高光谱分类EM算法密度模型参数初值的选取

Determination of Initial Parameter Values for EM Algorithm Applied in Hyperspectrum Imagery Classification

在高光潜影像的分类过程中，往往无法获取足够数量的训练样本点，这是由于影像数据维数的大幅增加以及地面信息的复杂程度所决定的。EM算法的应用可以有效地缓解训练样本点数量与数据维数比率过小的矛盾，但该算法只能保证获得各参数的局部极大估值，因此选取合适的起始点就成为获得理想分类结果的前提条件。由于类别可分性对EM算法的估值精度有直接影响，本文论证了通过对训练样本点进行低通滤波，可以使类别可分性得到改善。实验表明，在此基础上进行EM算法，可以得到较为理想的处理结果。

As a result of increment of imagery data dimensions and complexity of ground truth, it is very difficult for us to get enough training samples for hyperspectral imagery classification. The problem of undersize ratio of training points and data dimensions could be solved by applying EM algorithm. However, EM algorithm can only get local maximum estimators of parameters. To determine suitable initials, thus, would be a precondition to achieve an ideal classification. Since the accuracy of the parameter estimates by using EM algorithm is directly influenced by the mend classifibility, a lowpass filtering is applied for processing the training points. By this way,the classifibility is improved,aud the EM algorithm.