波段聚类和改进递归滤波的高光谱图像分类  

Band Clustering and Improved Recursive Filtering for Hyperspectral Image Classification

作　　者：渠慎明[1,2] 刘煊 梁胜彬[1] QU Shenming;LIU Xuan;LIANG Shengbin(School of Software,Henan University,Kaifeng,Henan 475001,China;Institute of Intelligence Networks System,Henan University,Kaifeng,Henan 475001,China)

机构地区：[1]河南大学软件学院,河南开封475001 [2]河南大学智能网络系统研究所,河南开封475001

出　　处：《遥感信息》2021年第3期25-31,共7页Remote Sensing Information

基　　金：河南省科技发展计划项目(212102210538);河南大学研究生教育创新与质量提升计划项目(SYL20040121)。

摘　　要：针对高光谱图像分类任务中的Hughes现象及噪声问题,提出了联合波段聚类和改进递归滤波的高光谱图像分类方法。首先,利用相对熵对高光谱图像的光谱波段进行K-means聚类,对聚类后的光谱波段进行高斯滤波,得到模糊化图像,将其作为递归滤波的引导图像;然后,对聚类后的光谱波段递归滤波处理,从而增强高光谱图像的轮廓特征;最后,利用支持向量机对递归滤波后的特征图像进行分类。在2个真实数据集上的实验结果表明,该方法降低了高光谱图像的维度,去除了噪声并阻止了信息跨越强边缘传播,与传统高光谱图像分类方法相比,提高了分类精度。Aiming at the Hughes phenomenon and noise problem in hyperspectral image classification,in this paper,a hyperspectral image classification method based on band clustering and improved recursive filtering is proposed.Firstly,the spectral bands of hyperspectral images are clustered by K-means using relative entropy,and then,Gaussian filtering is applied to the clustered spectral bands to get the blurred image,which is used as the guide image of recursive filtering.Next,the clustering spectral bands are filtered recursively to enhance the contour features of hyperspectral images.Finally,support vector machine is used to classify the feature image after recursive filtering.Experimental results on two real datasets show that the proposed method can reduce the dimensions of hyperspectral images,remove noise and prevent information from spreading across strong edges.Compared with traditional hyperspectral image classification methods,the proposed method improves the classification accuracy.

关 键 词：相对熵 K-MEANS聚类 高斯滤波 递归滤波 高光谱图像分类 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]