This study proposes a novel framework, GeoGMI, to generate virtual 89 GHz horizontally polarized brightness temperatures (TBs) of the GPM Microwave Imager (GMI) on the geostationary Geo-Kompsat-2 A (GK-2 A) platform. The study aims to overcome the temporal sparsity of low-Earth orbit microwave (MW) observations during tropical cyclone (TC) events by leveraging the high-temporal resolution of GK-2 A. In contrast to conventional statistical techniques, GeoGMI employs a conditional generative adversarial network (cGAN) that integrates optimized input feature selection with a composite loss function to learn storm-scale spatial structures effectively. Quantitative comparisons demonstrate the clear advantage of the proposed deep learning framework: the multiple linear regression (MLR) baseline shows limited predictive skill, with a correlation coefficient (CC) of 0.531 and a root-mean-square error (RMSE) of 15.599 K, whereas GeoGMI achieves a higher CC of 0.647 and a lower RMSE of 14.429 K. These improvements indicate that GeoGMI more faithfully reconstructs physically meaningful deep convective cores, capturing nonlinear relationships that linear models do not represent. Nonetheless, despite the enhanced performance, the reconstruction of fine-scale TC structural features remains less accurate than that from direct instrument observations, highlighting the need for further refinement to resolve sub-storm-scale variability. However, the generated 10-min interval virtual GMI-like TB fields effectively bridge the temporal gaps between GMI overpasses and enable continuous MW-like monitoring of TCs. GeoGMI offers a promising approach to improve near-real-time monitoring of a TC's internal structures using only geostationary satellite observations, thereby enhancing nowcasting capability and reducing disaster-related risks.



Keywords : Microwave, Infrared, Brightness temperature, Tropical cyclones, Geostationary satellite, Deep learning