遥感数据监督分类中训练样本的纯化

PURIFICATION OF TRAINING SAMPLES IN SUPERVISED CLASSIFICATION OF REMOTE SENSING DATA

本文分析了训练样本对遥感数据监督分类结果的影响，提出了训练样本纯化的理论与方法，即根据样本像元的光谱和空间信息来剔除训练样本中不合要求的样本像元。一个例子的试验研究表明，训练样本纯化后，各类型间的发散度、样本像元的概率密度函数与高斯分布的拟合度以及分类结果的精度都得到不同程度提高。

This paper analyses the effect of training samples on supervised classification of remote sensing data, proposes a theory and method for purincation of training samples, which uses spectral and spatial information to remove the undesirable sample pixels. An example shows that divergence between classes, goodness of fit to Gaussian distributinn and classification accuracy can be improved after purification of training samples.