基于多特征多核学习的全极化SAR溢油提取方法

全极化合成孔径雷达(Synthetic Aperture Radar,SAR)数据具有丰富的极化信息,能够提取出大量异构性特征。核学习方法在解决小样本、高维特征分类问题上具有优势,但异构特征对不同核函数具有响应差异。本文利用一种引入先验标签的多核学习方法进行全极化SAR的溢油信息提取,即基于分析结果对特征集进行遴选与组合,分别在每个特征组合中训练得到一个预备层核函数,以新获取的预备层核函数作为新的底层核函数,对全部特征进行学习分类。通过提取与分析溢油和海水的统计特征、物理散射特征和纹理特征,建立溢油全极化SAR特征谱,并利用引入先验标签的多核学习分类器进行溢油提取实验。结果表明,该方法能够利用全极化SAR多维异构特征的互补特性有效提高溢油分类提取精度。

Full polarimetric SAR oil-spill extraction method based on multi-feature and multi-kernel learning

Full polarimetric SAR data contains a wealth of polarization information, so a large number of heterogeneous features can be extracted from it. The kernel learning method has advantages in solving small-sample problems and high-dimensional feature classification, but heterogeneous features have different responses to different kernel functions. In this paper, we use a multi-kernel learning method based on an a priori label to extract oil-spill information from full polarimetric SAR data. Specifically, it selects and combines feature sets based on analysis results and trains a preliminary kernel function on each feature combination. This newly acquired preliminary kernel function is used as a new underlying kernel function to classify all the features. The full polarimetric SAR characteristic spectrum of an oil spill is determined by extracting and analyzing the statistical, physical-scattering, and texture features of the oil spill and seawater. We conducted an oil-spill extraction experiment using the above multi-kernel learning classifier. The results show that this method can effectively improve the precision of oil-spill classification by using the complementary characteristics of multi-dimensional heterogeneous features of full polarimetric SAR data.