鲁棒性监督等距特征映射方法

A robustness supervised isometric feature mapping method

提出一种鲁棒的监督Isomap算法(RS-Isomap算法).该方法首先在标准PCA基础上,为样本邻域点引入权值因子,产生新的优化问题,使用加权迭代最小二乘法求解.然后利用加权主成分分析,遍历每一个样本点,计算归一化权值之和,得到样本的可信度.接着融合样本的可信度、类别和邻域信息,重新定义样本点之间的测地距离,计算最短距离矩阵,采用多维标度分析和广义回归神经网络分别构建训练样本和测试点的嵌入坐标.实验表明:新方法比传统的Isomap方法有较强的抗噪声能力,能有效地提高高光谱图像的分类精度,在运行时间上具有可行性.鲁棒性的监督Isomap算法是一种有效的高光谱遥感图像特征提取方法.

A robustness supervised Isomap（RS-Isomap） algorithm was proposed. On the basis of standard PCA, the RS-Isomap algorithm first introduced weight factors into Sample neigh-bourhood points to generate new optimization problems, and adopted weighted iterative least square method to solve these problems. Then a weighted principal component analysis was used to calculate sample reliability. After that, the sample reliability, category and neighbour- hood information were integrated, and the distance between the samples was redefined. Finally multidimensional scaling and generalized regression neural network were applied to construct the embedding coordinates of training samples and test points by means of the new distance ma- trix. Experimental results demonstrate that the proposed method has higher anti-noise per- formance than the traditional one. It can effectively improve the classification precision of hyperspectral image with feasible runtime, which is an effective feature extraction method for hy-perspectral remote sensing image.