LoCA Combines Sentinel-1 and Sentinel-2 and Wins 68.8% of Comparisons Against SatCLIP

Coordinate encoders such as SatCLIP and GeoCLIP turn a (latitude, longitude) pair into a vector, which is then used to forecast wildfires, assess climate, map land use and count biodiversity. They have worked well, but until now nobody has properly explained why, or how to add a second data source to them. A group of authors led by Panagiotis Koromilas has proposed a theory and a working scheme called LoCA (Location Contrastive Alignment). Across 17 tasks, the model won 68.8% of all pairwise comparisons against five state-of-the-art competitors.

The two-dimension theorem

The main theoretical result is uncomfortable for the whole field. Any Lipschitz coordinate encoder produces an output with intrinsic dimension of at most 2, regardless of architecture and frequency bandwidth. After all, the input is two-dimensional. It follows that in this setup the contrastive loss stays strictly above its asymptotic optimum and never reaches the minimum.

The empirical picture looks similar. If you train an image encoder and a coordinate encoder together from scratch, the embedding covariance collapses to a very low rank. So why do SatCLIP-like models work? The authors explain it this way: the coordinate encoder learns against a frozen image encoder. The frozen encoder acts as a covariance regularizer and sets a high-rank target onto which the coordinate encoder "folds."

Two sensors, two requirements for the target

This trick is known for a single modality. With several, it gets harder: the same point on the map must receive multiple descriptors at once, and they may contradict each other. The authors formulated two conditions for the regularizer:

• Co-alignment across modalities. Adapters cannot reconcile independently pretrained encoders, so the targets must be aligned in advance.
• High effective rank. The coordinate encoder inherits the spectrum of its target, and a poor target will yield poor embeddings.

How LoCA works

The scheme has two stages.

• Stage 1. The Sentinel-1 (radar) and Sentinel-2 (optical) encoders are trained jointly with contrastive learning. The result is a co-aligned, high-rank multimodal target.
• Stage 2. The image encoders are frozen. The coordinate encoder and a tabular metadata encoder are trained as projection adapters with multi-view contrastive learning.

At inference, no imagery is needed; coordinates are enough. This matters in practice: fresh satellite scenes are not required, and the embedding is computed from a point on the map.

Against whom, and with what result

LoCA was compared with five state-of-the-art encoders:

• SatCLIP;
• GeoCLIP;
• GT-Loc;
• AlphaEarth;
• Copernicus-FM.

The tests covered 17 tasks across six areas: land cover, wildfire forecasting, climate, geophysics, socioeconomic indicators and biodiversity. The results:

• LoCA has the most task wins;
• it ranked in the top 2 on 11 of 17 tasks;
• it is the only model that beat each of the others on a majority of tasks;
• in pairwise task comparisons, it won 68.8%;
• its embeddings use 62% of dimensions, compared with 35% for SatCLIP.

The last figure is consistent with the theory. If the target's effective rank is higher, the embedding also uses the space more fully. Fewer dimensions sit idle.

Where this is useful

Geoembeddings are typically mixed into tabular and time-series models as additional features. For teams that forecast wildfires, crop yields or regional economic indicators, this is a way to get a "feel for the terrain" without storing and processing imagery. Sentinel-1 radar sees through clouds and shows moisture and surface structure, while Sentinel-2 optics provide the vegetation spectrum. Combining them at a single point gives a richer feature than either sensor alone.

The practical takeaway for developers: when adding modalities, watch the rank and consistency of the target first, not the size of the coordinate encoder. Code and pretrained weights are promised in the repository github.com/eleannavali/LoCA, and independent testing on your own tasks will show how well the 68.8% carries over beyond the authors' set of 17 tests.