3D-attention and residual dense-based generative adversarial network for cloud removal

Remote sensing (RS) technology plays an increasingly dominant role in Earth observation. However, cloud contamination is a serious hinder in analysis of RS images. Aiming at the problems that the present cloud removal methods remain cloud residues and lose ground scene details in the restored image, We propose a 3D-attention and residual dense based generative adversarial network (3DA-RDGAN) to remove the cloud. We first introduce the residual dense block (RDB) into the generator, so it learns plentiful local characteristics via dense connection and integrats different levels of features via residual learning to restore ground object information. secondly, a 3D attention module (3DAM) is inserted to each RDB to infer the 3-D attention weights for the feature maps without adding parameters of the original network. Under the guidance of attention loss, 3DAM effectively helps the network pay more attention to the cloud areas and discover the difference between cloud and ground scenes. The proposed 3DA-RDGAN is tested on the open source RICE dataset, and its effect is compared with several other existing deep learning methods. The results indicate the superiority of 3DA-RDGAN in cloud removal for RS images.