Prithvi V1.0: Generalist Geospatial Foundation Model on Global HLS Data

Abstract

Significant advancements in the development of adaptable and reusable artificial intelligence (AI) models will greatly influence Earth science and remote sensing. Foundation models, which are pre-trained on extensive unlabeled datasets through self-supervision, can be fine-tuned for various downstream tasks using small labeled datasets. Additionally, the scientific community is increasingly interested in constructing generalist AI models that leverage multi-sensor data for Earth observation applications. Here, we present Prithvi, a transformer-based geospatial foundational model pre-trained on over 7 years of multispectral satellite imagery from the Harmonized Landsat Sentinel-2 (HLS) global dataset. We used approximately 4.2 million samples for training and 45,568 samples for validation. We utilize Land Use and Land Cover (LULC) classes to ensure representation of all land use categories, ensuring that >3,000 tiles cover approximately 60% of all urban regions. By sampling from each LULC class separately, we guarantee the inclusion of every class, using LULC entropy to enhance landscape diversity further. Additionally, we ensure that around 800 ecoregions are represented, with each having at least three tiles. This approach results in a dataset that is more than three times larger than the previous global version, significantly increasing global coverage and the diversity of landscapes and ecoregions. The pre-trained model along with the fine-tuning workflows, has been made publicly available as open-source contributions to the global Earth sciences community through Hugging Face.