电子学报 ›› 2020, Vol. 48 ›› Issue (1): 131-136.DOI: 10.3969/j.issn.0372-2112.2020.01.016

• 学术论文 • 上一篇    下一篇

自适应类别的层次高斯混合模型遥感影像分割

石雪, 李玉, 赵泉华   

  1. 辽宁工程技术大学测绘与地理科学学院遥感科学与应用研究所, 辽宁阜新 123000
  • 收稿日期:2018-12-10 修回日期:2019-08-11 出版日期:2020-01-25 发布日期:2020-01-25
  • 作者简介:石雪 女.1992年1月出生,辽宁人.2014年毕业于辽宁工程技术大学测绘工程系,现为硕博连读研究生,从事遥感影像处理有关的研究.E-mail:374636252@qq.com;李玉 男.1963年3月出生,吉林人.教授、博士生导师.1984年在西北电讯工程学院获得学士学位,1991年在东南大学获得硕士学位,2006年在瑞尔森获得硕士学位,2010年在滑铁卢大学获得博士学位.现为辽宁工程技术大学教授,主要从事遥感数据处理理论与应用研究.E-mail:liyu@lntu.edu.cn
  • 基金资助:
    国家自然科学基金(No.41301479,No.41271435);辽宁省自然科学基金(No.2015020090)

Remote Sensing Image Segmentation Based on Hierarchy Gaussian Mixture Model with Self-adaptive Number of Classes

SHI Xue, LI Yu, ZHAO Quan-hua   

  1. The Institute for Remote Sensing Science and Application, School of Geomatics, Liaoning Technical University, Fuxin, Liaoning 123000, China
  • Received:2018-12-10 Revised:2019-08-11 Online:2020-01-25 Published:2020-01-25

摘要: 为了实现自动确定类别数的高精度遥感影像分割,提出一种自适应类别的层次高斯混合模型(Hierarchical Gaussian Mixture Model,HGMM)遥感影像分割算法.提出算法采用多个高斯混合模型加权和定义HGMM,用于建模具有非对称,重尾和多峰等复杂特性的影像统计模型.采用期望最大化算法(Expectation Maximization,EM)求解模型参数.为了实现自动确定类别数,采用贝叶斯信息准则(Bayesian Information Criterion,BIC)求解最优类别数,其中惩罚项采用加权像素数定义.为了验证提出算法可行性和有效性,对模拟和全色遥感影像进行分割实验,并对分割结果进行定性和定量分析.结果表明HGMM具有准确建模复杂统计分布的能力,提出算法具有高精度和高效率,同时可自动确定最优类别数.

关键词: 高分辨率遥感图像分割, 层次高斯混合模型, 贝叶斯信息准则, 自适应类别, 期望最大化

Abstract: To accurately segment remote sensing image and automatically determine the number of classes,a Hierarchical Gaussian Mixture Model (HGMM) based remote sensing image segmentation algorithm with self-adaptive number of classes is proposed.In the proposed algorithm,HGMM is defined by the weighted sum of several Gaussian Mixture Models (GMM),which is used to model the asymmetric,heavy-tailed and multimodal distributions of image.Expectation Maximization (EM) is used to estimate the model parameters.Bayesian Information Criterion (BIC) is used to solve the optimal number of classes,where penalty term is defined by the weighted number of pixels.To test the feasibility and availability of the proposed algorithm,segmentation experiments are carried out on simulated and panchromatic remote sensing images.Segmentation results are analyzed qualitatively and quantitatively,which show that HGMM can accurately model the complicated statistical distributions.The proposed algorithm can obtain high accuracy and has good efficiency,as well as determine the optimal number of classes.

Key words: high resolution remote sensing image segmentation, hierarchical Gaussian mixture model, Bayesian information criterion, self-adaptive number of classes, expectation maximization

中图分类号: