FAHRNet: Frequency-Aware Hybrid Refocusing Network for Remote Sensing Image Super-Resolution

High-resolution remote sensing imagery is essential for land-cover analysis, urban monitoring, environmental observation, and disaster-response mapping, yet its acquisition is constrained by sensor resolution, cost, atmospheric effects, and revisit frequency. This paper proposes FAHRNet, a Frequency-Aware Hybrid Refocusing Network for ×4 remote sensing image super-resolution. FAHRNet integrates Wavelet-Token Bidirectional LSTM Guidance (WT-BiLG), a Gated Row–Column Directional Mixer (GRDM), CNN local refinement, Haar-frequency correction, and gated feature fusion and is trained with Charbonnier reconstruction, Haar high-frequency, and Sobel edge losses without perceptual or adversarial supervision. Under the primary protocol, FAHRNet (C=128) achieves 29.0292 dB PSNR-Y on RSSCN7 and 29.3791 dB on UAVid. With the same composite objective, EDSR reaches 28.8457 dB and 29.1235 dB, respectively, confirming that objective alignment narrows the original gap while FAHRNet retains higher PSNR-Y on both datasets. On UCMerced, a same-objective, parameter-matched EDSR-CM attains 28.8026 dB versus 28.6254 dB for FAHRNet, but FAHRNet uses approximately 36.9% fewer standardized MACs and achieves higher CLIPIQA. Controlled GRDM tests show that the parallel formulation outperforms parameter-identical serial 1×9→9×1 mixing and a closely matched 9×9 depthwise control in PSNR-Y. Under Gaussian-blur and blur-plus-noise mismatch, FAHRNet retains higher absolute reconstruction quality than same-objective EDSR, although its relative degradation from the bicubic operating point is larger. A fixed HR-trained classifier yields 0.8636 accuracy and 0.8621 macro-F1 on FAHRNet reconstructions without domain fine-tuning. Overall, FAHRNet provides a competitive local–directional–frequency operating point with dataset-, objective-, metric-, and computation-dependent trade-offs.