Sparse autoencoders decompose Earth embeddings into sparse features
Original title: Decomposing Earth Embeddings with Sparse Autoencoders
On AlphaEarth, trained SAEs leave eight times less variance unexplained than a frozen random decoder.
Earth observation models such as AlphaEarth and Tessera produce opaque metre-resolution embeddings. The authors decompose them with sparse autoencoders and score reconstruction, concept alignment, and spatial scale against baselines, including an SAE with a frozen random decoder. Trained SAEs reconstruct better at matched sparsity, while sparsity alone explains finer spatial scale and narrower concept alignment. Land-cover sub-concept discovery and out-of-distribution detection are shown but not yet reliable for practical use.
- On AlphaEarth, 88% of features align with at most one of 20 concept groups.
- BatchTopK and Matryoshka carry 23 to 32% of variance within a 640 m patch.
- Raw embedding dimensions have a median alignment with 12 concept groups.
- A frozen random decoder shares sparsity properties but reconstructs poorly.
- AI models
- AlphaEarthTessera
From the original article
1 min read at arxiv.orgEarth observation foundation models produce global metre-resolution embeddings that underpin work across agriculture, ecology, hydrology, and beyond. Yet these embeddings are opaque: each dimension entangles unrelated physical concepts, which makes them difficult to interpret and their robustness hard to assess.
By Vitus Benson · 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.