一种用于遥感影像信息提取的改进Relief算法  

An Improved Relief Algorithm for Remote Sensing Image Information Extraction

作　　者：王宇豪 曹红新 秦增忍 王帆 朱镇 WANG Yuhao;CAO Hongxin;QIN Zengren;WANG Fan;ZHU Zhen(China Energy Engineering Group Shaanxi Electric Power Design Institute Co,Ltd.,Xi’an 710054,China)

机构地区：[1]中国能源建设集团陕西省电力设计院有限公司,西安710054

出　　处：《遥感信息》2025年第1期71-78,共8页Remote Sensing Information

基　　金：中国能源建设集团陕西省电力设计有限公司重大科技项目(K2022-10)。

摘　　要：针对Relief算法的特征权重受样本随机性影响较大、不适用于多类别间的特征选择且无法自动确定特征阈值等不足,提出了一种改进的Relief算法,用于面向对象的遥感影像信息提取。新算法基于正态分布改进Relief算法的样本抽样策略和特征权重更新方式,提出特征组合评判指标Cω用于多类别间特征选择,最后根据高斯混合模型自动确定特征的阈值。基于陕西省周至县地区无人机影像和湖北省巴东县地区高分六号影像的实例测试结果表明,IRelief算法不仅能有效对特征进行降维和精选,还能够自动计算特征在分类时的阈值,总体精度、Kappa系数以及其他精度评价指标均优于Relief和ReliefF算法,是一种更优的特征选择算法。Aiming at the shortcomings of Relief algorithm,such as feature weights are greatly affected by sample randomness,not applicable to feature selection among multiple categories,and unable to determine feature thresholds automatically,an improved Relief algorithm is proposed for object-oriented remote sensing image information extraction.The new algorithm improves the sample sampling strategy and feature weight updating method of Relief algorithm based on normal distribution,and then proposes the feature combination judging index Cωfor feature selection among multiple categories,and finally determines the threshold value of features automatically based on Gaussian mixture model.Example test results based on UAV images in Zhouzhi county area of Shaanxi province and GF-6 images in Badong county area of Hubei province show that IRelief algorithm can not only effectively downscale and select features,but also automatically calculate the threshold value of features in classification,and the overall accuracy,Kappa coefficient,and other accuracy evaluation indexes are all superior to that of Relief and ReliefF algorithms,making it a better feature selection algorithm.

关 键 词：RELIEF 特征选择 面向对象 信息提取 正态分布 高斯混合模型 

分 类 号：P237[天文地球—摄影测量与遥感]