基于对象模型与多点空间统计的高分辨率遥感影像分类策略

项目名称： 基于对象模型与多点空间统计的高分辨率遥感影像分类策略

项目编号： No.41501489

项目类型： 青年科学基金项目

立项/批准年度： 2016

项目学科： 天文学、地球科学

项目作者： 唐韵玮

作者单位： 中国科学院遥感与数字地球研究所

项目金额： 20万元

中文摘要： 随着地理空间信息需求的增长，高分辨率遥感的优势越发明显。高分影像蕴含了更丰富的地物细节，决定了其分类的复杂性。高分遥感一般采用面向对象的分类方法，即对有限数据支撑的网格数据的对象进行模式识别。尽管利用空间信息在像素级分类方法中已较常见，大部分面向对象的分类方法并没有充分利用这些空间信息。空间统计学尤其是多点空间统计能有效描述和量化影像对象这种特殊网格数据的空间相关性和异质性，与面向对象分类方法相结合，以提高分类精度。本课题提出利用空间统计学的理论，进行对象数据的空间上下文信息的量化与分析；理清不同数据支撑的空间信息计量方法；完善和扩展面向对象的分类方法。关键技术包括：影像分割与对象网格形成，基于空间统计的对象空间信息量化，影像对象的尺度建模与尺度转换，分类器融合，精度评定的新指标等。研究成果将促进空间统计学与遥感影像分析的融合，挖掘高分遥感数据的信息潜力，从而为地学应用提供更好的服务。

中文关键词： 基于对象的影像分析；空间相似与变异；可变面元问题；训练样本选择；高空间分辨率

英文摘要： With the growing requirement of geo-spatial information, high-spatial resolution remote sensing technique demonstrates great advantages. Rich spatial details in high-spatial resolution imagery decide the complexity of the relevant classification. The object-based classification is to identify the objects in the form of lattice data with a finite data support. This method is used commonly for classifying high-spatial resolution remotely sensed imagery. Although classifiers involved with spatial information are common in pixel-based methods, most object-based classification algorithms do not fully utilize the spatial information in the imagery. Geostatistics, particularly multiple-point geostatistics, can explain and quantify spatial correlation and spatial variability of image objects, which can be seen as lattice data. It can be employed to classification methods to improve classification accuracy. This project proposes methods based on geostatistics to quantify and analyze spatial contextual information of object data, to measure spatial correlation between different data supports, and to improve and expand the current object-based classification techniques. The key techniques of the proposal include image segmentation and object-based lattice data formation, object-based spatial information quantification using geostatistics, scale modeling and transformation of image object, classifiers fusion, and the new method for accuracy assessment. The outcome of this research can promote integration of geostatistics and remote sensing, and dig more information potentials of high-spatial resolution data to better serve the related geographical applications.

英文关键词： object-based image analysis;spatial similarity and variability;modifiable areal unit problem;training sample selection;high spatial resolution