基于均值漂移法进行多光谱遥感聚类研究

传统的聚类方法大都是基于空间划分的方法,一般都假设数据符合混合高斯模型。这在实际应用中往往是不成立的。在大部分模式分类的问题中,常见的参数形式不适合实际遇到的概率密度,特别是所有经典的参数密度都是单峰的,而一般遥感图像都是包含多峰的密度,因此分类结果往往不够精确。用于模式分类的非参数方法正是解决这类问题的一个重要途径,可以从本质上克服这一缺陷,而且可以发现任意形状的聚类。均值漂移方法是基于密度估计的非参数聚类方法,遥感图像的聚类分析可以通过均值漂移方法来实现,而且均值漂移过程不需要预先给出地物的类别数目,在聚类过程中自动确定类别数,这对于图像中类别数目不易确定的情况,给非监督遥感图像聚类带来方便。

Mean Shift Based Clustering Analysis of Multi-spectral Remote Sensing Imagery

Traditional clustering techniques are always based on partitioning of the feature space. An usual assumption used in these methods is that the feature space can be modeled as a mixture of Gaussians. However, the assumption is not true in many real remote sensing images. Furthermore, many usual parametric forms of probability do not accord with real data. Especially the feature space of remote sensing image is multivariate, multi-modal probability distribution, while the classical parametric density functions have only one local maximum. Therefore, incorrect classification results are easily obtained by parametric methods. Nonparametric methods in feature space analysis avoid the use of the normality assumption. Arbitrarily structured feature spaces can be analyzed only by nonparametric methods since these methods do not have embedded assumptions. Mean shift algorithm is a simple nonparametric technique for estimation of the density gradient. The mean shift algorithm can be used to cluster multi-spectral remote sensing imagery. The number of clusters is not specified a priori and determined in the mean shift procedure automatically. This will be convenient in unsupervised classification of remote sensing image when the number of clusters das not easy to get.