基于多尺度注意力机制的PolSAR深度学习超分辨率模型  

PolSAR image deep learning super-resolution model based on multiscale attention mechanism

作　　者：林镠鹏 李杰[2] 沈焕锋[1,3] LIN Liupeng;LI Jie;SHEN Huanfeng(School of Resource and Environmental Sciences,Wuhan University,Wuhan 430079,China;School of Geodesy and Geomatics,Wuhan University,Wuhan 430079,China;Hubei Luojia Laboratory,Wuhan 430079,China)

机构地区：[1]武汉大学资源与环境科学学院,武汉430079 [2]武汉大学测绘学院,武汉430079 [3]湖北珞珈实验室,武汉430079

出　　处：《遥感学报》2024年第9期2362-2371,共10页NATIONAL REMOTE SENSING BULLETIN

基　　金：湖北珞珈实验室开放基金(编号:220100041);国家自然科学基金(编号:62071341,42301417)。

摘　　要：全极化合成孔径雷达影像(PolSAR)可提供丰富的极化信息,但成像系统限制使其空间分辨率受到制约。为解决此问题,本文基于深度学习框架,提出一种基于多尺度注意力机制的超分辨率重建网络(MSPSRN),通过对低分辨率PolSAR影像进行分辨率增强,生成高分辨率的PolSAR影像。在该模型框架下,采用多尺度注意力模块对不同尺度下的地物目标进行特征提取;提出联合式与分离式两种内嵌方式,在模型中嵌入通道注意力与空间注意力,利用注意力机制的权值重校准特性,增强PolSAR影像的极化信息与空间信息;引入残差信息蒸馏机制,提取判别性特征并对模型参数进行压缩;提出自适应损失函数对网络训练过程进行约束以提升模型的数值拟合能力以及边缘信息保持能力。最后,采用RADARSAT-2卫星的模拟数据与真实数据两个数据集对提出的方法进行验证。空间信息实验结果表明本文方法在目视结果与定量指标中均优于对比算法,具有更高的空间纹理细节重建精度与较低的重建误差;极化信息保持测试表明,本文方法可在提升空间分辨率的同时,有效保持PolSAR影像的极化信息。Fully Polarimetric Synthetic Aperture Radar imagery(PolSAR) can provide rich polarimetric information;however,given the limitations of the system's signal bandwidth and the physical size of the antenna,the spatial resolution of the SAR imaging system is restricted while acquiring multiple polarization information.To solve this problem,on the basis of the deep learning framework,this study proposes a multiscale attention-based PolSAR super-resolution network(MS-PSRN),which performs super-resolution reconstruction on the low-resolution full-polarimetric SAR images to generate the fully polarimetric SAR images with high spatial resolution.Under this superresolution reconstruction framework,this study uses a multiscale architecture to fully extract the feature information of objects at different scales.On this basis,the spatial attention mechanism and the channel attention mechanism are introduced to recalibrate the feature maps,which are used to enhance the reconstruction performance of spatial details and improve the ability to maintain polarization information,respectively.Two attention mechanism embedding methods,i.e.,joint and separated,are proposed to cope with the spatial size and quantity changes of the feature maps processed by the encoder and decoder.This study introduces a residual information distillation mechanism,extracts discriminative features through feature distillation,and compresses model parameters at the same time.In addition,the adaptive loss function is proposed to constrain the network training process and improve the model's numerical fitting ability and edge information preservation ability.In this study,the proposed method is verified by two datasets,namely,the simulated data and the real data produced by RADARSAT-2 images.The experimental results of spatial information show that the proposed method is superior to the comparison algorithms in terms of visual results and quantitative indicators and has higher texture detail reconstruction accuracy and lower reconstruction error.The polarime

关 键 词：遥感 全极化合成孔径雷达 超分辨率重建 深度学习 遥感 多尺度 注意力机制 

分 类 号：TP75[自动化与计算机技术—检测技术与自动化装置] P2[自动化与计算机技术—控制科学与工程]