【作者】 骆成凤；

【导师】 王长耀；

【作者基本信息】 中国科学院研究生院（遥感应用研究所） ， 地图学与地理信息系统， 2005， 博士

【摘要】 土地覆盖和土地覆盖的自然变更,在地球系统的气候和生物化学全球尺度模式中扮演着重要角色,对全球生态环境产生着巨大的影响。地表土地覆盖信息是全球气候和陆地水文地理模型的基本输入。遥感为获取全球性和区域性土地覆盖信息提供了有效的技术手段。我国幅员辽阔,地表土地覆盖分布状况复杂多样,因此对我国地表土地覆盖状况进行研究对全球环境变化研究具有重要意义。 本文以MODIS植被指数和陆地表面温度为主要遥感数据源,进行了全国地表土地覆盖信息提取研究。首先提出基于NDVI-Ts特征空间的温度植被角度和距离(TVA&TVD)作为表征地表土地覆盖状况的生物物理参数。选择TVA和TVD作为主要分类特征,并用主成分分析(PCA)提取主要信息。用人工神经元网络算法进行大尺度低分辨率的全国土地覆盖分类研究,提出基于GA优化的BP神经网络遥感数据分类算法,用其对全国土地覆盖进行分类,并将该算法与模糊ARTMAP算法和最大似然法分类结果进行分析比较。本文还利用变化矢量方法对地表植被和土地覆盖状况进行变化监测研究。 本文的方法创新点归纳为: 1) 提出表征地表土地覆盖信息的生物物理参数:温度植被角度和距离(TVA&TVD)。 不同的地表土地覆盖类型在NDVI-Ts特征空间中空间分布差异非常明显,基于NDVI-Ts特征空间的TVA和TVD,能够唯一表达不同地表土地覆盖类别在NDVI-Ts特征空间中的位置。无论从信息含量的角度,还是不同类别的年内点轨迹,TVA和TVD对不同地表土地覆盖类别差异的体现都要优于NDVI和Ts。 2) 选择和提取了一组适合全国土地覆盖分类的特征 选择TVA和TVD时间序列数据、空间结构信息、数字高程作为分类特征,并用主成分变换(PCA)对TVA和TVD时间序列数据进行特征提取,得到既利于区分不同类别又含有主要信息的特征作为分类输入参数。为了比较,选择NDVI和Ts时间序列数据、空间结构信息、数字高程作为分类特征,经过同样的特征提取,得到另外一组输入参数。用相同的样本和分类算法对全国土地覆盖进行分类,结果表明:以TVA和TVD为主要参数提取的分类特征,其全国土地覆盖分类结果要优于以NDVI和TS为主要参数提取的分类特征。 3) 用遗传算法(GA)对反向传播神经网络算法(BP)进行优化,并将其用于大尺度低分辨率土地覆盖分类应用中更多还原

【Abstract】 Land cover and its natural alteration, which play an important role in climate system and biochemistry pattern in globe scale, have brought tremendous effect to global entironment. The land cover information of the earth’s surface is the primary input parameters for globe climate and terrestrial hydrography models. Remote sensing offered an effective method for acquiring surface land cover information. China is a country with a vast territory and diversiform, complicated surface land cover; upon that the research about Chinese surface land cover condition is one of the import parts of globe environment change.The article studied acquiring Chinese surface land cover information which the 1km MODIS NDVI&Ts data product which were used as main remote sensing data. Temperature-Vegetation Angel and Temperature-Vegetation Distance (TVA&TVD) based on NDVI-Ts space were proposed as biophysical parameters which can be thought as a token of surface land cover condition. TVA&TVD were selected as main classification features, and Principal Component Analysis was used to extract the primary information. Artificial Neural Network (ANN) algorithm was adopted to identify the Chinese surface land cover types at coarse spatial scales. One side, BP neural network classifier based on genetic algorithm for remotely-sensed data was proposed; on the other hand, classification of Chinese surface land cover was made with the algorithm, and other algorithms: fuzzy ARTMAP ANN and Maximum likelihood classifier (MLC), subsequently the three algorithms and their classification results were compared and analyzed. The surface vegetation and land cover condition detection using Change-Vector analysis is another important part of the article.Main research results and initiatives in this thesis include as following:1. Two index were proposed as biophysical parameters : TVA&TVDThere is an obvious diversity for distribution of different surface land cover classes in NDVI-Ts space, and TVA&TVD based on NDVI-Ts space can locate the spacial distribution uniquely. TVA&TVD can be in the person of different surface land cover classes much better than NDVI &Ts, which can be demonstrated by the information entropy and the different change trajectory per annum of each class.2. Several suitable features for Chinese surface land cover classification were selected and extracted.After TVA&TVD temporal series data, spacial structure information and Chinese digital elevation model were selected classification features, Primary Components Analysis (PCA) was更多还原