改进M-training算法的高光谱图像分类

Hyperspectral image classification based on improved M-training algorithm

为了解决高光谱数据有标签样本数量有限的分类问题,提出将M-training算法应用于高光谱图像分类。采用两个SVM、一个K近邻(KNN)以及一个随机森林(RF)进行分类器组合,对传统M-training算法进行改进,增强分类器的多样性和差异性。为了充分考虑大量无标签样本的影响,采用有标签样本与无标签样本错误率加权作为有标签样本集更新的限制条件,从而有效地扩大了有标签样本集。实验结果表明:改进算法和传统的M-training算法相比较,在总体分类精度与Kappa系数上分别提高1. 85%～12. 10%与0. 021 5～0. 141 3,从而验证了该算法的有效性。

To solve the problem of limited hyperspectral image classification of labeled samples,a modified M-training algorithm was applied to such classification.The algorithm employs two Support Vector Machines(SVMs),one K Nearest Neighbor(KNN),and one Random Forest(RF)classifier to enhance the diversity of classifiers,so as to improve the traditional algorithm of M-training.Taking the impact of a large number of unlabeled samples into account,the error rate of labeled and unlabeled samples was weighted as the limiting condition to updating the set of labeled samples,effectively enlarging the labeled set.The results showed that,compared with the traditional M-training algorithm,the proposed algorithm improves both overall classification accuracy and the Kappa coefficient by 1.85%～12.10%and 0.021 5～0.141 3,respectively,verifying the effectiveness of the improved algorithm.