Abstract: [Purposes] Aiming at the problem that traditional polarization synthetic aperture radar image interpretation methods are difficult to accurately identify ground objects through limited physical models,this paper proposes an improved module based on atrous spatial pyramid pooling convolution.[Methods] The improved module consists of four parallel branches.These four branches process data through different convolution methods,reducing certain computational parameters while ensuring model performance.The improved four branches can effectively extract multi-scale features and complete the learning ability of key information through the convolutional block attention module.[Findings] The experimental results show that the average row accuracy rate of the improved method in water body classification is 91.2%,which is superior to the comparison network.It also performed well in vegetation classification,with an average row accuracy rate of 82%.[Conclusions] This indicates that by increasing the depth of the network,the learning ability of deep learning models for complex data can be enhanced.

Key words: PolSAR, semantic segmentation, convolutional neural networks, atrous spatial pyramid pooling convolution, convolutional block attention module