Wiener Filter-Based Mamba for Remote Sensing Image Super-Resolution With Novel Degradation

In remote sensing image super-resolution (SR), the difficulty in acquiring precisely pixel-aligned high-resolution (HR) and low-resolution (LR) image pairs significantly hinders deep learning-based methods. Existing degradation models, typically relying on simplistic assumptions, often fail to capture the complex real-world degradations, leading to suboptimal SR performance. Moreover, existing SR networks face challenges in simultaneously reducing blur and noise while efficiently extracting global features, primarily due to high computational complexity. To address these issues, we introduce a blind SR framework featuring: 1) a novel high-order degradation model tailored for remote sensing, incorporating a Gaussian-based atmospheric turbulence kernel and Butterworth low-pass filter to simulate real-world degradations more accurately; and 2) a Wiener filter-integrated Mamba architecture (WMSR) that leverages minimum mean squared error constraints for frequency-domain optimization and state-space modeling for linear-complexity global dependency capture. Experimental results demonstrate that the proposed WMSR achieves a PSNR gain of 0.1194 dB over the suboptimal method in the 2× SR task on the UCMerced dataset with only 62% of the parameters compared to the baseline model. The proposed method demonstrates superior performance over existing solutions in blind SR tasks without LR images.