和影响,一定程度上加大了不同草种之间的光谱可分性,利用不同波段组合特征进行草地类型识别能够达到较好的效果；总体上,利用地物光谱信息的面向对象分类在分类效果上有了很大的提高,弥补了传统的基于像素统计特征分类方法的不足。同时光谱信息参与面向对象分类,在信息提取之前,通过野外采样数据,找出草地之间的区分规律,较光谱未参与面向对象分类的目标性强,提高了分类精度。和传统的监督分类法相比,基于波段组合特征的决策树分类法具有较高的识别精度：采样决策树分类的总体精度可以达到62.5%,采用监督分类的总体精度47.1%；决策树分类的Kappa系数0.493,监督分类的Kappa系数为0.268。和监督分类相比,决策树分类的总分类精度提高15.4%,Kappa系数提高0.225。本文采用的决策树分类法仅是基于影像的光谱特征、波段间的相互运算以及高程等信息,并没有加入其它分类特征,其分类精度不是特别高；若在以后的分类决策树模型中加入纹理等信息,则决策树分类法的优势会更明显,分类精度会更高。(2)草地盖度实测值与对应的TM4/TM5值变化趋势基本相同,且TM4/TM5增强了不同退化程度草地植被的光谱反射值的差异,以TM4/TM5为因子构建的草地盖度估测模型能够准确反映出源区草地盖度的基本分布趋势,模型的整体预测精度较高：RMSE为0.074,相对误差为19.6%,草地植被盖度模型验证精度达到0.91,达到模型验证精度要求。更多还原

【Abstract】 A new earth observation system, which is formed by remote sensing system, global positioning system and geographic information system, has been paid attention to by all industries in the scientific research and application in the national economy.The study area of this paper is the source region of Yarlung Zangbo river, which is the important study area in the environmental protection special public-"Research on Ecological Degradation and Envirmental Management".According to quadrat investigating, we recognize the types of grassland and inversion the vegetation coverage, which server the research on ecological degradation in the source region of Yarlung Zangbo River.We selected the Landsat5 TM images of high quality in August 1st,2009. According to different features of spectral combination, we build the recognition rules of grass identification with the data of 1:400million vegetation map, DEM, and NDVI. We did the research on grass recognition in the source region of Yarlung Zangbo river based on decision tree classification. This paper analyzed the NDVI、RVI、VI3、PVI、DVI、MSAVI、SAVI and TM4/TM5, combining quadrat investigating, and we select the TM4/TM5 as the main factor to construct model with vegetation coverage. We calculate the vegetation coverage of the image with this model and create the image of the distribution of grassland, which is of great important for grasping the desertification of grassland. It was shown that:(1)As a result of different habitat,it is increased the separability in some extent, inflected by soil type and moisture. We can achieve good results of remote sensing recognition of grass on spectral combination features. Using the spectral information,the classification effect of object-oriented classification has been improved greatly overall. It has made up the deficiency of traditional classification based on statistical characteristics of pixels.Compared with traditional supervised classification, the decision tree classification based on spectral combination has high precision of identification, overall classification accuracy has improved by 15.4% and Kappa coefficient has increased by 0.225. Decision tree classification used in this paper only consider the spectral characteristics of images, the inter-band operations, elevation and so on. We did not join the other categories, and theclassification accuracy is not particularly high.If we can consider texture information, the advantage of decision tree classification will be more obvious and its classification accuracy will be higher.(2)It has the same trend with the measured data of grassland coverage and TM4/TM5, which increase the separability of spectral inflection value of different degraded grassland.The grassland coverage estimation model built by TM4/TM5 as the main factor can reflect the distribution of of grassland accurately. The prediction accuracy of model is high:RMSE reaches to 0.074 and relative error is 19.6%. The verification accuracy of model reaches to 0.91 and it can meet the requairement of model validation.更多还原