Earth embeddings are decomposable records of the land surface
Abstract
Ecological monitoring, carbon accounting and biodiversity assessment increasingly rely on annual Earth embeddings--learned vectors that compress a year of satellite observation for every ten-meter pixel on Earth, consumed in place of the imagery itself. What these vectors mean has not been established. Here we decompose two widely used embedding products and show that they contain individual ecological features, including fire, forest clearing and crop type, validated against independent reference data on events across six continents. The recovered concepts date abrupt canopy loss to the first post-disturbance growing season and resolve severity well enough to separate prescribed fire from wildfire. Small fires that recover quickly are detectable in one product but not the other, which instead separates crops more strongly.
Related
- Integrating Land-Cover Dynamics and Management Records for Recovery Monitoring of Hainan Eld's Deer in Datian National Nature Reserve.
- A Context-Aware 14-3-3 Binding Predictor Enabled by Data Augmentation, Multi-Scale Biological Features, and Protein Language Model Embeddings
- iNaturalist records reveal shifts in phenology of caterpillar (Lepidoptera) activity in response to warmer spring temperatures
- Correcting and georeferencing natural history occurrence records with calibrated uncertainty
- Some temporal and geographic records of Abatus species (Echinoidea, Schizasteridae) in Antarctic and Subantarctic environments.
BibTeX
@misc{young2026earth,
title = {{Earth embeddings are decomposable records of the land surface}},
author = {Robin Young and Srinivasan Keshav},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.32617},
eprint = {2610.32617},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.32617},
}
RIS
TY - UNPB
TI - Earth embeddings are decomposable records of the land surface
AU - Robin Young
AU - Srinivasan Keshav
PY - 2026
DA - 2026/10/23
PB - arXiv
DO - 10.48550/arxiv.2610.32617
UR - https://arxiv.org/abs/2610.32617
AB - Ecological monitoring, carbon accounting and biodiversity assessment increasingly rely on annual Earth embeddings--learned vectors that compress a year of satellite observation for every ten-meter pixel on Earth, consumed in place of the imagery itself. What these vectors mean has not been established. Here we decompose two widely used embedding products and show that they contain individual ecological features, including fire, forest clearing and crop type, validated against independent reference data on events across six continents. The recovered concepts date abrupt canopy loss to the first post-disturbance growing season and resolve severity well enough to separate prescribed fire from wildfire. Small fires that recover quickly are detectable in one product but not the other, which instead separates crops more strongly.
ER -