Foundation models can change the landscape of remote sensing (RS) data analysis by enabling the pre-training of large computer vision models on vast amounts of remote sensing data. These models can be fine-tuned with a small amount of labeled training and applied to various mapping and monitoring applications. Because most existing foundation models are trained entirely on cloud-free satellite imagery, they are limited to ground-level applications or require atmospheric correction. SatVision-TOA is trained on all-sky conditions enabling applications involving atmospheric variables (e.g. cloud or aerosol).
SatVision TOA is a 3 billion parameter model trained on 100 million images. Moderate Resolution Imaging Spectroradiometer (MODIS). To our knowledge, this is the largest foundation model trained entirely on satellite remote sensing imagery. By including “all-sky” conditions during pre-training, the team included a range of cloud conditions often excluded in traditional modeling. It enables 3D cloud reconstruction and cloud modeling in support of Earth and climate science, offering significant enhancements to large-scale Earth observation workflows.
With an adaptable and scalable model design, SatVision-TOA can integrate diverse Earth observation datasets and reduce dependence on specific operational models. SatVision-TOA leverages one of the largest public datasets to capture global context and robust features. The model may have wide applications for investigating spectrometer data, including MODIS, VIIRS, and GOES-ABI. The team believes this will enable transformative advances in atmospheric science, cloud structure analysis, and Earth system modeling.
Model architectures and model weights are available. GitHub And A huggable faceFor more information, respectively, including a detailed user guide, see the respective white papers: SatVision-TOA: A Geospatial Foundation Model for Coarse-Resolution All-Sky Remote Sensing Imagery.