Advancing Geospatial Modeling with Prithvi-EO-2.0: Development and Validation of a Foundation Model for Earth Observation

Abstract

This thesis explores the development and evaluation of Prithvi-EO-2.0, a geospatial foundation model designed for Earth Observation (EO) applications. Leveraging Vision Transformer (ViT) architectures and Masked Auto-Encoders (MAEs), Prithvi-EO-2.0 was pre-trained on a global dataset carefully derived from the Harmonized Landsat-Sentinel (HLS) archive, with certain variants incorporating spatiotemporal embeddings to enhance its capacity for scalable and data-efficient geospatial modeling. The model was rigorously validated on a suite of downstream tasks, including two classification challenges from the Sen4Map dataset: land-cover and crop classification. Fine-tuning experiments revealed that the large Prithvi-EO-2.0 variants significantly outperformed both baseline methods and the earlier Prithvi-EO-1.0 model, achieving superior sample-weighted F1-scores. Notably, the 600M parameter ViT-H variant demonstrated remarkable performance retention under data-limited conditions, outperforming even Prithvi-EO-1.0 on the land-cover classification task. Data efficiency experiments highlighted that foundation models, particularly Prithvi-EO-2.0, better withstand reductions in training data compared to baseline models, showcasing their robustness in data-scarce scenarios. These findings underscore the transformative potential of Prithvi-EO-2.0 to address diverse EO challenges, enabling high-performance geospatial modeling with reduced labeled data and lower computational requirements.