Remote sensing video super-resolution via various real-world degradation modeling

Remote sensing video super-resolution (VSR) aims to restore high-resolution (HR) sequential images from their low-resolution (LR) counterparts. Due to the complex and diverse degradations inherent in real-world satellite and drone imaging, existing methods often perform suboptimally compared to their success on ideal or simulated data. To bridge this domain gap, we propose a novel VSR framework, termed RWD-VSR, tailored to tackle various real-world degradations in remote sensing videos. Specifically, we construct a comprehensive real-world dataset combining Unmanned Aerial Vehicle (UAV) videos with public satellite videos, systematically formulating nine authentic degradation models under non-ideal conditions. To handle these intricate artifacts, we employ a multi-scale vector-quantized generative adversarial network to implicitly learn a fused degradation manifold, synthesizing reliable pseudo LR-HR training pairs without relying on explicit mathematical assumptions. Furthermore, to restore sharp structural details from severely degraded sequences, we design an enhanced unidirectional recurrent network featuring dense cross-scale interactions. Extensive experiments demonstrate that RWD-VSR significantly outperforms state-of-the-art VSR models, achieving superior perceptual quality on both high ground sample distance satellite and low ground sample distance UAV datasets. The source code and dataset are available at https://shiaoo.github.io/RWD_VSR/ .