变化检测是通过对地物特征或者现象进行多次观察，并识别其状态变化的过程。随着影像数据源获取手段的增加以及获得影像数据质量的提高，为人们实现高质量的变化检测提供了更多可能。并且，随着社会飞速的发展，城市化、工业化的进程加快，人们对变化检测的应用需求也不断增长。目前，已有广泛的变化检测技术应用，如森林植被的变化检测，土地利用及覆盖类型的变化检测，灾害地区的分析及评估，城镇变化实时监控等。虽然在变化检测领域已拥有诸多成果，但由于变化情况的复杂性及多变性，没有一种方法可以解决所有情况。如何针对具体的实验情境，实现变化检测的自动化，智能化仍是人们不断追求的目标。 本文围绕变化检测进行了如下研究： （1）首先，采用一种增长式分割方法实现图斑对象的获取，图斑对象为本文变化检测方法的基本处理单位，通过此方法，既可避免传统多时相变化检测如通过合并影像分割结果等方式产生的边界描述不一致的问题，又可避免传统矢量-影像法通过栅矢套合所获得的图斑对象对局部变化情况描述不完整的情境。通过增长式分割方法获得图斑对象后，提取图斑在新时期影像中的光谱特征，纹理特征及自定义特征，最后为避免特征冗余，应用主成分分析法实现特征的降维，完成特征空间的构建。 （2）针对传统多时相变化检测易受拍摄季节、拍摄 角度及太阳高度角等因素影响，应用一种矢量-影像方法，不同于以往影像-影像法，应用前后两个时相影像，对其应用一定的相似性度量方法计算新旧影像的变化差异，并在一定阈值的基础上进行变化检测，或在统计前后时相分类结果基础上进行变化检测的研究。矢量-影像法应用旧时相矢量图与新时相影像数据进行变化检测。 （3）因变化图斑与未变化图斑相比，具有“少且不同”的特点，故可将数据挖掘方法中的异常检测方法引入到变化检测中，采用如基于距离的方法，基于密度的方法，基于模式的方法进行变化指数的计算，后应用贝叶斯阈值计算方法获得图斑的变化阈值，实现变化/未变化图斑的判别。对比马氏距离、局部离群因子与孤立森林三种变化指数计算方法，验证本文将数据挖掘方法引入变化检测中的可行性，并得出基于矢量图与孤立森林方法的结果最为理想。

Change detection is made through multiple observations of ground objects or phenomena to identify its’ change process. With the increase of image data access methods and the improvement of data quality, which make it possible to realize high quality change detection. Meanwhile with the rapid development of society, speed up the process of industrialization urbanization, the change detection application demand is also increasing at present. The change detection technology has widely application, such as the dynamic monitoring of forest vegetation, land use and cover change detection, hazard analysis and assessment of the disaster area, real-time monitoring of the town change and so on. Although many achievements have been made in the field of change detection, due to the complexity and variability of changes, there is no one method can cover all the cases. How to realize the automation of change detection according to then specific experimental conditions, and the high precision is still a constant pursuit of people. This article around the change detection were studied as follows: (1) First of all, using an incremental segmentation method to obtain objects, through this method, which can avoid the traditional multi-temporal change detection, such as by combining image segmentation result way to produce the problem of not consistent boundary description, It can also avoid the incomplete description of local changes of the image object obtained by traditional vector-image method through directly combine vector data with image. After image object is obtained by means of incremental segmentation method, spectral features, texture features and customized features of image are extracted. Finally, to avoid feature redundancy, principal component analysis is applied to realize feature dimension reduction and complete the construction of feature space. (2) In view of the traditional multi-temporal change detection is susceptible to shooting season, shooting angle, etc. Using a vector-image method is different from the previous image-image method. With the application of those two phase images, the change differences between the old and new images were calculated by using a certain similarity measurement method, and the change detection was carried out on the basis of a certain threshold value, or the change detection was carried out on the basis of the classification results of the time phase before and after the statistics method. The vector-image method using the vector data and the new time image data detect the changes. (3) The anomaly detection method in data mining method can be introduced into the change detection because the change objects has fewer and different characteristics compared with the unchange objects. For example, the distance-based method, the density-based method and the pattern-based method are adopted to calculate the change index. Then, the Bayesian threshold method is applied to obtain the change threshold of the object, so as to realize the discrimination of the changed/unchanged objects. The feasibility of introducing the data mining method into the change detection is verified by comparing the three calculation methods of Mahalanobis distance, local outlier factor and isolated forest method.