Multidiscriminator Supervision-Based Dual-Stream Interactive Network for High-Fidelity Cloud Removal on Multitemporal SAR and Optical Images

Optical remote sensing images have the advantages in clear visual characteristics and strong interpretability. Unfortunately, cloud coverage limits the quality and availability of optical images in practical applications. In contrast, Synthetic Aperture Radar (SAR) images provide all-day and all-weather imaging, which can serve as effective auxiliary information for cloud removal. Existing cloud removal methods are difficult to obtain high-fidelity cloud-free results due to the insufficient spectral and spatial information exploration in the multitemporal SAR and optical images. In this paper, we propose a multi-discriminator supervision-based dual-stream interactive network (MDS-DIN) for cloud removal. Specifically, we first design a dual-stream interactive learning module to take full advantage of the complementary information between multitemporal SAR and optical images. Moreover, we specially design an adaptive weight fusion module to adaptively allocate fusion weights to the dual-stream results by considering the discriminative features in spectral and spatial levels. In addition, multi-discriminator is employed to jointly optimize overall networks for high-fidelity cloud removal. Experiments on simulated and real data sets demonstrate the competitive performance of our proposed method.