EBRNet: Lightweight Enhanced Bidirectional Recurrent Network for Satellite Video Super-Resolution

Efficiently aggregating spatial-temporal information is critical in satellite video super-resolution tasks. However, the limited receptive field of convolutional neural networks constrains their ability to learn global and diverse representations needed to aggregate long-range spatiotemporal information in large-scale satellite video scenes. To address this limitation, this paper proposes EBRNet, an enhanced bidirectional recurrent network for satellite video super-resolution. Specifically, we design a novel multi-branch spatial-frequency enhancement module that performs aggregation of diverse spatiotemporal information. The global feature enhancement branch adopts a dual-branch spatial–frequency feature enhancement strategy for global and long-range spatiotemporal aggregation, the energy–frequency enhancement branch adaptively models energy- and frequency-aware feature representations, and the auxiliary enhancement branch performs local spatial detail enhancement. The outputs of three branches are fused to capture diverse and complex spatiotemporal characteristics. Moreover, a second-order multiscale alignment is constructed to obtain rich and diverse offset information for satellite video content. Extensive qualitative and quantitative experiments on satellite video datasets demonstrate that EBRNet achieves state-of-the-art performance and improved robustness compared with competitive methods, by a 1.39 dB gain in terms of average PSNR than the second-ranked model, with fewer parameters (6.8M). Code will be available at https://github.com/dz0602pshr/EBRNet.