一种模型选择优化准则及其在高光谱图像非监督分类中的应用

An Unsupervised Classification Method Based on a Model Selection Criterion for Hyperspectral Data

选择合适的类别数是非监督分类中的一个关键问题 .针对采用高斯混合建模的高光谱图像非监督分类问题 ,该文提出了一种基于主成分分析 (PCA)的最小描述长度 (MDL)型模型选择准则 (文中简称为PMDL)来确定分类类别数 ,即根据PCA变换后保留的各主成分表达的数据方差不同而应具有不同的编码长度这一事实 ,在计算描述长度时对各维进行加权 .分类过程中 ,论文采用期望最大化 (ExpectationMaximization)算法在合并的策略下对PCA变换后的数据求解混合模型 ,并应用所提出的准则进行模型选择从而确定待分类的类别数 .仿真数据实验证实了新准则的有效性和优良的性能 。

A new principal component analysis (PCA)-based minimum description length (MDL) type criterion (termed PMDL) is proposed in this paper to solve the key problem of class number selection in unsupervised classification.It is based on the fact that data of different dimensions after PCA transform should be encoded with different code length,as they represent different amount of data variance.We perform unsupervised classification to hyperspectral image by Gaussian mixture modeling,estimate the parameters of the mixture model using the expectation maximization (EM) algorithm in merged operations to data after PCA linear projection,and select the number of components according to the proposed criterion.Experiments on a set of synthetic data verify the new criterion.The whole algorithm performs quite effectively and gives proper class number without any prior information applied to real data.