Accurate and realistic nowcasting of geostationary (GEO) weather satellite observations is critical for improving early warning capabilities for natural hazards such as heavy rainfall and wildfires. This study proposes a deep learning framework, Simpler Yet Better Video Prediction-Generative Adversarial Network (SimVP-GAN), to advance the nowcasting of Level-1 infrared (IR) observations up to 6 hours from the advanced meteorological imager onboard the GEO-Kompsat-2A satellite. The proposed framework couples a modified SimVP generator, in which shortcut connections are removed to improve long-term prediction stability, with a three-dimensional convolutional PatchGAN discriminator designed to capture spatiotemporal characteristics. Comprehensive evaluations across nine IR channels against established convolutional, generative, and recurrent baselines reveal distinct performance variations intricately aligned with the physical characteristics of the spectral bands. Specifically, standalone convolutional models excelled in smoother water vapor channels, where power spectral density analysis elucidated that adversarial training introduces unnecessary high-frequency noise and a detrimental over sharpening effect. Conversely, recurrent architecture achieved the lowest quantitative errors in intermediate channels dominated by continuous advection but struggled to maintain structural fidelity in complex window channels. In contrast, SimVP-GAN demonstrated superior capability in primary window and longwave absorption bands, successfully enforcing structural integrity against localized, non-linear thermodynamic changes. These findings highlight the absolute necessity of exploring channel-adaptive loss weighting strategies and provide critical insights into the domain-specific applicability of convolutional, generative, and recurrent models for operational satellite nowcasting.



Keywords : Modeling, Generative adversarial networks, Satellites, Measurement, Device-to-device communication, Architecture, Computer architecture, Grounding, Timing, Licenses