摘要
针对遥感影像分类对初始训练集要求严格且数据含有大量未标记样本的特点,提出基于预聚类和主动半监督学习算法。首先利用网格聚类筛选初始训练样本,避免决策边界偏倚;然后,利用主动学习方法挑选最有价值的未标记样本交由专家标记后加入训练样本集;最后,通过半监督学习充分利用未标记样本信息,使得减少样本标记成本的同时也可以获得良好的分类效果。UCI数据集及遥感影像数据集上的仿真结果表明,所提算法能够获得较好的分类效果。
Aiming at the characteristics that remote sensing image classification are strict with the initial training samples and the data contains a lot of unlabeled samples,this paper proposes active learning based on grid-clustering and semi-supervised supper vector machine algorithm. Firstly,the initial training samples are screened with the grid clustering method to avoid the decision-making boundary bias. Then,the most valuable unlabeled samples are selected with the method of active learning,labeled by experts and put into the training sample group. Finally, taking full advantage of unlabeled sample information through semi-supervised learning,we can achieve good taxonomies while reducing sample markup cost. Simulation results on UCI data sets and remote sensing image data sets show that the proposed algorithm can help obtain higher classification accuracy with fewer training samples.
作者
汪婵
王磊
丁西明
WANG Chan;WANG Lei;DING Xi-ming(College of Electrical and Electronic Engineering,Anhui University of Science and Technology,Fengyang Anhui 233100,China)
出处
《湖北第二师范学院学报》
2018年第2期58-64,共7页
Journal of Hubei University of Education
基金
安徽省自然科学基金项目(1708085QF146)
安徽科技学院人才引进项目(DQYJ201602)
关键词
主动学习
半监督
网格聚类
遥感影像分类
active learning
semi-supervise
grid clustering
remote sensing image classification
