Abstract:

In recent years, Synthetic Aperture Radar（SAR） has been widely used in improving oil spill detection. However, due to its special imaging mechanism, dark spots caused by multiplicative coherent speckle noise and other physical phenomena have always affected the accuracy of oil spill detection. It is difficult to distinguish the oil film and oil-like film phenomenon on the image by using a single feature. To solve this problem, this paper proposes to use multi-feature fusion combined with deep Residual Network（ResNet） to distinguish the oil film from the fully polarized image oil-like phenomenon. In the experiment, the three polarization characteristics of the C-band：polarization scattering entropy（Entropy）, average scattering angle（Alpha） and single scattering characteristic value relative difference（SERD） are combined to form an optimized feature subset, and then a plurality of regions of interest are selected as the training set and test set of the ResNet network on the feature map corresponding to the determined three polarization features. Finally, the training set used in the experiment in this paper is composed of 3 600 crude oil samples, 3 600 biological oil samples and 3 600 emulsified oil samples （a total of 10 800）. The test set is composed of 600 crude oil samples, 600 bio-oil samples and 600 emulsified oil samples （a total of 1 800）, and the final classification accuracy is 97.56%. Finally, using the same experimental data, the VGG and AlexNet classification algorithms that are also deep learning are used to classify the oil film and oil-like film, and compared with the ResNet algorithm classification results. In order to reduce the phenomenon of overfitting and obtain more reliable experimental results,[K]-cross validation and ROC curve experiments are also conducted in this paper. The results show that the algorithm proposed in this paper is effective.

Key words: full polarimetric SAR images, multi-feature fusion, oil film, oil-like film, deep residual network

摘要：

近年来，合成孔径雷达（SAR）在改善溢油检测方面得到了广泛应用。然而，由于其特殊的成像机理，乘性相干斑噪声和其他物理现象引起的暗斑一直影响着溢油检测的精度。单独使用一种特征很难对图像上的油膜和类油膜现象进行区分，针对这一问题提出了利用多特征融合结合深度残差网络（ResNet）的方式来区分全极化图像上的油膜和类油膜现象。实验中将C波段的三种极化特征：极化散射熵（Entropy）、平均散射角（Alpha）和单次散射特征值相对差异度（SERD）组合在一起，形成一个优化特征子集，在确定的三种极化特征对应的特征图上选取多个感兴趣区域作为ResNet网络的训练集和测试集。该实验所用训练集由3 600个原油样本、3 600个生物油样本和3 600个乳化油样本（共计10 800个）组合而成。测试集由600个原油样本、600个生物油样本和600个乳化油样本（共计1 800个）组合而成，最终得到97.56%的分类精度。用同样的实验数据采用同是深度学习的VGG和AlexNet分类算法进行油膜和类油膜的分类，并与ResNet算法分类结果进行对比分析。为了减弱过拟合现象以及获得更可靠的实验结果，分别进行了[K]-交叉验证和ROC曲线实验。结果表明所提出的算法是有效的。

关键词: 全极化SAR图像, 多特征融合, 油膜, 类油膜, 深度残差网络