融合地物语义的多尺度对象马尔可夫模型的高分辨率遥感影像分割研究

项目名称： 融合地物语义的多尺度对象马尔可夫模型的高分辨率遥感影像分割研究

项目编号： No.41201463

项目类型： 青年科学基金项目

立项/批准年度： 2013

项目学科： 地理学

项目作者： 洪亮

作者单位： 云南师范大学

项目金额： 25万元

中文摘要： 遥感影像空间分辨率的显著提高，使得遥感影像具有更丰富的地物细节信息，也使得传统的遥感影像解译方法面临巨大挑战。针对高空间分辨率遥感影像地物语义层次分明、随机性强、纹理和几何结构特征明显的特点，以影像对象作为基本处理单元，将层次地物语义信息融合到多尺度马尔可夫随机场（MRF）理论框架中，研究高分辨率遥感影像解译问题。首先，采用多特征TurboPixels算法对高分辨率遥感影像进行区域过分割生成基本处理单元-影像对象；结合不同层次地物的语义、光谱及空间特征，有区别的建立多尺度合并规则，建立与地物语义层次对应的多尺度对象邻接图；最后，在多尺度对象邻接图上，基于多尺度MRF模型，研究融合多尺度特征与多层次语言信息的高分辨率遥感影像解译方法。在上述基础上，研究基于多尺度对象MRF模型的影像信息提取算法，为高分辨率遥感影像的自动解译提供一条新的思路。

中文关键词： 高分辨率遥感影像；影像分割；马尔可夫模型；地物语义层次模型；多尺度对象邻接图

英文摘要： Remote sensing image is experiencing remarkable improvement in spatial resolution, which makes much more detail of land observed, at the same time, brings a huge challenge to conventional remote sensing image interpretation methodology. High spatial resolution remote sensing images are characterized by hierarchical semantic structure, strong randomicity, and being rich in texture and geometric structure information. In order to benefit high spatial resolution remote sensing image interpretation, image objects are used as processing units, hierarchical semantic information is fused into multiscale markov random field (MRF) theory framework and a multiscale object MRF based method for information extraction in high resolution image is explored. Firstly, the multi-feature based TurboPixels segmentation algorithm is employed to generate basic analysis units-image objects. Then, according to merging rules defined by combinations of semantic information in every level and the spectral and spatial feature , a hierarchical image object adjacency graph, which matches the hierarchical land semantic object, is established level by level. Finally, the multiscale MRF model is operated on the multiscale object adjacency graph and an image interpretation method fusing multiscale features and multilevel land semantic informati

英文关键词： High Resolution Remote Sensing Image；Image Segmentation；Markov Random Field；Object Semantic Hierarchical Model；Multi-Scale Object Adjacency Graph