Abstract

Satellite imagery provides a critical lens for monitoring Earth’s dynamic systems, yet integrating multi-source, multi-temporal data into globally consistent, high-resolution representations remains a challenge. Traditional remote sensing vision models, which process patches or images as inputs, often struggle to capture fine-grained spatiotemporal-spectral relationships critical for downstream tasks like land classification, climate modeling, and change detection. We present a self-supervised framework leveraging Barlow Twins to train an Earth Foundation Model that outputs pixel-level representations from diverse satellite data sources. Preliminary results demonstrate that the resulting representation map encodes high-quality spatiotemporal patterns, outperforming other earth foundation models (including current SOTA model, Google’s Embedding Fields Model) in classification and regression tasks. By bridging multi-modal satellite data into a harmonized latent space, our approach unlocks new opportunities for monitoring planetary-scale processes with higher precision.