TY - JOUR
T1 - Thin cloud removal with residual symmetrical concatenation network
AU - Li, Wenbo
AU - Li, Ying
AU - Chen, Di
AU - Chan, Jonathan Cheung Wai
N1 - Publisher Copyright:
© 2019 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS)
PY - 2019/7
Y1 - 2019/7
N2 - Thin cloud removal is important for enhancing the utilization of optical remote sensing imagery. Different from thick cloud removal, the pixels contaminated by thin clouds still preserve some surface information. Therefore, thin cloud removal methods usually focus on suppressing the cloud influence instead of replacing the cloudy pixels. In this paper, we proposed a deep residual symmetrical concatenation network (RSC-Net)to make end-to-end thin cloud removal. The RSC-Net is based on the encoding-decoding framework consisting of multiple residual convolutional layers and residual deconvolutional layers. The feature maps of each convolutional layer are copied and concatenated to the symmetrical deconvolutional layer. We used real cloud-contaminated and cloud-free Landsat-8 data very close in time for both training and testing. The RSC-Net is trained to take cloudy images as input and directly produce corresponding cloud-free images as output with all the bands together except the cirrus band and the panchromatic band. Compared with other traditional and state-of-the-art deep learning based methods, the experimental results show that our method has significant advantages in removing thin cloud contaminations in different bands.
AB - Thin cloud removal is important for enhancing the utilization of optical remote sensing imagery. Different from thick cloud removal, the pixels contaminated by thin clouds still preserve some surface information. Therefore, thin cloud removal methods usually focus on suppressing the cloud influence instead of replacing the cloudy pixels. In this paper, we proposed a deep residual symmetrical concatenation network (RSC-Net)to make end-to-end thin cloud removal. The RSC-Net is based on the encoding-decoding framework consisting of multiple residual convolutional layers and residual deconvolutional layers. The feature maps of each convolutional layer are copied and concatenated to the symmetrical deconvolutional layer. We used real cloud-contaminated and cloud-free Landsat-8 data very close in time for both training and testing. The RSC-Net is trained to take cloudy images as input and directly produce corresponding cloud-free images as output with all the bands together except the cirrus band and the panchromatic band. Compared with other traditional and state-of-the-art deep learning based methods, the experimental results show that our method has significant advantages in removing thin cloud contaminations in different bands.
KW - Convolutional neural network
KW - Residual blocks
KW - Symmetrical concatenation
KW - Thin cloud removal
UR - http://www.scopus.com/inward/record.url?scp=85065761555&partnerID=8YFLogxK
U2 - 10.1016/j.isprsjprs.2019.05.003
DO - 10.1016/j.isprsjprs.2019.05.003
M3 - 文章
AN - SCOPUS:85065761555
SN - 0924-2716
VL - 153
SP - 137
EP - 150
JO - ISPRS Journal of Photogrammetry and Remote Sensing
JF - ISPRS Journal of Photogrammetry and Remote Sensing
ER -