Reproduce “Better Together: Evaluating the Complementarity of Earth Embedding Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Thijs L van der Plas, Jacob JW Bakermans, Vishal Nedungadi, Gabrielė Tijūnaitytė, Marc Rußwurm, Ioannis N Athanasiadis
Earth embedding models capture overlapping but nonidentical information about locations, and combining them can improve downstream geospatial prediction beyond the strongest individual representation. Across six tasks, fusion helped on crop classification, land-cover, bioclimatic, and road-distance prediction, but offered little benefit for biomass or population density; the best gains included 76.7% R² for land cover, 89.2% for bioclimatic variables, and 59.0% for distance to roads.
Key insights
- The embeddings share substantial structure but are not interchangeable: pairwise CCA similarities range from 0.18 to 0.44 and CKA similarities from 0.42 to 0.60.
- Standalone weakness does not rule out downstream value; SatCLIP, for example, often improved land-cover predictions when fused with pixel-based embeddings despite weak average performance alone.
- Complementarity is target- and location-dependent. Adding all available models is not always optimal, and the best fusion can be a selected pair rather than the full ensemble.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.