基于空间邻域信息的高光谱遥感影像半监督协同训练  

Semi-supervised Tri_training Hyperspectral Image Classification Approach based on Spatial Neighborhood Information

作　　者：朱济帅 尹作霞 谭琨[1] 王雪[1] 李二珠[3] 杜培军[3] 

机构地区：[1]中国矿业大学江苏省资源环境信息工程重点实验室,江苏徐州221116 [2]济南市城市规划咨询服务中心,山东济南250099 [3]南京大学,卫星测绘技术与应用国家测绘地理信息局重点实验室,江苏南京210023

出　　处：《遥感技术与应用》2016年第6期1122-1130,共9页Remote Sensing Technology and Application

基　　金：国家自然科学基金项目(41471356);卫星测绘技术与应用测绘地理信息局重点实验室项目(KLAMTA-201410)

摘　　要：针对tri_training协同训练算法在小样本的高光谱遥感影像半监督分类过程中,存在增选样本的误标记问题,提出一种基于空间邻域信息的半监督协同训练分类算法tri_training_SNI(tri_training based on Spatial Neighborhood Information)。首先利用分类器度量方法不一致度量和新提出的不一致精度度量从MLR(Multinomial Logistic Regression)、KNN(k-Nearest Neighbor)、ELM(Extreme Learning Machine)和RF(Random Forest)4个分类器中选择3分类性能差异性最大的3个分类器;然后在样本选择过程中,采用选择出来的3个分类器,在两个分类器分类结果相同的基础上,加入初始训练样本的8邻域信息进行未标记样本的二次筛选和标签的确定,提高了半监督学习的样本选择精度。通过对AVIRIS和ROSIS两景高光谱遥感影像进行分类实验,结果表明与传统的tri_training协同算法相比,该算法在分类精度方面有明显提高。In the process of hyperspectral image classification using the tri_training algorithm, the labels of unlabeled samples have error labels when the amount of initial training samples is small.In this paper,we propose a novel tri_training based on spatial neighborhood information（tri_training SNI） to solve the problem for the tri training algorithm. Firstly, we choose three basic classifiers from MLR（Multinomial Logistic Regression）, KNN（k-Nearest Neighbor）, ELM（Extreme Learning Machine） and RF（Random Forest） classifier based on disagreement measure and disagreement-accuracy.These classifiers are redefined using unlabeled samples in the tri_training_SNI process.In detail,in each round of tri_training_SNI,unla beled samples are labeled for a classifier by the following two steps. samples is constructed under certain conditions that the other two cl Step assifi l：the first selection of unlabeled ers have the same labels.Step 2： spatial Neighborhood Information of initial training samples based on 8-neighborhood is applied in this pro posed approach to construct the secondary selection of unlabeled samples and the labels of unlabeled sam pies.Then the final classification results are produced via majority voting by the classification results of three classifiers.Experiments on two real hyperspectral data indicate that the proposed approach can effec- tively improve classification performance.

关 键 词：空间邻域信息(SNI) 协同训练 半监督 高光谱遥感影像分类 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]