【作者】 郑毅；

【导师】 武法东；

【作者基本信息】 中国地质大学（北京） ， 地学信息工程， 2010， 博士

【摘要】 遥感影像周期性、现势性、客观性和系统性特点使其成为土地利用信息获取的重要途径,但由于土地利用/土地覆盖的复杂性,一直以来多是通过目视解译方式从遥感影像上获取地类信息。显然这种方式是不能满足人类快速、准确获取地类信息的要求,虽然在计算机自动解译或半自动解译方面已有众多研究成果,但投入实际应用的方法并不多,用于大区域范围的就更少见报道。对于遥感影像分类来说,主要有两种基本方法:基于像元和基于对象的方法。大多数基于像元分类方法能被专门聚类分析方法所包含。如,借助于统计方法,借助于模糊逻辑或借助于神经网络且被指定为一类,被分类的特征通常是像元的光谱特性。基于像元的分类方法是迄今为止最为普遍的遥感影像分类方法。近年来不论如何发展,这些纯像元方法已经到了极限。基于对象分析的方法提供了更大的潜力,由于对于分类它有庞大的特征要素(形状、尺寸、纹理和邻域关系),容易集成来自于其他数据源的附加信息并可进行分析。正因基于对象分析的极大潜力,近年在此方向研究工作发展迅速。本研究在全面、系统地分析、研究国内外面向对象分类法在遥感影像地类信息获取研究成果的基础上,提出了自动特征分析方法——SaT法,并构建起了面向对象分析法+SaT方法的遥感影像地类信息获取新方法。该方法能较好地解决面向对象分类法中特征选取速度慢,易漏选最佳特征值等问题。应用此方法,采用SPOT5全色影像对剑川县2238平方公里土地的地类进行了实证研究应用,对所拟定地类进行分类处理,分类总精度92.85%,Kappa系数0.91,同时可获得90%或更高的生产精度和用户精度。对比研究表明其分类精度远远高于通常使用的:⑴最大似然法;⑵含纹理信息的最大似然法;⑶人工神经网络法。实际成果表明面向对象分析法+SaT方法具有较好的应用前景,特别是特征分析方法——SaT具有较强的特征选择能力。更多还原

【Abstract】 Remote Sensing Images is an important approach of obtaining information of land use due to its high-resolution, easy-obtaining and economy. Because of the complexity of land use/land cover, human had to obtain information of land use by visual interpretation. By all appearances the approach is not able to satisfy the requirement of obtaining information fast and exactly. Although human had obtained many research achievements of computer interpretation, few of them are applied to practice.The classification of Remote Sensing images includes two basic methods: the pixel-based and the object-based method. Most pixel-based classification methods can be subsumed under the cluster analysis. For example, with the aid of statistical methods, with fuzzy-logic techniques or with neural networks, them can be assigned to one class. The features to be classified are generally the spectral signatures of pixel. Pixel-based methods are hitherto the most commonly used type of classification in remote sensing. In recent years, these purely pixel-based methods have increasingly reached their limits. Object-based image analysis shows greater potential, because it has a very large feature basis (shape, size, texture and neighbourhood) for classification, and can integrate additional data from other data sources easily. Just because of the great potential of object-based analysis, recent research work increasingly goes in this direction.Based on comprehensive study on a large number of research results of object-oriented classification of Remote Sensing images, this study puts forward a new automatic feature analysis method-SaT method. This method can solve the problems of slow speed of feature selecting and missing optimal eigenvalues. This segmentation object-oriented processing classifier has been applied to land-cover classification for Jianchuan County of 2238Km2. This classifier outperforms an overall accuracy of 92.75% and Kappa of 0.91, and can get producer accuracy and costumer accuracy of 90% or higher. This approach excel common classifiers, such as :⑴the maximum likelihood classifier;⑵the maximum likelihood classifier with texture analysis;⑶artificial neural network. Study result shows that the object-oriented analysis + SaT has good application perspective.更多还原