海上溢油合成孔径雷达探测研究  

Study of Oil Spill Detection on SAR Images

作　　者：张彦敏[1] 徐卓[1] 旭锋 

机构地区：[1]中国海洋大学信息科学与工程学院,山东青岛266100

出　　处：《中国海洋大学学报（自然科学版）》2017年第2期106-115,共10页Periodical of Ocean University of China

基　　金：海洋公益性科研专项(201505002);国家自然科学基金项目(61501520)资助~~

摘　　要：在合成孔径雷达(SAR)图像中区分溢油和类油现象是溢油SAR探测的关键任务。实现该任务一般可分为3步:首先是提取油膜和类油膜的特征;然后筛选出有助于油膜和类油膜分类的关键特征;最后构造有效的分类器进行模式识别以便做出准确的判别。本文基于2011年蓬莱19-3油田溢油事故期间的15景SAR图像提取了138个油膜和类油膜样本的几何特征、背景特征、散射特征和纹理特征,将Fisher判别率和序列前向选择方法相结合,筛选出背景后向散射系数标准差、逆差距、能量和后向散射系数的均值四个关键特征组成的特征子集。在此基础上,为提高分类器的精度,将决策树模型CART算法与Bagging技术相结合,通过随机抽样给出多个维数相同大小的训练数据集从而建立多个决策树模型,以投票的方式对油膜和类油膜样本进行分类;最后,文中采用了五折和十折交叉验证方法对油膜和类油膜的分类结果进行评估,研究显示基于Bagging的决策树方法的油膜和类油膜分类的平均精度在85%以上,且将文中所用基于Bagging的CART决策树分类算法与经典CART决策树分类算法及神经网络分类算法相比较,发现本文所用方法的分类精度较高,从而表明了该方法在溢油SAR探测方面的可行性。Discriminating oil spills from lookalike phenomena is a crucial procedure in oil spill detection. To achieve this purpose, three-step approach is taken in general： firstly, features of oil spills and looka- likes are extracted; then, key features which are beneficial to the oil spill classification are screened out; finally, effective classifier is built and pattern recognition method is used to conduct classification. In this paper, 16 kinds of features which include geometric features, surrounding features, backscattering fea- tures and textural features of 138 oil spills and lookalikes are extracted from 15 SAR images. The images were acquired during Penglai 19-3 Platform oil spill accident in 2011. The 16 features are sorted from big to small based on the FDR value of the single feature. We find that the standard deviation of backscattering coefficient of the backgrounds has larger FDR value. Therefore, it can be selected as the first feature. Then, the forward selection method of sequential search method are used to determinate the optimal fea- ture subset for oil spill detection. We find that the standard deviation of backscattering coefficient of the backgrounds, inverse difference moment, energy and the mean value of backscattering coefficient can be selected as the optimal feature subset in this work. CART（Classification And Regression Tree） is a kind of binary decision tree which is helpful to improve efficiency of generating tree. However, the disadvan- tage of the decision tree classifier is that the variance of classification results is quite high. So the decision tree classifier is an unstable classifier. While for bagging algorithm, the only real training set in practice are divided into different training sets through resampling methods. And that is benefit for improving the unstable classifier performance. The bagging method based on decision-making tree combines massive cal- culation of single classifiers which is helpful to improve the accuracy of oil spill detection. In this paper, we co

关 键 词：合成孔径雷达 特征选择 油膜分类 CART决策树 BAGGING 

分 类 号：TP722.6[自动化与计算机技术—检测技术与自动化装置]