结合FCN和多特征的全极化SAR土地覆盖分类

Land cover classification of fully polarimetric synthetic aperture radar with fully convolution network and multi-feature

针对极化合成孔径雷达(PolSAR)影像地物分类特征表征性弱,以及传统全卷积网络(FCN)分类精度较低、效果差的问题,该文提出了一种结合FCN和多特征的全极化SAR土地覆盖分类算法。首先,根据PolSAR影像和极化目标分解获取散射特征参数构建特征空间,利用主成分分析(PCA)对特征空间实现降维,以优化特征组合;接着,以SegNet建模思想为基础,在网络中层嵌入多层多尺度非对称卷积单元(MACU)结构,并在中层添加代价函数构建双代价收敛(DC)结构,基于此设计了DC-MA-FCN网络;然后,以优化后的特征组合为输入,通过DC-MA-FCN网络进行多层自主学习训练网络,并利用训练好的网络进行PolSAR影像初始分类;最后,组合DC-MA-FCN网络分类结果和形态学方法实现最终分类。该方法对两地区的PolSAR影像进行取样和试验,并使用多种评价指标定量分析,表明了算法的可行性和有效性。

Aiming at the problem that the object classification in polarimetric synthetic aperture radar(PolSAR)images is week in feature representation and the traditional fully convolution network(FCN)has low accuracy and poor effect on classification,a land cover classification of fully polarimetric SAR with FCN and multi-feature algorithm was proposed.Firstly,the feature space was constructed based on the PolSAR image and target decomposition to obtain the scattering feature parameters,and the feature space was reduced by principal component analysis(PCA)to optimize the feature combination.Then,based on the SegNet network modeling idea,the multi-scale asymmetric convolution unit structure and cost function were embedded in the network layer to design the DC-MA-FCN network.Using the DC-MA-FCN learn the multi-level feature autonomously of optimized feature combination,the initial classification results were obtained.Finally,the DC-MA-FCN network classification results and morphological methods were combined to achieve final classification.The method sampled and tested PolSAR images from two regions,and used a variety of evaluation indicators to quantitatively analyze,which showed the feasibility and effectiveness of the algorithm.