摘要
遥感影像的云修复是改善影像质量、降低数据成本的一种重要手段.使用Landsat 8影像研究卷积神经网络在云修复中的应用,提出一种影像信息重建的新式网络结构——边缘辅助的门控卷积网络(edge-guided gated convolutional network,EGCN).该网络以多时相数据作为含云影像上被遮挡信息的辅助数据,主干网络为多时空门控卷积网络(spa-tial-temporal based gated convolutional network,STGCN),在多尺度特征融合模块引入一种改进的非局部(non-local,NL)模块——门控非局部(gated non-local,GNL)来替代传统的卷积层,并以边缘特征提取网络(edge network,ENet)为分支,从边缘信息层面进行特征引导.实验结果表明,GNL模块和ENet的加入均有助于提升云修复效果.
Abstract
Cloud removal of remote sensing images is one of the important technologies to improve data quality and reduce data cost.A novel network structure to reconstruct missing in-formation in images,which is called edge-guided gated convo-lutional network(EGCN),is put forward via the applications of convolutional neural network in cloud removal task using Landsat 8 images.This network uses multi-temporal data as auxiliary data for padding cloudy images.Spatial-temporal based gated convolutional network(STGCN)is used as the main trunk and an improved non-local(NL)block called gated non-local(GNL)is introduced to replace the traditional convolution layers.Besides,an edge network(ENet)is used as a branch to guide the feature from the edge information level.The experimental results show that GNL and ENet both bene-fit cloud removal task.
NETL
NSTL
万方数据