【摘要】 陆地探测一号（LT-1） 01组A/B双星作为中国首个民用L波段分布式合成孔径雷达（SAR）卫星系统，有望进一步推动中国雷达数据产品的应用普及化。LT-1主要支持的双极化模式能够保持高分辨率和大幅宽成像，然而与全极化观测系统相比，LT-1缺失了部分极化信息，不利于土地精细分类等应用。因此，本文面向LT-1常用的两种双极化组合，设计了多头多分支卷积神经网络及损失函数，旨在以双极化数据输入复原缺失的极化信息，从而实现LT-1全极化SAR数据重构。考虑到极化、空间特征互补性与幅度、相位数据特性差异，本文采用不同分支网络分别提取像元级的极化信息和局部感知野内的空间信息，并通过两个输出头独立预测幅度、相位值以降低输出阶段数据间的串扰风险，同时提出了顾及相位周期性的幅度/相位组合损失作为优化目标指导重构网络训练。基于LT-1在英国城镇区域获取的两景全极化影像，对本文方法进行了重构质量和地物分类实验验证。验证结果表明，本文方法的幅度项预测误差为1—2 dB、相位项预测误差低于0.5 rad，优于参数规模相当的UNet模型重构精度，且重构数据与真实数据的统计分布和极化分解效果基本一致，全极化信息得以良好保持；与利用真实双极化、真实全极化数据的分类结果相比，利用本文方法生成重构全极化数据后再进行地物分类，总体精度与后者相当且比前者提高了约5%。本文初步验证了利用LT-1双极化数据兼顾高分辨率、大幅宽、全极化的可行性，可为SAR土地覆盖/利用分类等应用提供更为丰富的输入信息。更多还原

【Abstract】 The LuTan-1(LT-1) 01A/B satellite constellation, China’s first civilian L-band distributed Synthetic Aperture Radar(SAR) satellite system, is expected to further promote the widespread application of domestic radar data products. While LT-1’s dual-polarization mode supports high-resolution and wide-swath imaging, it lacks certain polarization information compared to full-polarization observation systems, which limits its applications such as land cover classification. To address this, this study designs a multi-head multi-branch Convolutional Neural Network(CNN) and a tailored loss function for two commonly used dual-polarization modes of LT-1. The goal is to restore missing polarization information from dual-polarization inputs, thereby achieving full-polarization SAR data reconstruction. Considering the complementary nature of polarization and spatial features, as well as the distinct characteristics of amplitude and phase data, the network employs separate branches to extract pixel-level polarization information and local spatial features within receptive fields. Two independent heads predict amplitude and phase values to mitigate data crosstalk during output. Additionally, an amplitude/phase combined loss function incorporating phase periodicity is proposed to guide network training. Experiments conducted on two full-polarization SAR images acquired by LT-1 over urban areas in the UK demonstrate that, the proposed method achieves amplitude prediction errors of 1—2 dB and phase errors below 0.5 rad. These results surpass the reconstruction accuracy of a U-Net model with comparable parameter scale. The reconstructed data exhibit statistical distributions and polarization decomposition effects nearly identical to the ground-truth data, confirming effective preservation of full-polarization information. Furthermore, land cover classification experiments show that, classification maps using the reconstructed data achieve overall accuracy comparable to those using real full-polarization data, with a 5% improvement over dual-polarization based classification maps. This study preliminarily validates the feasibility of leveraging LT-1’s dual-polarization data to achieve high-resolution, wide-swath imaging while retaining full-polarization capabilities, providing richer input information for SAR land cover/use classification applications.更多还原