【作者】 祝锦霞；

【导师】 王珂； Thomas C.Harmon；

【作者基本信息】 浙江大学 ， 农业遥感与信息技术， 2011， 博士

【摘要】 高分辨率遥感影像包含更多、更丰富的地理和地形信息,已经广泛应用到社会经济的很多领域。尽管如此,高分辨率遥感影像在变化检测中仍面临很多问题：地物内部的异质性、丰富的纹理；阴影及投影差；混合像元问题等。其中,几何配准误差、相对辐射校正精度、阴影、像元尺度上较为破碎的变化检测结果、噪声等最为关键,本论文从这几个问题出发,针对变化检测中的变化与未变化两种类别对象,对高分辨率遥感影像变化检测的关键技术进行研究,主要包括以下几个方面：(1)提出了一种基于多变量变化检测(MAD)得到的差异影像为基础,做面向对象后分类处理的OB-MAD方法,减小了高分辨率遥感影像变化检测中不可避免的几何配准误差和阴影等造成的“伪变化信息”。OB-MAD方法结合了多变量变化检测(MAD)和基于对象的后分类方法。一方面,充分利用了面向对象和面向像元两种方法的特点,提高了变化检测精度。另一方面,几何配准误差和阴影等在高分辨率的遥感影像中具有较好的形状特征,MAD变换得到的差异影像集中了所有的变化信息,为面向对象的后分类处理,剔除“伪变化信息”奠定基础。航拍影像通常具有1m以内的空间分辨率,其几何校正、阴影和噪声等问题比卫星影像更为明显,是多时相航拍影像变化检测的一个瓶颈。湿地生态系统又是全球环境变化检测中的一个重要领域,湿地植被的变化、阴影等问题是湿地变化检测研究的重点和难点。因此,选用2006和2007年5月的航拍影像(分辨率0.15m)对美国加州中部的Gadwall North湿地生态系统进行变化检测研究,详细比较了基于差异影的OB-MAD和其他几种方法(OB-traditional, Threshold-MAD和PB-MAD)在变化检测中的差异。研究结果表明,针对变化检测研究中几何配准等造成的“伪变化信息”,三种基于MAD差异影像的方法取得的精度较高。其中,OB-MAD方法得到的总体精度最高(93.54%),其次是Threshold-MAD (90.07%)和PB-MAD (86.09%),相应的地面未变化像元的用户精度也是OB-MAD最高(90.57%),PB-MAD (82.2%)和Threshold-MAD (81.49%)。(2)多时相高分辨率遥感影像的变化检测预处理过程中,不可避免的几何配准误差、混合像元等都会带来一定的影响。针对这些问题,提出将广泛用于纹理特征提取、滤波除噪等方面的窗口用于相对辐射校正,用窗口区域范围内计算的相对平稳的值作为该区域的特征提高相对辐射校正的精度。研究选择广东番禺区两处具有不同空间异质性的区域,选择2005和2006的SPOT5影像(分辨率2.5m)和2006及2007的航拍影像(分辨率0.15m),3×3,5×5,7×7像元大小的窗口,自动的相对辐射校正算法MAD和稳健回归,进行基于窗口的相对辐射校正方法的敏感性分析。研究结果表明窗口的应用较好的提高了相对辐射校正的精度,同时1)不同空间异质性的遥感影像获得最佳辐射校正所需的窗口大小不一：异质性影像选择3×3像元,而均质性影像则选择5×5像元；2)local size尺度上,稳健回归在不同异质性的遥感影像中差异显著：异质性影像中稳健回归优于传统的手动方法,而在均质性影像中则表现的没有传统方法稳定,而MAD方法在两种空间特性的影像中均表现出比稳健回归更强的鲁棒性；3)几何配准误差在具有较高空间分辨率的航拍影像的变化检测中更为明显,提高相对辐射校正所需窗口较大。(3)基于波段相减的图像差值法对影像的质量和预处理要求相对较高,通常只用于来自同一传感器、同一时相的遥感数据；预处理中不可避免的几何配准误差、相对辐射校正精度、阴影等都是基于差异影像的高分辨率遥感影像变化检测中的典型问题；像素尺度上得到的变化检测结果较为破碎。在这样的背景下,提出一种面向对象的高分辨率差异图像的变化检测方法(OB-EM),即在对象尺度上,对基于MAD变换和最小噪声比率变换(MNF)得到的差异影像做阈值分析,将像元—像元之间的差异影像推广到对象—对象。研究选择广东番禺区2005和2006的SPOT5影像(分辨率2.5m),分析比较对象的不同特征选择、不同尺度的差异影像(对象和像元)、不同的方法(DFPS、Gams、EM、OB-MAD)等对变化检测的影响。研究结果表明：1)基于决策树的特征选择极大的提高了面向对象的变化检测精度；2)提出的OB-EM方法使得来自不同传感器的遥感影像用于变化检测变的可能；减小了对原始数据的要求和多时相影像的相对辐射校正的要求；改善了“椒盐效应”,提高了高分辨率遥感影像的变化检测精度。3)基于Kappa系数的显著性分析表明OB-EM和OB-MAD两者之间差异显著,说明基于对象的差异影像(OB-EM)比基于像元的差异影像(OB-MAD)能更好的减小甚至避免多时相高分辨率遥感影像变化检测中不可避免的几何配准误差和阴影引起的“伪变化信息”,具有较强的鲁棒性。更多还原

【Abstract】 High resolution imagery can provide more execute detail information, and have been widely used in scientific application. However, due to heterogeneity, texture, shadow and mixed pixel problem, there are big challenges in change detection using high resolution remotely sensed imageries. Additionally, spatial mis-registration, shadow and noise are common drawbacks of pixel-based method and particularly conspicuous in high resolution imagery, which can exaggerate or make thematic changes in change detection analysis. Considerable effort had been expended on minimizing the influence of these problems. In this study, we focused on these big issues and did the research on the key technology of change detection using high resolution remotely sensed imagery.(1) With respect to the inevitable mis-registration and shadow effects on change detection analysis, we proposed an object-based post-classification of the Multivariate Alteration Detection components method (OB-MAD), which can take advantage of pixel and object-based method. The extended difference image generated from MAD components can enhance the change information as well as amplyfying the false change caused by geometric distortion and shadow to an extent. On basis of the good shape information of false change, we proposed to do the object-based post-processing on this extended difference image. Obviously, aerial photograph with a high resolution less than 1 meter were useful in change detection, which can provide the real time multi-temporal imageries. However, compared with widely usd satellite imageries, the geometric distortion, shadow and noise problems of aerial imagery were significant. Change detection on wetland is a big topic in environmental monitoring and it is very difficult to detect the changes of vegetation and exclude the false information. Thus, very high spatial resolution images of drained managed wetland ponds (Gadwell North) were used to compare the proposed OB-MAD method with three commonly used classification methods in terms of minimizing the influence of mis-registration and shadow on the change detection analysis:(1) the traditional MAD method with thresholds (Threshold-MAD), (2) a pixel-based post-classification of MAD components with decision tree analysis (PB-MAD), and a traditional object-based post-classification method (OB-traditional). The proposed OB-MAD method, which utilized shape and textural information of objects derived from MAD components, produced the highest accuracy with respect to wetland change detection and successfully minimized the influence from the geometric distortion and shadow on the changed area. Overall accuracy was best for the OB-MAD method (93.54%), followed by the Threshold-MAD (90.07%), and PB-MAD (86.09%). The OB-MAD method also resulted in the greatest user’s accuracy for no-change pixels (90.57%), compared to 82.2% and 81.49% for PB-MAD and Threshold-MAD, respectively.(2) Inevitable geometric distortion of multi-temporal imageries and mixed pixel can significantly influence the accuracy of pre-processing in change detection analysis. New pixel represented by the averaged features within the window size can provide more reliable information, which had been widely used in texture extraction and filtering, since we proposed to combine the window size into the automatic relative radiometric correction methods. In our study, we tried to find out whether the proposed window size based relative radiometric correction method can improve accuracy. In addition, due to the fact that the accuracy of relative radiometric correction is different at homogeneous and heterogeneous landscape, we tried to analyze the sensitivity of window size at different landscapes. Homogeneous and heterogeneous landscapes generated from SPOT-5 imageries of 2005 and 2006 of PanYu city, China was preferred as our study areas. Three window size with 3 x 3,5×5 and 7 x 7 pixel, and two relative radiometric correction methods named MAD and robust regression were used to compare with the traditional manual method using two spatial resolution imageries of aerial image and SPOT satellite iamge. All of the results indicated that the optimun window size of proposed relative radiometric correction was significantly different at homogeneous and heterogenous landscapes. Specifically, for the homogeneous landscape, the optimun window size was 3 x 3 pixel, while 5×5 pixel window size was selected at heterogeneous landscape. Moreover, compared with traditional manual method at local size, robust regression was significantly different at both landscapes, while MAD method got robust better results. The proposed window size based relative radiometric correction method was sensitive to the spatial resolution of images.(3) Traditional pixel-based change detection method using difference image is difficult to perform change on multi-sensor imageries, because of the different number of bands and the impact of inter-channel correlation. In addition, pixel-based change detection methods are more sensitive to the inevitable geometric distortion, shadow and noise in pre-processing for high resolution imageries, and often suffer from "salt and pepper" effect in the resulting map. Note that another bottleneck is how to reasonably determine the threshold of change and no-change of difference image. In this context, we proposed an object-based method based on difference image generated from objects (OB-EM). Firstly, we got the object-based difference image using MAD and MNF method, and then found the change and no-change information using EM method. The proposed OB-EM method can take advantage of salient aspects of the MAD, MNF and object-based methods. SPOT-5 imagery of 2005 and 2006 in a case study of PanYu city, China were used to validate the proposed OB-EM method by comparing with other methods (DFPS, Gams, OB-traditional and OB-MAD). All the results indicated that feature selection can significantly improve the accuracy of change detection. Second, the proposed OB-EM method can take advantage of MAD, MNF and EM, which could deal with the data getting from different sensors. The "salt-and-pepper" effect could be improved well. Moreover, the accuracy of pre-processing of change detection was not required as high as OB-MAD. Finally, Z-test got a significant result between OB-EM and OB-MAD. It demonstrated that OB-EM was such a better method that can obviously recognize the false change from mis-registration and shadow.更多还原