中文摘要：

为解决高光谱图像中高维数据和有标记训练样本不足的矛盾导致“维度灾难”问题,提出一种无监督的基于流形学习的波段选择（MLBS）方法。首先通过流形学习方法,得到原始数据的流形嵌入映射;然后通过LASSO优化过程,运用顺向坐标下降算法,得到原始波段对每个流形结构维度的贡献度;最后统计每个波段的贡献度,选取贡献度大的波段形成波段子集。用真实的AVIRIS高光谱图像对算法进行仿真实验的结果表明,本文方法在小样本下的高光谱地物分类识别问题上具有良好的效果。

英文摘要：

For solving the ＂curse of dimensionality＂ problem. Caused by the contradiction between high dimensional data and insufficient labeled training samples in huperspectral image, this paper presentes an unsupervised manifold learning based band selection （MLBS） method. Firstly, the embedding projection of each pixel is calculated through manifold learning algorithm;secondly, the contribution degree of each dimensional of manifold structure is calculated through a lasso optimization process and coordinate-wise descent algorithm; at last,the contribution degree of each band is calculated, while the bands with large contribution degree form band subset. The proposed band selection approach is experimentally evaluated using real AVIRIS hyperspectral data set. Experimental results in AVIRIS data demonstrate that the proposed algorithm can yield good performance in hyperspectral land-cover classification with small sampies.