Satellite Video Super-Resolution With Implicit Interpolation-Based Alignment and Spatiotemporal Fusion

Satellite video super-resolution (SVSR) is a critical task in remote sensing, aiming to reconstruct high-resolution videos from low-resolution satellite observations. Compared to natural-scene video super-resolution, SVSR faces unique challenges, such as weak texture regions, large-scale interframe displacements, and complex motion patterns, which limit the performance of existing methods in terms of alignment accuracy, detail restoration, and computational efficiency. To address these issues, we propose implicit interpolation-based alignment and spatiotemporal fusion satellite video super- resolution (ISSVSR), a lightweight SVSR network tailored for remote sensing scenarios. First, we introduce an implicit interpolation mechanism into the classical flow-guided alignment framework and design the implicit interpolation-based alignment module to replace traditional warp methods based on explicit interpolation. This enhancement improves alignment accuracy and preserves more high-frequency details under complex motion conditions. Second, we design the spatiotemporal fusion module, which combines temporal and spatial attention mechanisms for more effective multiframe feature aggregation, thereby improving reconstruction quality. Finally, by combining lightweight architecture and module-level optimizations, we significantly reduce the computational complexity of the model. Extensive experiments demonstrate that ISSVSR achieves a better balance between reconstruction accuracy, perceptual quality, and computational efficiency, verifying its effectiveness in SVSR tasks.