摘要
全极化合成孔径雷达(SAR)影像准确分类的一个重要前提是充分提取反映地物实际物理性质的特征。然而现有的全极化SAR特征提取算法和分类算法众多,却均存在各种各样的问题。无论极化特征提取方法还是分类算法,都会影响最终的分类精度。针对此问题,在多次实验的基础上,提出一种综合Pauli极化特征分解和支持向量机(SVM)的分类策略,简称为Pauli-SVM算法。首先通过经典的Pauli分解法提取全极化SAR影像的奇次散射、偶次散射、体散射等极化特征;并将这些信息组合成一个特征向量,然后引入高精度的SVM分类算法,选择训练样本后对全极化SAR影像进行监督分类。在江苏溧水和南京横溪镇两个研究区,以ALOS卫星的PALSAR影像为研究数据,进行监督Wishart分类算法、Freeman特征提取法结合SVM的分类算法、Yamaguchi特征提取法结合SVM的分类算法、Pauli-SVM算法的分类对比实验。结果表明,新提出的PauliSVM算法可以有效地提高分类的准确性。
An important precondition of accurate classification on Polarimetric SAR image is sufficient feature extraction which can reflect ground objects＇ physical attributes. However, there＇s many feature extraction and classi- fication algorithms for polarimetric SAR image, which have all kinds of problems. Both polarimetric feature extrac- tion methods and classification algorithms can affect the final classification accuracy. Aiming at this problem, on the basis of many experiments, a new classification strategy called Pauli-SVM for short is proposed by synthesizing Pauli polarimetric feature decomposition and SVM algorithm. Firstly polarimetric features extracted from classic Pauli decomposition including odd scattering, double scattering and volume scattering are used to form an eigenvec- tor. Secondly, after training samples are selected, supervised classification can be done on polarimetric SAR image by importing SVM algorithm which can get high classification accuracy. Finally, experiments of contrasting super- vised Wishart algorithm, SVM algorithm combined by Freeman feature extraction method, SVM algorithm combined by Yamaguchi feature extraction method and Pauli-SVM algorithm are done on two research plots including Lishui in Jiangsu province and Hengxi Town in Nanjing city with PALSAR image from ALOS satellite. The result turns out that new proposed Pauli-SVM algorithm can efficiently promote classification accuracy.
出处
《科学技术与工程》
北大核心
2014年第17期104-108,142,共6页
Science Technology and Engineering
基金
国家自然科学基金(41171323)
中国地质调查局地质调查工作项目(1212011120229)
江苏省自然科学基金(BK2012018)
地理空间信息工程国家测绘地理信息局重点实验室开放基金(201109)资助
作者简介
陈军（1978-），男，汉族，博士研究生。研究方向：全极化合成孔径雷达遥感图像处理、机器学习在遥感影像分析中的应用。
通信作者简介：杜培军E-mail：dupjrs@126．com。
