【作者】 王茹； 聂银琚； 耿杰；

【Author】 Ru Wang;Yinju Nie;Jie Geng;School of Electronics and Information, Northwestern Polytechnical University;

【机构】 西北工业大学电子信息学院；

【摘要】 极化合成孔径雷达(Polarimetric Synthetic Aperture Radar, PolSAR)凭借其优异的观测能力受到越来越多的关注,极化SAR图像分类已成为遥感解译的重要任务之一。高分辨率极化SAR图像通常呈现场景复杂度高,地物目标空间分布差异大等特点,存在语义信息建模难度高、分类精度低的问题。为了解决上述问题,提出了一种基于边界约束长短程语义增强的极化SAR图像语义分割方法。使用卷积神经网络和Transformer对极化SAR图像进行长程和短程语义信息建模,以增强观测目标的语义信息表达。然后,设计长短程语义交互模块,合理平衡图像短程和长程语义信息对最终语义表征的贡献度。此外,构建边界约束语义感知模块,增强极化SAR图像同类目标语义表达的一致性,进一步提升极化SAR图像语义分割的精度。更多还原

【Abstract】 Polarimetric Synthetic Aperture Radar(PolSAR) has attracted increasing attention due to its excellent observation capabilities, and the classification of PolSAR images has become an important task in remote sensing interpretation. High-resolution PolSAR images typically exhibit characteristics, such as high scene complexity and large spatial distribution differences of targets, which lead to difficulties in modeling semantic information and low classification accuracy. To address these issues, boundary constrained long-short range semantic enhancement network is proposed for PolSAR image semantic segmentation. In this method, convolutional neural network and Transformer are used to model short and long range semantic information of PolSAR images, which is good to enhance the semantic information representation of targets. Subsequently, a long-short range semantic interaction module is designed to balance the contribution of short and long range semantic information to the final semantic representation. Additionally, a boundaryconstrained semantic perception module is constructed to enhance the consistency of semantic representation for similar targets in PolSAR images, and then the accuracy of PolSAR image semantic segmentation can be further improved.更多还原