针对现有非监督分类方法不能自动确定最佳分类数、对包含噪声的大数据集适应性差的问题,提出了一种基于几何概率的聚类分析方法,即按照先分大类、后分小类、逐层细分的顺序来确定分类方案,其同一分类层次上不同子类进一步细分的步骤相同,但执行过程彼此相互独立。在每一分类层次上,以几何概率为理论基础,根据样本在特征空间中的分布结构确定类的数目、提取类的中心位置、搜索类的边界。通过TM遥感影像的分类实例及其与ERDAS中的监督、非监督分类方法进行对比的结果表明,基于几何概率的聚类分析方法能明显提高分类精度。

Current unsupervised classification methods unable to determine optimum classes and poorly suitable for large and noise-included datasets.A cluster analysis approach based on geometric probability has been put forth,which gradually generates a hierarchical classification scheme in top-down order.The step of further classification for different sub-classes at the same classification level is identical but the executive process is independent for each other.Determining class number,extracting class central position,and searching the boundaries between different classes are performed at each classification level according to the distribution structure of samples in the feature space under the theoretical support of geometric probability.The algorithm of cluster analysis based on geometric probability is contrasted to ERDAS built-in supervised and unsupervised classification algorithms via a case of classifying a TM(thematic mapper) remotely sensed image.It turns out that the one based on geometric probability can obviously improve the classification accuracy.