GAIA: A Foundation Model for Operational Atmospheric Dynamics

TL;DR

GAIA is a hybrid self-supervised foundation model trained on 15 years of satellite imagery that captures atmospheric dynamics and improves performance on various atmospheric analysis tasks.

Contribution

This work introduces GAIA, a novel hybrid self-supervised model combining MAE and DINO, specifically designed for atmospheric data, enabling better transfer to downstream atmospheric tasks.

Findings

GAIA outperforms MAE-only baseline in atmospheric river segmentation.

GAIA improves tropical cyclone detection recall and early detection.

GAIA demonstrates robust gap-filling capabilities in satellite imagery.

Abstract

We introduce GAIA (Geospatial Artificial Intelligence for Atmospheres), a hybrid self-supervised geospatial foundation model that fuses Masked Autoencoders (MAE) with self-distillation with no labels (DINO) to generate semantically rich representations from global geostationary satellite imagery. Pre-trained on 15 years of globally-merged infrared observations (2001-2015), GAIA learns disentangled representations that capture atmospheric dynamics rather than trivial diurnal patterns, as evidenced by distributed principal component structure and temporal coherence analysis. We demonstrate robust reconstruction capabilities across varying data availability (30-95% masking), achieving superior gap-filling performance on real missing data patterns. When transferred to downstream tasks, GAIA consistently outperforms an MAE-only baseline: improving atmospheric river segmentation (F1: 0.58 vs 0.52), enhancing tropical cyclone detection (storm-level recall: 81% vs 75%, early detection: 29% vs 17%), and maintaining competitive precipitation estimation performance. Analysis reveals that GAIA's hybrid objectives encourage learning of spatially coherent, object-centric features distributed across multiple principal components rather than concentrated representations focused on reconstruction. This work demonstrates that combining complementary self-supervised objectives yields more transferable representations for diverse atmospheric modeling tasks. Model weights and code are available at: https://huggingface.co/bcg-usra-nasa-gaia/GAIA-v1.