Earth embeddings encode fire, forest clearing, and crop type
Original title: Earth embeddings are decomposable records of the land surface
Two annual embedding products differ: one detects small recovering fires, while the other separates crops more strongly.
Annual Earth embeddings compress a year of satellite observation into learned vectors for every ten-meter pixel and are used in place of imagery for ecological monitoring, carbon accounting, and biodiversity assessment. The paper decomposes two widely used products and reports individual ecological features, including fire, forest clearing, and crop type, checked against independent reference data on events across six continents. Recovered concepts date abrupt canopy loss to the first post-disturbance growing season and separate prescribed fire from wildfire by severity. Small fires that recover quickly appear in one product but not the other, which instead separates crops more strongly.
Unconfirmed: The two embedding products are not named in the abstract.
- Embeddings stand in for a year of satellite imagery at ten-meter pixels.
- Validation uses independent reference events across six continents.
- Canopy loss is dated to the first post-disturbance growing season.
- Severity resolution is reported to separate prescribed fire from wildfire.
From the original article
1 min read at arxiv.orgEcological 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.
By Robin Young · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.