地球信息科学学报 ›› 2022, Vol. 24 ›› Issue (3): 508-521.doi: 10.12082/dqxxkx.2022.210237

• 遥感科学与应用技术 • 上一篇    下一篇

基于自适应半监督模糊C均值的遥感变化检测

邵攀1,2(), 范红梅1,2,*(), 高梓昂1,2   

  1. 1.三峡大学湖北省农田环境监测工程技术研究中心, 宜昌 443002
    2.三峡大学计算机与信息学院, 宜昌 443002
  • 收稿日期:2021-04-29 修回日期:2021-07-05 出版日期:2022-03-25 发布日期:2022-05-25
  • 通讯作者: *范红梅(1997— ),女,湖北咸宁人,硕士生,主要从事遥感图像处理,变化检测研究。E-mail: fanhm@ctgu.edu.cn
  • 作者简介:邵 攀(1985— ),男,河南安阳人,博士,副教授,主要从事遥感图像处理,人工智能,信息融合研究。E-mail: panshao@whu.edu.cn
  • 基金资助:
    国家自然科学基金项目(41901341)

An Adaptive and Semi-Supervised Fuzzy C-means Clustering Algorithm for Remotely Sensed Change Detection

SHAO Pan1,2(), FAN Hongmei1,2,*(), GAO Ziang1,2   

  1. 1. Hubei Engineering Technology Research Center for Farmland Environment Monitoring, China Three Gorges University, Yichang 443002, China
    2. College of Computer and Information Technology, China Three Gorges University, Yichang 443002, China
  • Received:2021-04-29 Revised:2021-07-05 Online:2022-03-25 Published:2022-05-25
  • Supported by:
    National Natural Science Foundation of China(41901341)

摘要:

遥感变化检测是开展对地观测应用的关键技术之一,在城市研究、灾害评估以及资源调查等领域发挥着重要的作用。本文针对基于差分影像的遥感变化检测展开研究,提出一种基于自适应半监督模糊C均值(Adaptive and Semi-Supervised Fuzzy C-means, ASFCM)聚类的变化检测技术。① ASFCM根据“差分影像像素灰度值越大,则对应区域发生变化可能性越大”的性质,通过阈值技术自适应、自动地分析差分影像灰度直方图并将其划分为两部分:几乎可确定类别区域和不确定区域;② 将差分影像像素灰度值、空间上下文信息和几乎可确定类别区域像素的伪类别标签信息集成到模糊聚类过程,生成差分影像的模糊隶属度函数;③ 通过最大隶属度原则生成变化检测图。ASFCM通过半监督策略利用伪类别标签信息指导聚类过程,通过空间引力模型优化的模糊因子自适应地利用差分影像的空间相关性,能够得到更加准确的模糊隶属度函数和更优的变化检测结果。3组真实遥感数据的实验结果验证了ASFCM的有效性:ASFCM在Bangladesh数据上的kappa系数为0.9188,比其它方法提高3.5%~16.4%;在Madeirinha数据上的Kappa系数为0.9379,比其他方法提高2.18%~7.13%;在黑龙江数据上的Kappa系数为0.8696,比其它方法提高2.88%~22.02%。

关键词: 遥感, 变化检测, 模糊聚类算法, 自适应, 伪标签, 空间信息, 模糊因子, 半监督

Abstract:

Remote sensing change detection plays an important role in earth observation. This paper presents a novel Adaptive and Semi-Supervised Fuzzy C-Means (ASFCM) clustering algorithm for remote sensing change detection. The proposed ASFCM method integrates semi-supervised technique, spatial attraction model, and fuzzy factor into fuzzy clustering process, and thus can make full use of the characteristics of difference image. ASFCM consists of three main steps. Firstly, according to the property “the larger the gray values of the pixels in difference image, the more likely the pixels belong to the change region”, the difference image is partitioned into two parts-the nearly certain part and the uncertain part-by analyzing the histogram of difference image using thresholds, which are adaptively and automatically obtained. The pixels in the nearly certain part possess a high probability of belonging to the unchanged or changed class. Secondly, the gray values of pixels, the spatial-contextual information, and the pseudo labels of the nearly certain pixels are jointly exploited in the well-designed ASFCM algorithm to compute the membership functions of difference image. Finally, a change detection map is generated by applying the maximum membership principle to the computed membership functions. On the one hand, ASFCM uses the labeling information of the nearly certain pixels to guide the clustering process through semi-supervised scheme. As a result, the change information is enhanced and more accurate membership is achieved. On the other hand, ASFCM employs the spatial attraction model to improve the traditional fuzzy factor. The improved fuzzy factor is then used to integrate the spatial information adaptively, for reducing the effect of noise and outliers. Owing to the above two aspects, the proposed ASFCM method can obtain more accurate change detection result. Three real remote sensing datasets were used to evaluate the performance of the proposed ASFCM method. Eight state-of-the-art change detection methods were used as the comparative methods. The first dataset consists of two synthetic aperture radar images acquired by Envisat in April 2007 and July 2007. For the first dataset, the proposed ASFCM produced the lowest overall error rate (1.99%) and highest Kappa coefficient (91.88%). The second dataset is made up of two Landsat Thematic Mapper images acquired in July 2000 and July 2006. For the second dataset, ASFCM generated the lowest overall error rate (1.69%) and highest Kappa coefficient (93.79%). The third dataset contains two Landsat-7 ETM+ images acquired in August 2001 and August 2002. For the third dataset, ASFCM also gave the lowest overall error rate (2.71%) and highest Kappa coefficient (86.96%). Experimental results demonstrated the effectiveness of the proposed ASFCM method for change detection.

Key words: remote sensing, change detection, fuzzy clustering algorithm, adaptive, pseudo labels, spatial information, fuzzy factor, semi-supervised