Overview of the EO data infrastructure landscape

Researchers and practitioners must navigate a complex and rapidly evolving ecosystem of satellite data sources and access platforms. The choices made in the early stages (which mission to use, which processing level to start from, and which platform to retrieve the data through) have downstream consequences that are easy to underestimate. This section maps the landscape of available EO data, and discusses the practical considerations of accessing them.

NoteDefinition · Processing levels, analysis-ready, ML-ready

Processing levels. How far a product is from the raw signal. Level 0 is unprocessed instrument data; Level 1 adds time referencing, calibration and geolocation (e.g., Sentinel-1 GRD, Sentinel-2 L1C top-of-atmosphere reflectance); Level 2 holds geophysical variables at the resolution and location of the Level 1 source (e.g., Sentinel-2 L2A surface reflectance, Landsat Collection 2 Level-2); Level 3 maps variables onto a uniform space-time grid, often by compositing (e.g., the Sentinel-1 monthly mosaics) [1].

Analysis-ready data (ARD). Data processed to standardized physical specifications (calibrated, geometrically orthorectified, radiometrically corrected, with per-pixel quality masks and documented provenance) so that it can be used for analysis with a minimum of additional user effort.

ML-ready. Data formatted into standardized tensors (e.g., spatial or temporal patches) and organized in high-throughput layouts to optimize I/O performance during model training.

While institutional standards (e.g., NASA, ESIP) often use AI-ready as a broader term encompassing FAIR metadata, licensing, and documentation, ML-ready specifically denotes the structural formatting required for efficient GPU ingestion.

[1]

NASA Earthdata, “Data processing levels.” https://www.earthdata.nasa.gov/learn/earth-observation-data-basics/data-processing-levels.