AlphaEarth is a foundation model that encodes every 10-meter land pixel as a 64-component vector, intended to support a variety of global mapping tasks. But many scientific applications operate at coarser spatial resolutions, where the full 64-dimensional representation may be more than necessary.
The paper, posted on arXiv, investigates this mismatch. Its title points to 'emergent sparsity': as you coarse-grain the spatial support, the number of effective variables that survive drops below 64. The abstract stops short of giving an exact count, but the framing suggests that most of the 64 dimensions are redundant at coarser scales.
This matters for researchers using AlphaEarth embeddings in downscaling, land-cover classification, or climate studies, where computational cost and interpretability are important. If the sparsity is confirmed, it could lead to simpler, cheaper representations without much loss of information.