Satellite video super-resolution via frequency domain decomposition and multidomain attention fusion

With the rapid advancements in remote sensing technology and imaging devices, high-resolution satellite videos have become increasingly critical for dynamic earth observation tasks. However, effectively fusing spatiotemporal information in remote sensing videos remains challenging due to complex scene dynamics and multi-scale features. This paper proposes a novel satellite VSR approach based on frequency-domain decomposition and multi-domain attention fusion, aiming to improve the efficiency of spatiotemporal feature integration. The proposed method introduces a SpatioTemporal-Frequency Multi-domain Information Fusion Module (STF-MIFFM), which decomposes features into highfrequency and low-frequency components using high-pass and low-pass filters. A Frequency Decomposition Module (FDM) is designed to extract discriminative features from both frequency domains via adaptive spatial filtering, while a multi-domain attention mechanism adaptively weights and merges spatial, temporal, and frequency-domain information. Extensive experiments on the Jilin-1 satellite dataset demonstrate that the proposed method outperforms state-of-the-art VSR algorithms, achieving an average PSNR of 35.42 dB and SSIM of 0.9532 for ×4 scaling. Qualitative results further validate its superiority in reconstructing sharp edges and fine textures, making it particularly suitable for applications requiring high-precision remote sensing analysis.