In the field of remote sensing, the quest for enhanced spatial resolution remains paramount across numerous applications, even amid the surge of satellite imagery. Single Image Super-Resolution (SISR) addresses this need by improving the detail and clarity of lower-resolution (LR) images based on reference high-resolution (HR) images. While state-of-the-art (SOTA) deep learning models have been developed for this purpose, their reliance on synthetically generated LR image pairs from HR images raises concerns. Typically, these datasets are created through bicubic down-sampling or by modeling degradation with blur kernels and imaging noise, limiting the models’ ability to accurately model real-world scenarios. This study introduces a novel Real World Super-Resolution Generative Adversarial Network (RWSRGAN), utilizing an Activated Residual-in-Residual Network (ARRDNet) combined with a weighted loss function. The training has been done on the WorldStrat data, a real-world dataset containing LR and HR image pairs from Sentinel-2 and SPOT 6/7 satellites, respectively, with a resolution ratio of approximately 6. Experiments demonstrate that the proposed RWSRGAN achieves superior quantitative results compared to the existing SOTA ESRGAN, which lacks attention mechanism, and outperforms it in various distortion-based and perception-based metrics.