关于LLE算法的监督型参数设置方法及应用

The method and application of supervised parameter setting of LLE algorithm

采用局部线性嵌入(Locally Linear Embedding,LLE)算法进行数据降维时,不仅能保持数据分布的局部线性特征,同时还能保存数据分布的流形结构,因此该算法常用于高光谱影像的数据降维。其中,关于最近邻像元个数K的设置是执行该算法的关键。然而,关于K值的设置,目前尚无一个行之有效的方案。针对这一问题,文中基于监督型特征提取的思想,从"线性预测误差均值最小化"的角度出发,提出了一个监督型参数设置方法。同时,为了验证该方法的可行性和优越性,结合两个实验区Hyperion影像关于第26至57波段包含的32维光谱数据,进行了降维实验。最后,通过分析对比实验结果,证明了:采用LLE算法进行高光谱影像数据降维时,若依据文中所提方法设置的K值,能获得噪声点少且地物细节信息更加丰富的低维影像数据。

When the Locally Linear Embedding( Locally Linear Embedding,LLE) algorithm was used to reduce the dimension of data,not only the local linear characteristics but also the manifold structure about distribution of data could be preserved. So this algorithm had been often used to reduce the dimension of data on hyperspectral imagery consequently. Among them,the setting of the nearest neighbor pixel number K is the key to executing the algorithm. However there is no effective scheme for setting K currently. Aim at this problem,based on the principle of supervised feature extraction,a supervised method of setting parameter was proposed in this paper from the perspective of "minimizing the mean of linear prediction error". Meanwhile,in order to verify the feasibility and superiority of the proposed method,the experiments on data dimension reduction were carried out by combining the 32 D spectral data( i. e. from the 26 th band to the 57 th band) on Hyperion imagery of two experimental areas. Finally,by analyzing and comparing the experimental results,it showed that: According to the K which is set by using the proposed method in this paper,the low-dimensional imagery with less noise points and more abundant cultural details could be obtained by using LLE algorithm to reduce the dimension of spectral data on hyperspectral imagery.