全卷积网络和条件随机场相结合的全极化SAR土地覆盖分类

针对经典全卷积网络(fully convolution network,FCN)分类精度低、效果差,以及传统的极化合成孔径雷达(PolSAR)土地覆盖分类方法未充分考虑地物散射特性的问题,提出了一种结合改进FCN和条件随机场(conditional random field,CRF)的全极化SAR土地覆盖分类算法。首先,利用Freeman分解和Pauli分解建模全极化SAR影像,同时提取各分解对应的散射特征,参考Freeman分解散射功率获取其主散射分量对应的主散射地物;同时,借鉴在图像分类领域中具有卓越表现的FCN-Vgg19-8s网络,考虑其高层卷积参数量大和低层卷积模型参数优化程度不足,通过在高层和中层分别构建多尺度卷积组和代价函数设计了FCN-MD-8s网络,保证对整体模型参数进行降维和优化;以Freeman分解散射机理特征为基准,采用级连式迁移学习结构,实现FCN-MD-8s网络的模型训练和测试;然后,根据主散射分量所对应的主散射地物,在各分量预测图中提取出主特征地物,得到分量地物分类结果,并将其进行叠加得到全局粗分类;最后,利用全连接CRF结合Pauli相干分解重建假彩色图,对全局粗分类进行全局像素类别转移获得细分类结果。通过对分类结果定性和定量分析,可知提出算法具有有效性和可行性。

Land cover classification of polarimetric SAR with fully convolution network and conditional random field

Aiming at the problems of low classification accuracy and poor effect in the traditional fully convolution network (FCN), and insufficient consideration on the scattering characteristics of ground object features in the traditional polarimetric synthetic aperture radar (PolSAR) land cover classification methods. To overcome this limitation, this paper proposes a land cover classification algorithm of polarimetric SAR with improved FCN and conditional random field (CRF). First of all, the Freeman and Pauli decompositions are used to model the full-polarimetric SAR image to obtain the scattering features of scattering mechanisms, and Freeman decompositions are referenced to obtain the main scattering object corresponding to the main scattering component. Learning from the FCN-Vgg19-8s network with excellent performance in the field of image classification, and considering the large amount of high-level convolution parameters and the insufficient optimization of low-level convolution model parameters, Then an improved FCN, named FCN-MD-8s, is designed though constructing multi-scale convolution group and cost function in the upper and middle layers based on FCN-Vgg19-8s to guarante dimensionality reduction and optimization of overall model parameters. Additionally, FCN-MD-8s network is trained and tested for scattering mechanisms from Freeman decomposition by Cascade-migration-learning structure. Afterwards, according to the main scattering feature corresponding to the main scattering component, the main feature object is extracted from each component prediction image to obtain a component classification result. The result of each component classification is superimposed to gain a global rough classification. Finally, the fully-connected CRF with false color image, which is visualized by Pauli coherent decomposition, is used to transfer full image information over global rough classification for fine classification. The qualitative and quantitative analyses of classification results demonstrate that the proposed algorithm has effectiveness and feasibility.