Downstream Applications on Prithvi: A Foundation Model for Global Geospatial Data

Abstract

Foundation models, which are deep-learning models pre-trained on large quantities of unlabeled data, have rapidly risen to prominence in the Earth and atmospheric sciences. Foundation models' power lies in their ability to rapidly fine-tune to new downstream tasks using small amounts of labeled data. In this presentation, we introduce three such tasks using the Prithvi geospatial foundation model: CO2 flux estimation, forest disease and infection detection, and land use/land cover classification. In the CO2 flux estimation task, remote sensing imagery is combined with meteorology from the Modern-Era Retrospective Analysis for Research and Applications, Version 2 (MERRA-2) to estimate CO2 uptake by vegetation, which is a crucial component of the climate system. The forest disease detection and land cover classification tasks are traditional remote sensing tasks using only satellite imagery, but nevertheless are expensive to carry out manually, providing the opportunity for AI to improve the efficiency and reduce costs of those products. In particular, tracking forest health is of growing importance as regional environments and climates respond to the changing global climate. In the land cover classification task, we ingest time-series from Sentinel-2 and evaluate the model on hierarchical classes, ranging from broad categories to detailed information, such as crop type classification, where the temporal component plays a crucial role. Our preliminary results indicate that global geospatial AI models, such as Prithvi, can accurately assist in these tasks.