基于深度学习的高分辨率卫星遥感影像条带噪声去除  

Stripe noise removal in high resolution satellite remote sensing images based on deep learning

作　　者：高浩博 卜桐 李欣[1] 陆世东 钟慧敏[3] 崔林[3] GAO Haobo;BU Tong;LI Xin;LU Shidong;ZHONG Huimin;CUI Lin(School of Remote Sensing and Information Engineering,Wuhan University,Wuhan 430079,China;Hubei Provincial Research Institute of Land and Resources,Wuhan 430071,China;China Centre for Resources Satellite Data and Application,Beijing 100094,China)

机构地区：[1]武汉大学遥感信息工程学院,武汉430079 [2]湖北省国土资源研究院,武汉430071 [3]中国资源卫星应用中心,北京100094

出　　处：《遥感学报》2023年第3期610-622,共13页NATIONAL REMOTE SENSING BULLETIN

基　　金：国家重点研发计划(编号:2018YFA0605500);湖北省国土资源科研计划“省级国土资源卫星应用示范关键技术研究与系统研制”(编号:844-11)。

摘　　要：受到成像环境、硬件条件等因素的限制,高分辨率卫星遥感影像上普遍存在条带噪声的现象,其严重影响了影像的辐射质量和可用性。本文针对传统条带去除方法存在的适应性差、去噪效率低、依靠先验知识等不足,提出了一种基于深度学习卷积神经网络的条带噪声去除方法。本方法首先利用不同尺度的卷积层进行特征提取,然后对多尺度的特征图进行特征融合得到去噪底图,通过残差学习的方法在底图上预测存在的噪声分量,最后用噪声影像减去条带噪声分量实现噪声的去除。以模拟和真实获取的噪声影像为实验数据,将本文提出的方法与一些经典的去噪方法进行实验结果对比分析,实验结果表明本文提出的基于深度学习的条带噪声去除方法能够在保留影像地物细节的情况下,能以优异的速度达到最高的定量指标和最好的视觉效果,充分证明了本文方法的优越性。Affected by imaging conditions,data transmission,and other factors,stripe noise is common in satellite remote sensing images.It seriously restricts the quality and further use of images.In early studies,various denoising methods,such as statistics-based methods,filtering-based methods,and optimization-based methods,have been proposed to overcome the above problems.These proposed methods have achieved inspiring results in some aspects.However,they still suffer from poor adaptability,low denoising efficiency,and the need for prior knowledge.Therefore,stripe noise removal remains a challenging task.In this study,we take advantage of the convolutional deep network while considering the characteristics of the stripe noise image itself.A deep-learning-based method is proposed,which includes three parts:a feature extraction module,a feature fusion module,and a stripe denoising module.The feature extraction module uses the convolutional layer of the same channel with different strides to extract features.As a result,different-scale feature maps of the noisy image are obtained for the following feature fusion module.The feature fusion module upsamples different-scale feature maps.It fuses these upsampled feature maps through the element-wise addition method.Finally,a denoising network is used to predict the components of stripe noise.The stripe component is subtracted from the noise image based on predictions.Given the difficulties in obtaining real noise samples,the network is trained by simulation samples.Then,it is extended to denoise real images.Experiments on simulation and real images show the excellent performance of our network.In the quantitative assessment,the PSNR and the SSIM of our network when simulated images are used are higher than those of the four methods.In the visual assessment,our network performs well on homogeneous and nonhomogeneous objects.Our network denoises more efficiently and retains more details of ground features than traditional methods and other denoising networks.In real noise images,ou

关 键 词：高分影像 深度学习 条带噪声 卷积神经网络 特征融合 

分 类 号：P2[天文地球—测绘科学与技术]