A team of scientists from NASA’s Goddard Space Flight Center has published the SatVision Top-of-Atmosphere Foundation Model for remote sensing. NASA said Wednesday that the geospatial foundation model can support the 3D cloud reconstruction and modeling for Earth and climate science studies.
NASA Scientists Develop SatVision TOA
The foundational model is trained using 100 million images from NASA’s Moderate Resolution Imaging Spectroradiometer, or MODIS, an instrument on the Terra and Aqua satellites. Terra and Aqua orbit the Earth from north to south across the equator in the morning and south to north across the equator in the afternoon, respectively, to capture images of the planet’s surface. The MODIS instruments collect data in 36 spectral bands in three different resolutions to enable scientists to better understand environmental processes around the world.
According to NASA, the team used “all-sky” conditions during the SatVision TOA’s pre-training period, which enabled the 3D cloud reconstruction and cloud modeling capability. Similar foundation models that currently exist are trained solely on cloud-free satellite imagery, which presents limitations and often requires atmospheric corrections.
The GSFC Data Science Group behind the foundational model believes that their work can advance atmospheric science, cloud structure analysis and Earth system modeling.
The SatVision TOA model architecture is available via GitHub while the model weights are on Hugging Face.