基于半监督鉴别分析的高分辨率SAR影像建筑区提取  

Extraction of Building Areas from High-resolution SAR Images based on Semi-supervised Discriminant Analysis

作　　者：李婷[1,2] 程博[1] 尤淑撑[3] 

机构地区：[1]中国科学院遥感与数字地球研究所数据深加工部,北京100094 [2]中国科学院大学电子电气与通信工程学院,北京100049 [3]中国土地勘测规划院遥感所,北京100035

出　　处：《遥感技术与应用》2016年第2期239-246,共8页Remote Sensing Technology and Application

基　　金：国家自然科学基金项目"高分辨率SAR图像典型地物目标样本特征提取和识别研究"(61372189);国家科技合作专项项目(2012DFA20930)资助

摘　　要：针对建筑物在城市化发展规划、地理国情信息系统更新、数字化城市以及军事侦察等方面的迫切要求,提出将半监督鉴别分析(Semi-supervised Discriminant Analysis,SDA)算法应用于高分辨率SAR影像的建筑区提取中,实现快速提取建筑区信息以及提高城市地物目标识别能力。以Radarsat-2影像和TerraSAR-X影像为实验数据,基于灰度共生矩阵计算影像的各种纹理特征;结合SDA算法进行特征提取,并以新特征作为大津法(Otsu)的输入提取建筑区;最后对分类结果进行后处理。实验结果与线性鉴别分析(Linear Discriminant Analysis,LDA)算法和局部保持投影(Local Preserving Projection,LPP)算法进行比较,结果表明:SDA算法具有较强的泛化能力,在先验类别信息较少时,适用于高分辨率SAR影像的特征提取,可以快速有效地提取建筑区信息。Considering urgent request of buildings information for urban development and planning, upda- ting of geographical situation information system, digital city and military reconnaissance, Semi-supervised Discriminant Analysis （SDA） algorithm was employed for extraction of building areas in high resolution SAR image to improve the ability to target recognition and achieve fast extraction of building areas in city district.Radarsat-2 images and TerraSAR-X images were used as the experimental data.A variety of texture features of images were calculated based on gray level co-occurrence matrix, and then the SDA algorithm was employed for feature extraction.The new feature was used as input of Otsu method to extract building areas.Finally, the post-processing of image classification was performed.A comparison of the recognition re sult of building areas between SDA algorithm,Linear Discriminant Analysis （LDA） algorithm and Local Preserving Projection （LPP） algorithm was made.It is concluded that SDA algorithm has stronger general- ization ability SDA algorithm is applicable to feature extraction of high resolution SAR image under less prior class information, quickly and effectively extract the information of building areas in city district.

关 键 词：合成孔径雷达 建筑区 流形学习 特征提取 半监督鉴别分析 

分 类 号：TP75[自动化与计算机技术—检测技术与自动化装置]