Microwave radiometer imagers (MWRIs) possess the technical advantages of all-weather, all-time, and large-scale observation, and are widely used in atmospheric remote sensing, becoming a core payload in meteorological monitoring and data assimilation systems. Currently, due to the inherent design characteristics of the payload, MWRI-acquired image data suffers from low resolution, significantly negatively impacting the accuracy of subsequent meteorological data assimilation and weather forecasting. Furthermore, in the detection channels where MWRI degradation is most severe, the original effective high-frequency information has undergone irreversible attenuation, making it difficult for traditional neural networks to reconstruct this high-frequency information. To address these issues, this paper proposes a spatial resolution enhancement method based on a generative adversarial network (GAN) architecture. The core idea of this method is to decompose the inverse problem of spatial resolution enhancement into multiple progressive sub-tasks, gradually restoring high-frequency information to improve image resolution. This paper compares this method with traditional methods (such as Wiener filtering) through a resolution chart experiment. The results show that the proposed method achieves higher resolution than traditional methods, with a peak signal-to-noise ratio (PSNR) reaching up to 39.5 dB and a structural similarity (SSIM) of 0.9780. This method can recover clearer land-sea boundaries and has stronger noise suppression capabilities.