遥感影像变化检测方法的比较──以堰塞湖灾害点为例

Comparison of Change Detection Methods for Remote Sensing Imageries:Take Yanse Lake Disaster Area as an Example

一个对象(现象)在不同的时间发生的变化存在着一定的相关性,利用光谱和空间相关性信息提取出地物发生变化的方法──基于邻域相关性图(NCIs)和基于交互相关性分析(CCA)的变化检测方法分别取得了不错的效果。为了探究这两种方法在土地覆盖变化检测中的适用性,对其进行应用并评价。以堰塞湖灾害点为试验区,选取了同一年份的两期高分一号卫星影像,对比了监督的基于NCIs的方法和监督的分类后比较法及非监督的基于CCA的方法和非监督的分类后比较法。结果表明:基于NCIs的方法在邻域半径等于3像素时,总体精度最大,为93.05%,比监督的分类后比较法(采用支持向量机分类器,总体精度为93.03%)稍好,但其受邻域半径的制约较大;基于CCA的方法在阈值等于20像素时,总体精度最高,为93.95%,比非监督的分类后比较法(采用ISODATA分类器,总体精度为90.52%)好,但不足在于需要经验判断并试验求取最佳阈值,以及需要已知前一个时相的地表真实值。

A certain correlation exists between the changes of an object （phenomenon） at different time, which can be used as one of the important features to detect changes. Many traditional change detection methods like layer arithmetic, change vector analysis, and multi- data direct comparison take no consideration of context information and spectral and spatial correlation. Neighborhood correlation images （NCIs） and cross correlation analysis （CCA） change detection methods that use spectral and spatial correlation information to extract changes in land-use both have been verified to have a good accuracy. In this paper,these two methods are applied and evaluated in order to verify their applicability. Dual phase images from GF-1 satellite of yanse lake disaster areas are chosen to conduct the experiments, where the supervised NCIs and the supervised post-classification comparison （ PCC ） based on support vector machine （ SVM ） and maximum likelihood, and the unsupervised CCA and the unsupervised PCC based on ISODATA have been compared. The results indicate that the supervised NCIs method shows the highest overall accuracy with a 3-pixel radius of neighborhood ,93.05%, and it is a little higher than the supervised PCC method based on SVM whose overall accuracy is 90.52%. The demerit of the NCIs method is the constraint of neighborhood size. Moreover, the unsupervised CCA method performs best with a threshold of 20 pixels,where the overall accuracy is 93. 95%. It shows more advantages than the unsupervised PCC method based on ISODATA,the overall accuracy of which is 90.52%.However, it is not easy for the CCA method to choose the proper threshold and obtain the ground-truth of a single phase image.