基于多标签共享子空间学习和内核脊回归的空谱分类算法

针对高光谱图像维度高、地物间非线性可分造成的分类精度低等问题,提出一种基于多标签共享子空间和内核脊回归的空谱分类算法.该算法利用内核脊回归将地物相近像素在线性空间的不可分特征映射到高维空间中,实现分类特性在高维空间下的有效分离,以提高地物相近特性的区分精度;同时将高维样本数据映射到低维共享子空间中,在低维环境下以多类标为指导,引入低秩矩阵建立类别标签与共享空间的预测关系,挖掘多标签间的共同特性,提高融合利用多类别间的共同属性提高高光谱图像的分类精度;最后利用奇异值分解迭代法求解目标函数,一定程度上加速参数求解.在Indian Pines和Pavia University两组高光谱数据集上进行仿真实验,实验结果表明,与其他同类算法相比,在低样本比例下,本文算法在总体分类精度、平均分类精度和Kappa系数等评价指标上至少提高4.76%、4.24%和5.19%,与非内核化的算法相比,本文算法在基本不增加运行时间的情况下总体分类精度、平均分类精度和Kappa系数至少提高2.92%、2.8%和3.48%.

Space Spectrum Classification Algorithm Based on Multi-label Shared Subspace Learning and Kernel Ridge Regression

Aiming at the problems of high dimension of hyperspectral image and low classification accuracy caused by non-linear classification between objects,a space spectrum classification algorithm based on multi-label shared subspace and kernel ridge regression is proposed.The inseparable features of similar pixels in linear space are mapped to high-dimensional space using kernel ridge regression,which realizes the effective separation of classification characteristics in high-dimensional space,so as to improve the accuracy of the similarity of features.At the same time,the high-dimensional sample data is mapped into the low-dimensional shared subspace.In the low-dimensional environment,the multi-class label is used as a guide,and the low-rank matrix is introduced to establish the prediction relationship between the category label and the shared space,and the common characteristics among the multiple labels are mined.Improve the use of common attributes among multiple categories to improve the classification accuracy of hyperspectral images.Finally,the singular value decomposition iteration method is used to solve the objective function,which can speed up the parameter solution to a certain extent.Simulation experiments are carried out on two sets of hyperspectral datasets,Indian pines and Pavia University,compared with other similar algorithms,the overall classification accuracy,average classification accuracy and kappa coefficient of this algorithm are improved by at least 4.76%,4.24% and 5.19% at low sample ratios.Compared with the non kernel algorithm,the overall classification accuracy,average classification accuracy and kappa coefficient of the algorithm are improved by at least 2.92%,2.8%and 3.48% without increasing the running time.