基于改进的局部保持投影高光谱图像分类研究*

针对高光谱图像分类中基于流形的降维方法进行了研究。提出一种改进的局部保持投影（LPP）方法即MLPP方法。该方法利用标签信息避免了传统LPP在邻接图构建中很难确定的邻域大小的选择问题，同时采用更能反映高维数据间相关性的统计特征量相关系数来衡量数据之间的相似程度。设计的权重矩阵既保持类内数据的几何结构，又最大化类间距离。而且MLPP不依赖任何参数和先验知识。在两个高光谱图像上的实验结果表明MLPP增加了不同光谱特征地物之间的可分性，在提高分类性能上明显优于其他传统的降维方法。

The Research of Hyperspectral Image Classification Based on Modified Locality-Preserving Projection Approach

In view of the research of hyperspectral image (HSI) classification based on manifold dimensionality reduction (DR) method, we propose a modified version of the original Locality-preserving projection (LPP) called MLPP. This method use the tag information to avoid the problem of neighborhood size selection which is difficult to determine in adjacency graph construction by the traditional LPP. At the same time, it adopts a statistical characteristic, correlation coefficient, to measure the similarity between the data which can reflect the correlation between high dimensional data. The weighted adjacent matrix designed not only maintains the geometric structure of the data in the class, but also maximizes the distance between the classes. Moreover, MLPP does not depend on any parameters or prior knowledge. Experiments on two HSIs demonstrate that MLPP increases the separability between objects with different spectral characteristics and is remarkably superior to other conventional DR methods in enhancing classification performance.