From Spatial to Spatio-Temporal Learning: An Architectural Review of Prithvi-EO-1.0 and Prithvi-EO-2.0

This paper reviews the architectural evolution of the Prithvi series of geospatial foundation models, particularly focusing on the transition from spatial to spatio-temporal representation learning in Earth observation.Prithvi-EO-1.0proposed a 100M-parameter Vision Transformer with masked autoencoder pretraining on large-scale multispectral images, allowing for transferable spatial representations with limited supervision.Upon this foundation, Prithvi-EO-2.0increases model capacity to 300M-600M parameters and incorporates temporal and geographic embeddings to capture time-series patterns in global datasets.With 4.2 million multi-temporal samples in the training set, Prithvi-EO-2.0demonstrates improved robustness and generalization capabilities over varied geographic areas and spatial scales.Benchmark evaluations demonstrate steady performance improvements of up to 8% relative to Prithvi-EO-1.0 on a variety of downstream Earth observation tasks such as flood mapping, wildfire detection, and crop segmentation.