Tied and Anchored Stereo Attention Network for Cloud Removal in Optical Remote Sensing Images

The cloud removal technique for optical remote sensing (ORS) images can fuse synthetic aperture radar (SAR) images to enhance texture detail. Nevertheless, the differential features between SAR and ORS images, alongside speckle noise inherent in SAR images, significantly impede the fusion's quality and the resultant cloud-free ORS image's texture clarity. To solve the described problems, a tied and anchored stereo attention network was proposed for cloud removal in ORS images. The tied attention was proposed to mitigate differential features and enhance fusion quality. Tied attention establishes connections with differential features between SAR and ORS images by leveraging similar texture features as ties. The anchored stereo attention was also proposed to extract global features by leveraging spatial extensibility in ORS images. Anchored stereo, window, and channel attention are parallel used to extract global, regional, and local features to solve the impact of speckle noise. Quantitative and qualitative results on the open dataset demonstrate that the proposed network can restore fine textures of ORS images in different cloud covers and ground scenes. The source code is publicly available at https://github.com/ningjin00/TASANet.