基于元学习和K均值聚类的高分遥感影像变化检测  

Change Detection of High Resolution Remote Sensing Image Based on Meta Learning and K-means Clustering

作　　者：许自昌[1,2] XU Zichang(College of Earth Sciences,China University of Geosciences,Wuhan 430074,China;Fujian Geologic Surveying and Mapping Institute of Remote Sensing Center,Fuzhou 350011,China)

机构地区：[1]中国地质大学资源学院,武汉430074 [2]福建省地质测绘院遥感中心,福州350011

出　　处：《遥感信息》2023年第5期31-38,共8页Remote Sensing Information

基　　金：国家自然科学基金项目(41801324);福建省科技厅引导性项目(2017Y0055、2022H0009)。

摘　　要：为解决现有主流遥感影像变化检测方法在检测精度、自动化程度方面存在的局限性,提出一种基于元学习同/异质混合集成和K-means聚类的高分影像变化检测方法,可在较高检测精度下大幅缩减同/异质混合集成算法的运行时间。该方法首先以元学习为基础框架,选择同质集成的梯度提升树、随机森林和极端随机树作为元学习的初级学习器,快速重构原始样本的特征空间;然后利用K-means算法处理重构样本集,拟合多个逻辑回归次级学习器进行变化区域初检;最后采用超像元分割算法和空间邻域信息双重约束,滤除细小的“椒盐”碎斑。为验证该方法的有效性,选用两组不同地区的高空间分辨率遥感影像作为实验数据源。实验结果中,两组数据集上的Kappa系数分别为0.8492和0.8139,漏检率分别为0.1321和0.2152,误检率分别为0.1482和0.1017,处理耗时分别为65.217 s和700.441 s。结果表明,元学习算法结合K-means聚类的方法可有效提升变化检测精度,在算法效率方面也有良好的表现。In order to solve the limitations of existing mainstream remote sensing image change detection methods in terms of detection accuracy and automation,this paper proposes a high resolution image change detection method based on meta-learning with/without hybrid integration and K-means clustering,which can greatly reduce the running time of the same/with hybrid integration algorithm under high detection accuracy.Firstly,based on the framework of meta learning,this method selects the gradient lifting tree,random forest and extreme random tree as the primary learners of meta learning,and quickly reconstructs the feature space of the original samples.Then,the K-means algorithm is used to process the reconstructed sample set and fit multiple logistic regression secondary learners for the initial detection of change area.Finally,super pixel segmentation algorithm and spatial neighborhood information are used to filter out small“salt and pepper”spots.In order to verify the validity of the method,two groups of high spatial resolution remote sensing images from different regions are selected as the experimental data sources.In the experimental results,the Kappa coefficients on the two groups of data sets are 0.8492 and 0.8139,the missed detection rates are 0.1321 and 0.2152,the false detection rates are 0.1482 and 0.1017,and the processing time of the two data sets is 65.217 s and 700.441 s,respectively,showing that the meta learning algorithm combined with K-means clustering can effectively improve the accuracy of change detection.

关 键 词：变化检测 高分影像 元学习 K-MEANS 超像元分割 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]