基于混合高斯密度模型和空间上下文信息的遥感影像变化检测方法及扩展

Research and extension of remote sensing image change detection method based on gaussian mixture model and contextual information

在运用混合高斯密度模型对差分影像建模的基础上,分别采用顾及上下文信息的概率松弛迭代法和马尔可夫随机场模型法进行影像的变化检测。首先,提出一种运用遗传K均值算法与EM算法联合解算高斯混合密度模型参数的方法,该方法可以自动地解算出模型的统计参数,结果与手工选择样本的解算结果完全一致。然后,比较概率松弛迭代法以及马尔可夫随机场模型法的影像变化检测效果,得出基于模拟退火法的马尔可夫随机场法效果较好的结论。最后,对传统的基于模拟退火法的马尔可夫随机场方法进行改进,提出了一种变权马尔可夫随机场方法,检测结果能更好地保持影像的结构性,并有效去除了孤立噪声。

Multi-temporal remotely sensed imagery change detection is a hot topic in recent years. Most researchers pay attention to statistical pattern recognition principal to solve the problem. In this paper, we discuss the problem from three aspects. Firstly we focus on GMM model statistic coefficients resolve method. The Expectation Maximization （EM） is the most commonly method to calculate GMM coefficients. However, EM algorithm often converges to local value. So we combine genetic k-means algorithm （GKA） with EM to modify it. Using the initial clustering result obtained by GKA, we are able to initialize EM globally. The combination helps EM search out globally optimal solution and enhances the automatic degree. When we get the global optimal results, it is easy for us to obtain change detection result using Bayes Rule for Minimum Error （BRME）. However, the BRME doesn＇t take into account the image＇s contextual information. It is well known that one pixel belongs to ＂change＂ or ＂no change＂ depends not only on itself, but also on its neighbor pixels. There are two ways to model contextual information. The first is probability relaxation iteration, and the second is Markov Random Field （MRF）. MRF has two commonly used solution methods; one is Iterated Conditional Method （ICM） , and the other is Simulated Annealing （SA）. In this paper, we compare the three spatial contextual change detection methods using visual effect and kappa coefficient. The experiment shows that MRF based on simulated annealing has better performance than the other two. Through the above experiments, we find that traditional MRF deals all pixets equally and ignores the neighbor local features. In fact, in different region, the spatial information has different impact. We analyze the impact and propose variable weight MRF method. It can adaptively vary spatial impact according to different image local features. It has virtues of preserving structural change and filter noises. The experiment proves that variable weight MRF gets the best result.