摘要
针对不同时期高分辨率遥感影像变化检测中城区建筑物因投影差差异所产生的误检测现象,提出了一种综合应用光谱和纹理特征的建筑物变化检测方法。以变化和未发生变化地物影像的散度作为可分性依据,首先对光谱差分影像在混合高斯密度模型下建模,并采用马尔可夫最小错误概率准则提取初始变化区域,往往含有错判的建筑物。然后将误判建筑物影像类和真实变化影像类构成训练集,通过引入多通道Gabor滤波器,提取训练集的纹理差分特征,并采用分类别PCA变换实施纹理差分特征的选择。最后对选择出的纹理差分特征依据高斯混合密度模型建模,并用马尔可夫最小错误概率提取真变化区域,即可去除光谱信息检测所产生的伪变化。试验表明,本文方法能够较好地解决建筑物变化的错判问题,提高了影像变化检测的精度。
A change detection method for high-resolution remotely sensed imagery is designed by combined the texture features with spectral features. The method can decrease change detection errors caused by building projection difference between two period imageries. It utilizes transform divergence between change and no-change to establish change detection strategy. Firstly, the differential spectral features are modeled by using Gaussian mixture model （GMM） and initial change area is obtained by using MRF minimum error probability. But for high-resolution imagery, the projection difference will disturb change area＇s result. For decreasing these errors, the texture features are extracted by multi-channel Gabor filter. When using these texture features, it is necessary to reduce these texture features＇ correlation and relieve computation burden. A class-within PCA is adopted to select texture features. Using selected texture features, the initial change area is modeled by using GMM and ＂false change＂ is got rid of using MRF minimum error probability. The experiment has shows that the mentioned method can remove false change caused by projection difference and improve change detection accuracy.
出处
《武汉大学学报（信息科学版）》
EI
CSCD
北大核心
2007年第6期489-493,共5页
Geomatics and Information Science of Wuhan University
基金
国家973计划资助项目(2006CB701302)
全国优秀博士学位论文作者专项资金资助项目(200142)
教育部长江学者和创新团队发展计划--创新团队资助项目(IRT0438)
关键词
影像变化检测
多通道Gabor滤波器
分类别PCA变换
混合高斯密度模型
马尔可夫最小错误概率
remote sensing imagery change detection
multi-channel Gabor filter
divergence of class-within PCA
Gaussian mixture model
Markov random field minimum error probability
