Geostationary satellite observations provide continuous, high-temporal-resolution measurements that are essential for monitoring cloud evolution and convective systems, improving severe weather forecasting, and advancing our understanding of atmospheric processes. However, most artificial intelligence methods are developed for individual downstream tasks or rely on static satellite imagery, limiting their ability to fully exploit the rich spatial and temporal information contained in geostationary observations.
To address this challenge, we developed SatVision-Pix4D, a spatiotemporal foundation model trained on all-sky GOES-16 and GOES-17 Atmospheric Baseline Imager (ABI) observations. By learning the evolution of clouds and convection directly from multi-channel satellite imagery, SatVision-Pix4D produces transferable representations that can be efficiently adapted to diverse Earth science applications while reducing the need for large task-specific labeled datasets.
In this presentation we demonstrate the capabilities of SatVision-Pix4D through two downstream applications: three-dimensional cloud property reconstruction and convective system identification and nowcasting. Across both tasks, the pretrained model achieves competitive performance while requiring substantially less labeled training data than conventional deep learning approaches. These results demonstrate that a single pretrained model can support multiple atmospheric remote sensing applications, improving the scalability and efficiency of AI for Earth observation.
SatVision-Pix4D provides a reusable framework for cloud characterization, convective storm monitoring, satellite-based nowcasting, and many other science applications, offering a pathway toward operational AI systems that can accelerate Earth science research and enhance the use of current and future geostationary satellite missions.