0:00
/
0:00
Transcript

"EarthView: A Large Scale Remote Sensing Dataset for Self-Supervision"

Generated below podcast on this paper with Google's Illuminate.

EarthView combines 15 trillion pixels of satellite data to create better Earth monitoring models. 🌍

This paper introduces EarthView, a massive 15 tera-pixel dataset combining satellite imagery from multiple sources, designed specifically for self-supervised learning in remote sensing applications. It introduces EarthMAE, a modified Masked Autoencoder architecture optimized for Earth monitoring tasks.

→ Covers 437,682 km² with 2,967,663 high-resolution patches

https://arxiv.org/abs/2501.08111

This Paper 🛠️:

→ EarthView combines data from three major sources - Satellogic (1m resolution), NEON (hyperspectral), and Sentinel (multi-spectral).

→ The architecture incorporates distinct tokenizers for different data sources and special encodings for temporal and source information.

→ Dataset spans 5 years (2017-2022) with multiple revisits per location, enabling temporal analysis.

-----

Key Insights 💡:

→ Tube masking consistently outperforms random masking across tasks

→ Including temporal information significantly improves model performance

→ Combined Sentinel-Satellogic data yields better results than individual sources

→ Achieves consistent performance improvement over models pre-trained on Sentinel data alone

→ Shows linear performance scaling with dataset size, indicating room for further improvements

Discussion about this video