第16卷第6期 智能系统学报 Vol.16 No.6 2021年11月 CAAI Transactions on Intelligent Systems Now.2021 D0:10.11992/tis.202007037 网络出版地址：https:/kns.cnki.net/kcms/detail/23.1538.TP.20210922.1140.002.html 多视图主动学习的多样性样本选择方法研究 陈立伟，房赫，朱海峰 (哈尔滨工程大学信息与通信工程学院，黑龙江哈尔滨150001) 摘要：为了去除高光谱图像多视图主动学习分类中的所选样本的冗余，降低人工标记成本，本文提出了两种 用于多视图主动学习分类中的多样性样本选择方法。将高光谱图像进行超像素分割，将所选样本中属于不同 的超像素的样本加入训练集，其余样本加入候选集；比较各视图对样本的预测标签，将所选样本中预测标签不 完全相同的样本加入训练集，其余样本加入候选集。本文分别用这两种方法对传统多视图主动学习的样本选 择方法进行改进，并用两组高光谱图像数据进行实验。实验结果表明：使用这两种方法改进后，所得分类精度 不变，使用的训练样本数量大幅诚少。 关键词：高光谱图像分类：多视图主动学习：多样性；样本选择：超像素：训练样本数量：预测标签：分类精度 中图分类号：TP753文献标志码：A文章编号：1673-4785(2021)06-1007-08 中文引用格式：陈立伟，房赫，朱海峰.多视图主动学习的多样性样本选择方法研究J,智能系统学报，2021,16(6)： 1007-1014. 英文引用格式：CHEN Liwei,.FANG He,ZHU Haifeng.Diversity sample selection method of multiview active learning classifica- tion[J.CAAI transactions on intelligent systems,2021,16(6):1007-1014. Diversity sample selection method of multiview active learning classification CHEN Liwei,FANG He,ZHU Haifeng (College of Information and Communication Engineering,Harbin Engineering University,Harbin 150001,China) Abstract:To remove the redundancy of selected samples in the multiview active learning classification of hyperspec- tral images and reduce the cost of manual marking,this paper proposes two methods for the selection of diverse samples in the multiview active learning classification.First,hyperspectral images are divided into superpixel segments, then samples belonging to different superpixel segments are added to the training set,and the remaining samples are put back into the candidate set.Second,the prediction labels of the samples from each view are compared,then the samples with different prediction labels are added into the training set,and the remaining samples are put back into the candidate set.In this study,the two methods are used to improve the sample selection method in the traditional multiview active learning classification,and experiments are conducted in two groups of hyperspectral image data.The results show that the accuracy of classification is unchanged,yet the number of training samples is greatly reduced after using the two methods. Keywords:hyperspectral image classification;multiview active learning;diversity;sample selection;superpixel;num- ber of training samples,prediction labels;accuracy of classification 随着遥感技术的迅速发展，高光谱图像(hy-贵又费时。主动学习(active learning,.AL)方法 perspectral image,.HSI)在土地覆盖物分类中得到可以有效解决HSI标记样本少的问题s-刀。在AL 了广泛的应用1。训练一个HSI分类器，通常需 方法中，多视图主动学习(multiview active learning,. 要大量的标记样本，而标记样本的采集过程既昂 MVAL)方法可以从多个视图中提取互补信息，大 大减少训练样本的数量⑧-o。 收稿日期：2020-07-23.网络出版日期：2021-09-23. 基金项目：国家自然科学基金项目(61675051). 学者们对MVAL的样本选择方法展开了广 通信作者：朱海峰.E-mail:zhuhaifeng(@hrbeu..edu.cn. 泛研究：文献[1l]提出了自适应最大不一致(ad-DOI: 10.11992/tis.202007037 网络出版地址: https://kns.cnki.net/kcms/detail/23.1538.TP.20210922.1140.002.html 多视图主动学习的多样性样本选择方法研究 陈立伟，房赫，朱海峰 （哈尔滨工程大学 信息与通信工程学院, 黑龙江 哈尔滨 150001） 摘 要：为了去除高光谱图像多视图主动学习分类中的所选样本的冗余，降低人工标记成本，本文提出了两种 用于多视图主动学习分类中的多样性样本选择方法。将高光谱图像进行超像素分割，将所选样本中属于不同 的超像素的样本加入训练集，其余样本加入候选集；比较各视图对样本的预测标签，将所选样本中预测标签不 完全相同的样本加入训练集，其余样本加入候选集。本文分别用这两种方法对传统多视图主动学习的样本选 择方法进行改进，并用两组高光谱图像数据进行实验。实验结果表明：使用这两种方法改进后，所得分类精度 不变，使用的训练样本数量大幅减少。 关键词：高光谱图像分类；多视图主动学习；多样性；样本选择；超像素；训练样本数量；预测标签；分类精度 中图分类号：TP753 文献标志码：A 文章编号：1673−4785(2021)06−1007−08 中文引用格式：陈立伟, 房赫, 朱海峰. 多视图主动学习的多样性样本选择方法研究 [J]. 智能系统学报, 2021, 16(6): 1007–1014. 英文引用格式：CHEN Liwei, FANG He, ZHU Haifeng. Diversity sample selection method of multiview active learning classifica￾tion[J]. CAAI transactions on intelligent systems, 2021, 16(6): 1007–1014. Diversity sample selection method of multiview active learning classification CHEN Liwei，FANG He，ZHU Haifeng (College of Information and Communication Engineering, Harbin Engineering University, Harbin 150001, China) Abstract: To remove the redundancy of selected samples in the multiview active learning classification of hyperspec￾tral images and reduce the cost of manual marking, this paper proposes two methods for the selection of diverse samples in the multiview active learning classification. First, hyperspectral images are divided into superpixel segments, then samples belonging to different superpixel segments are added to the training set, and the remaining samples are put back into the candidate set. Second, the prediction labels of the samples from each view are compared, then the samples with different prediction labels are added into the training set, and the remaining samples are put back into the candidate set. In this study, the two methods are used to improve the sample selection method in the traditional multiview active learning classification, and experiments are conducted in two groups of hyperspectral image data. The results show that the accuracy of classification is unchanged, yet the number of training samples is greatly reduced after using the two methods. Keywords: hyperspectral image classification; multiview active learning; diversity; sample selection; superpixel; num￾ber of training samples; prediction labels; accuracy of classification 随着遥感技术的迅速发展，高光谱图像 (hy￾perspectral image，HSI) 在土地覆盖物分类中得到 了广泛的应用[1-3]。训练一个 HSI 分类器，通常需 要大量的标记样本，而标记样本的采集过程既昂 贵又费时[4-5]。主动学习 (active learning，AL) 方法 可以有效解决 HSI 标记样本少的问题[6-7]。在 AL 方法中，多视图主动学习 (multiview active learning, MVAL) 方法可以从多个视图中提取互补信息，大 大减少训练样本的数量[8-10]。 学者们对 MVAL 的样本选择方法展开了广 泛研究：文献 [11] 提出了自适应最大不一致 (ad- 收稿日期：2020−07−23. 网络出版日期：2021−09−23. 基金项目：国家自然科学基金项目（61675051）. 通信作者：朱海峰. E-mail：zhuhaifeng@hrbeu.edu.cn. 第 16 卷第 6 期 智 能 系 统 学 报 Vol.16 No.6 2021 年 11 月 CAAI Transactions on Intelligent Systems Nov. 2021