面向对象高分辨率遥感数据滑坡灾害信息提取研究

发布时间：2018-01-28 19:15

本文关键词： 面向对象高分辨率遥感信息提取滑坡变化检测 eCognition　出处：《兰州大学》2015年硕士论文　论文类型：学位论文

【摘要】：2008年汶川“5.12”地震不仅影响了四川省大部分地区,也对甘肃南部地区产生了很大影响,表现为由地震引发的滑坡、泥石流、崩塌等次生地质灾害急剧增多,亟需对地质灾害频发的甘肃南部地区进行滑坡灾害调查,为做好滑坡的防御减灾工作奠定基础。为了能够第一时间获取滑坡的相关灾害信息,如滑坡的性质、规模、分布、对桥梁、道路、建筑物等的影响,提出用高分辨率遥感数据使用面向对象的方法进行滑坡体灾害信息的提取。面向对象分析技术已在遥感领域的信息提取方面应用广泛,区别于基于像元的方法,分类过程更加符合人类认知原理,用于分类的特征不仅包含光谱信息,还包含了纹理形状、空间结构、上下文语义、拓扑关系等信息,避免了分类结果过于破碎,减弱了“椒盐”噪声的影响,分类精度有了明显的改善。选用2007年11月IKONOS数据(全色波段的分辨率为lm)和2013年4月QuickBird影像(全色波段的分辨率为0.61m)两个时相的高分遥感数据作为滑坡灾害信息提取的遥感数据源,结合甘肃陇南地区的分辨率为30m的数字高程模型(DEM),采用易康软件的面向对象分析方法,以分布有大量滑坡的甘肃南部地区为研究区,进行滑坡灾害信息的半自动提取研究。主要结论如下：(1)采用多尺度分割算法分别对建筑物、道路、植被和滑坡体设置不同的分割参数进行多次分割实验,得到最优分割参数并分析遥感数据的光谱特征、形状特征、纹理特征等信息,得出最优特征空间。(2)利用震前震后两个时相的遥感数据获得的建筑物信息提取结果进行变化检测得到建筑物的变化信息,间接反映地质灾害引发的建筑物损毁情况。(3)根据震后遥感数据受滑坡威胁的道路和植被的属性特征,利用面向对象模糊分类法和阈值分类法建立规则级进行信息提取,为滑坡易损性快速评估提供技术支持。(4)采用面向对象支持向量机方法,选取合适的光谱、形状、纹理等特征建立特征空间,通过径向基核函数对选定的200个滑坡样本和300个非滑坡样本进行机器学习训练,最终获得滑坡体提取结果。对滑坡灾害信息的各提取结果选用混淆矩阵定量评价方法进行精度评价,滑坡灾害信息的提取总精度达到85%以上。综上所述,面向对象的分类方法适用于滑坡灾害信息提取,提取结果与真实情况相符,也满足实际应用需求,为滑坡灾情及承灾信息的迅速评估提供依据。
[Abstract]:The 2008 Wenchuan "5.12" earthquake not only affected most areas of Sichuan Province, but also had a great impact on the southern part of Gansu Province, which was manifested by landslides and debris flows caused by the earthquake. The number of secondary geological disasters such as collapses is increasing rapidly, and it is urgent to investigate the landslide disasters in the southern part of Gansu, where geological disasters occur frequently. In order to be able to obtain the landslide disaster information in the first time, such as the nature, scale, distribution of landslides, the impact on bridges, roads, buildings and so on. An object oriented method is proposed to extract landslide disaster information using high resolution remote sensing data. Object-Oriented Analysis (OOA) technology has been widely used in the field of remote sensing information extraction, which is different from the pixel based method. The classification process is more in line with human cognitive principles. The features used for classification not only contain spectral information, but also contain texture, spatial structure, context semantics, topology and other information. The classification results are not too broken and the effect of "salt and pepper" noise is weakened. The classification accuracy has been significantly improved. The IKONOS data of November 2007 (resolution in panchromatic band) and QuickBird images of April 2013 (with a resolution of lm) were selected. The resolution of panchromatic band is 0.61m) the high-score remote sensing data of two phases are used as the remote sensing data source of landslide disaster information extraction. Combined with the digital elevation model with a resolution of 30 m in Longnan region of Gansu Province, the object oriented analysis method of Yikang software is adopted, and the study area is the southern part of Gansu, where a large number of landslides are distributed. Research on semi-automatic extraction of landslide disaster information. The main conclusions are as follows: (1) Multi-scale segmentation algorithm is used to set different segmentation parameters of buildings, roads, vegetation and landslide for many times. The optimal segmentation parameters are obtained and the spectral features, shape features, texture features of remote sensing data are analyzed. Get the optimal feature space. 2) using the remote sensing data of the two phases before and after the earthquake to extract the building information to detect the change information of the building. Indirectly reflects the geological hazards caused by the damage of buildings. 3) according to the remote sensing data after the earthquake landslide threatened by the road and vegetation attributes. Object oriented fuzzy classification and threshold classification are used to establish rule level for information extraction, which provides technical support for rapid assessment of landslide vulnerability. (4) Object-Oriented support Vector Machine (OSVM) method is used. Appropriate spectral, shape, texture and other features are selected to establish the feature space, and then 200 landslide samples and 300 non-landslide samples are trained by radial basis function (RBF). Finally, the results of landslide body extraction were obtained. The accuracy of landslide information extraction was evaluated by the method of confusion matrix quantitative evaluation. The total accuracy of landslide disaster information extraction was more than 85%. To sum up, the total accuracy of landslide disaster information extraction was more than 85%. The object-oriented classification method is suitable for the landslide disaster information extraction. The extraction results are consistent with the real situation and meet the practical application needs. It provides the basis for the rapid evaluation of landslide disaster situation and disaster information.
【学位授予单位】：兰州大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：P642.22;P237

【参考文献】