基于面向对象的高分辨率遥感影像变化检测方法研究

发布时间：2018-03-14 06:19

本文选题：高分辨率　切入点：变化检测　出处：《北京师范大学》2014年硕士论文　论文类型：学位论文

【摘要】：高分辨率遥感影像成像清晰、纹理丰富、定位精准、地物内部异质性大，这些特点都使得传统的基于像元级别的变化检测方法存在不能充分利用影像信息、分类精度低、速度慢等局限性，还容易造成空间数据的大量冗余和资源的浪费，不能够很好的适应高分辨率影像的特点。因此研究高分辨率影像的变化检测方法具有较强的现实意义。本文以北京市通州区2008年10月31日和2010年11月1日的IKONOS影像为例，采用基于面向对象的方法，对高分辨率遥感影像的变化检测方法进行了相关的试验研究。首先，针对eCognition软件的多尺度分割算法，采用与邻域绝对均值差分方差比方法（Ratio of Mean Difference to Neighbors(ABS) to StandardDeviation，RMAS）方法确定了各类地物的最佳分割尺度，并对其进行了验证。然后，利用分割结果，生成面向对象的相关关系影像（Object Correlation Images,OCIs），并分别采用多变量自动提取二值化阈值方法和决策树分类方法，探究该特征影像在变化检测中的作用。结果表明，RMAS方法可以比较准确的确定各类地物的分割尺度。与传统的单变量提取二值化阈值方法相比，多变量自动提取二值化阈值方法的变化检测精度有了很大的提高。同时， OCIs影像多变量二值化的变化检测结果也要好于多变量的差值影像。其中，对于差值影像，第一波段的变化检测精度的KAPPA系数为0.7，第二波段为0.69，第三波段为0.72，第四波段为0.76，而四个波段合起来的变化检测精度的KAPPA系数达0.87；对于OCI影像，单个相关系数波段的变化检测精度的KAPPA系数为0.86，，斜率的为0.7，截距的为0.63，而三者合起来的变化检测精度KAPPA系数为0.89。基于OCIs特征影像的决策树分类方法与其他方法相比，分类结果要好。其中，基于象元的决策树分类总体精度为81%，KAPPA系数为0.71；面向对象的决策树分类总体精度为88%，KAPPA系数为0.75；而基于OCIs的最邻近分类的总体精度为87%，KAPPA系数为0.81；基于OCIs的决策树分类的总体精度为88%，KAPPA系数为0.82。基于通过以上研究发现，基于面向的变化检测方法的变化检测精度要高于基于象元的方法，另外，面向对象的方法可以有效的避免椒盐现象。
[Abstract]:High resolution remote sensing images are characterized by clear image, rich texture, accurate location and large internal heterogeneity of ground objects. These characteristics make traditional pixel level based change detection methods can not make full use of image information, and the classification accuracy is low. Because of the limitation such as slow speed, large amount of redundancy of spatial data and waste of resources, it can not adapt to the characteristics of high-resolution image. Therefore, it is of great practical significance to study the change detection method of high-resolution image. In this paper, the IKONOS images of Tongzhou District in Beijing are taken as an example, and the change detection methods of high-resolution remote sensing images are studied based on object-oriented method. In view of the multi-scale segmentation algorithm of eCognition software, the ratio of Mean Difference to neighborhood absolute difference variance ratio method is used to determine the best segmentation scale of all kinds of ground objects, and it is verified. Then, the segmentation results are used. The object Correlation images are generated, and the binary threshold method and decision tree classification method are used to explore the role of the feature image in change detection. The results show that the RMAS method can accurately determine the segmentation scale of all kinds of ground objects. The change detection accuracy of multivariable automatic binary threshold extraction method has been greatly improved. At the same time, the change detection result of multivariable binarization of OCIs image is better than that of multivariable differential image. The KAPPA coefficient of the first band is 0.7, the second band is 0.69, the third band is 0.72, the 4th band is 0.76, and the KAPPA coefficient of the four bands combined is 0.87. For OCI images, The KAPPA coefficient, slope and intercept of a single correlation coefficient band are 0.86, 0.7, 0.63, respectively, while the KAPPA coefficient of variation detection accuracy is 0.89. The decision tree classification method based on OCIs feature image is compared with other methods. The classification results are good. Among them, The overall accuracy of decision tree classification based on pixel is 81kappa coefficient is 0.71; that of object-oriented decision tree classification is 88kappa coefficient 0.75; and that of nearest neighbor classification based on OCIs is 87kappa coefficient 0.81; decision tree classification based on OCIs is decision tree classification based on OCIs. The overall accuracy of KAPPA is 0.82.Based on the results of the above study, The accuracy of change detection based on object-oriented change detection method is higher than that based on pixel method. In addition, object-oriented method can effectively avoid salt and pepper phenomenon.
【学位授予单位】：北京师范大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：P237

【参考文献】