Robust SAR-assisted cloud removal via supervised align-guided fusion and bidirectional hybrid reconstruction

Cloud removal improves the quality of cloud-covered optical remote sensing images. The use of Synthetic Aperture Radar (SAR) images, which are known for their strong penetration capability and ability to reflect the geometrical features of ground objects, further enhances the reliability and accuracy of the reconstruction. However, significant variances between these two modalities, particularly misalignments from inadequate registration, disrupt the complementary information provided by the SAR. In addition, clouds in optical images often exhibit irregular shapes and uneven distributions, presenting substantial challenges for cloud removal. In this study, we propose a robust SAR-assisted cloud removal network. Specifically, the proposed align-guided fusion module optimally leverages the effective features from SAR, thereby enhancing the robustness of fusion. The bidirectional hybrid reconstruction module integrates global and local information via the parallel combination of convolution and transformer layers to ensure consistent filling in both the central and boundary areas. Consequently, our method achieves cloud removal with rich structure and accurate details. Experiments conducted on the SEN12MS-CR and GF-CR datasets demonstrate that our method outperforms six representative cloud removal methods, yielding an approximate gain of 0.4 in the peak signal-to-noise ratio and achieving satisfactory visualization performance. The code is available at https://github.com/phaeton2017/SAG-CR.