结合SOM神经网络和混合像元分解的高光谱影像分类方法研究

Research on the Classification Based on SOM and LSMA for Hyperspectral Image

本文对SOM神经网络算法进行改进,在标类的过程中采用3个策略加以控制,对初始产生的自组织映射图进行调整。通过改进,那些映射到可靠神经元的像素得到了很好的分类,而那些映射到不可靠神经元的像素都被作为不可分像元而提取出来。继而,从混合像元分解的角度来对这些不可分像元进行处理,按类型分解的思想确定混合像元的类别,实现对不可分像元的分类。将SOM神经网络和混合像元分解相结合的分类方法应用于高光谱图像的分类中,通过实验表明了该方法能较好地改善分类效果,提高分类精度。

In SOM algorithm it will create a map in output layer in which the cells are labeled class ID, e.g. 1, 2, 3, etc. It＇ s curial for correctly classifying the data to make map. In this paper, we focus our interests on analyzing the map and the process of creating map to improve the SOM. We take three measures to change the map. We can classify the pure pixels and find the mixed pixels through the changed map. Furthermore, we can process the unclassified pixels from the view of linear spectral mixture analysis （LSMA）. Furthermore, we consider the two constraints： unnegative and the sum one, so the constraint spectral mixture analysis（CSMA） is applied in this paper. After CSMA, we assign the class ID to the endmember which has largest proportion in the mixed pixel. So, the spectral unmixing classification based on category proportion is performed to the unclassified pixels. Thus, we can get the extreme classification combining the former results. The experiment shows that the classification combined SOM with LSMA can get better classification results and well improve the classification accuracy.