基于粗糙集和模糊聚类的超谱波段约简

Hyperspectral Band Reduction Based on Rough Sets and Fuzzy C-Means Clustering

由于超光谱图像数据量大,维数高给分类识别处理带来不便,该文提出一种可行有效的波段约简方法.通过FCM聚类将原始波段划分为若干等价波段组,然后根据最大隶属度原则只保留每组中具有代表性的波段,达到维数减小的目的。其中,模糊聚类中相似度的定义是基于超谱相邻波段间的相关性,利用粗糙集理论中的处理属性依赖性的方法合理表达出来。实验表明,这一方法既有效地缩减了高维数据,又尽可能少地损失有用信息,保持了原始波段的分类能力。

A method of hyperspectral band reduction based on Rough Sets (RS) and Fuzzy C-Means (FCM) clustering is proposed, which consists of the following two steps. First, Fuzzy C-Means clustering algorithm is used to classify the original bands into equivalent band groups, which employs the concept of attribute dependency defined in RS to define the distance between a group and the cluster center, viz. the correlatives of adjacent bands. Then the data is reduced by selecting the only one from each group with maximum grade of fuzzy membership. With this approach, great dimension of band is decreased while preserving much wanted information. Simulation results prove the effectiveness of this approach.