STPNet: adaptive fourth-order PDE-inspired regularization for satellite video super-resolution

Abstract

Satellite video super-resolution (SVSR) is essential for improving the spatial resolution and detail fidelity of satellite imagery. However, challenges such as low-quality frames, complex motion, and imaging noise make temporal alignment and multi-frame feature fusion difficult. To address these issues, this paper presents a novel SVSR framework, STPNet, which leverages three complementary mechanisms to enhance reconstruction quality. Specifically, a partial differential equation-inspired learnable regularization module is introduced to suppress reconstruction artifacts while preserving critical structural details. A multi-order temporal alignment strategy is designed to fully exploit long-range temporal information for accurate multi-frame feature alignment. In addition, a high-frequency enhancement module is incorporated to explicitly model and enhance high-frequency components, thereby improving texture reconstruction. Experimental results show that STPNet achieves a peak signal-to-noise ratio of 40.77 dB on the SAT-MTB-VSR dataset and demonstrates effective cross-dataset transfer capability on the Jilin-189 dataset. Both quantitative and qualitative evaluations indicate that STPNet outperforms existing methods, validating its effectiveness and robustness in handling complex dynamic scenarios in SVSR tasks.