一种基于核数据变换方法的遥感图像谱聚类算法  

A spectral clustering algorithm for remote sensing images based on kernel data transformation method

作　　者：赵海军[1] 陈华月[1] 崔梦天[2] ZHAO Haijun;CHEN Huayue;CUI Mengtian(School of Computer,China-West Normal University,Nanchong 637009,China;School of Computer Science and Technology,Southwest University for Nationalities,Chengdu 610041,China)

机构地区：[1]西华师范大学计算机学院,南充637009 [2]西南民族大学计算机科学与技术学院,成都610041

出　　处：《林业工程学报》2025年第2期130-137,共8页Journal of Forestry Engineering

基　　金：四川省自然科学基金(2022NSFSC0536);国家自然科学基金(12050410248)。

摘　　要：随着遥感图像在各行各业的日益广泛应用,遥感图像的处理变得愈来愈重要。为了实现谱聚类算法应用于林业工程中的遥感图像处理,本研究提出了一种基于核数据变换和角距离度量的谱聚类新算法。首先,通过对基于多变量核特征提取的一般核熵成分分析法的分析,并运用信息论概念和核密度估计密切相关的瑞利二次熵,提出了最佳特征提取和无监督降维方法,即最佳核熵成分分析法。它根据类或聚类信息方面的数据结构,采用一个额外的旋转,使得成分之间的独立性最大化;在这些成分中最佳地捕捉数据的高信息势部分,直接找到关于保留成分的数量的最大化信息势的基,以确保得到的解比标准的核熵成分分析得到的解保留更多(或相等)的信息势;并提出了采用梯度上升法来求解最佳核熵成分分析优化问题,具体实现是采用了一种简单的提前终止方案,以确保梯度达到一个额外迭代不会显著修改成本函数的区域。其次,通过对最佳核熵成分分析变换和样本外扩展的分析,构建了一种基于角距离度量的谱聚类算法,它采用角距离度量的核k-均值聚类目标,而不是采用基于欧氏距离的度量。优化过程采用最佳核熵成分分析空间中的角距离,以保证收敛到局部最优,从而实现图像的聚类。采用多光谱卫星图像的实验结果表明,本研究提出的谱聚类算法不仅适用于遥感图像的云筛选问题,而且相比目前其他先进的聚类算法有更好的分类性能。With the widespread application of remote sensing images across various industries,the processing of image has become increasingly important.To enable the application of spectral clustering algorithm to remote sensing image processing in the forestry engineering,this study proposes a new spectral clustering algorithm based on kernel data transformation and angular distance measurement.First,a new optimal feature extraction and unsupervised dimensionality reduction method,called the best kernel entropy component analysis(BKECA)method,is proposed.This method is developed by analyzing the general kernel entropy component analysis approach,which is based on multivariable kernel feature extraction.It incorporates concepts from information theory and Rayleigh quadratic entropy,which is closely associated with kernel density estimation.It uses an additional rotation based on the data structure in terms of class or cluster information to maximize the independence between components.The components optimally capture the high information potential parts of the data and the basis that maximizes this information potential concerning the number of retained components is directly determined.This approach ensures that the obtained solution retains as much or more information potential compared to what is achieved by the standard Kernel Entropy Component Analysis.A gradient-ascent approach is also proposed to solve the best kernel entropy component analysis optimization problem.The concrete implementation is that a simple early termination scheme is used to ensure that the gradient reaches a region where additional iterations do not significantly modify the cost function.In the second place,through the analysis of best kernel entropy component analysis transform and out of sample extension,a spectral clustering algorithm based on angular distance measurement is constructed,it adopts the kernel k-means clustering target of angular distance measure instead of the Euclidean distance measure.The angular distance in the best kernel en

关 键 词：遥感图像 非线性特征提取 概率密度函数 K-均值 瑞利熵 谱聚类 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]