LoCA turns satellite imagery into location embeddings
LoCA: Enabling Multi-Modality in Geospatial Location Encoders · arXiv:2610.32894
In plain words
Many maps and research tools need a useful description of a place from its latitude and longitude. Image-based descriptions can depend on which satellite image is available, so they do not always give a coordinate-only answer.
LoCA learns a location description from satellite images and geographic information. It first teaches radar and optical image encoders to agree, then trains a coordinate encoder to match them.
Across 17 tasks, LoCA beat each of five comparison models on most tasks shared with that model, though other models led on some individual tasks.
The research
A description based on a satellite image can change with the observation, and a precomputed map of descriptions is limited to its saved grid and observation period.
Can a system instead produce a useful representation for a location from its coordinates alone, while drawing on more than one kind of geographic information?
LoCA jointly trains radar and optical image encoders, then fits a coordinate encoder to their shared space so locations can be represented from coordinates alone.
LoCA first trains radar and optical image encoders together, then freezes them while adapting a coordinate encoder and a metadata encoder to their shared representation space.
The authors test the resulting fixed location embeddings on 17 tasks with small task-specific predictors, comparing them with five other models.
LoCA scores higher in most shared-task comparisons, with a 68.8% pooled win rate across pairwise model-task comparisons. This supports broad performance across the tested tasks, while individual results vary: other models lead on some biodiversity, wildfire and housing tasks.
The results cover a two-sensor, non-temporal system and coordinate-based evaluations that may be sensitive to nearby locations appearing in both training and test data. The authors identify evaluations that hold out whole regions as future work.
What the evidence shows
- The coordinate-only question
LoCA, or Location Contrastive Alignment, is a multimodal location encoder: it turns latitude and longitude into an embedding intended to represent a place. The motivating problem is that image-based representations depend on the particular satellite observation supplied, while precomputed grids cover only specified locations and observation periods.
The research asks whether a coordinate-based encoder can learn a useful representation from multiple kinds of geographic information.
- A structural constraint
The paper proves a constraint on smooth coordinate encoders trained to align with image encoders. Their outputs cannot fill a higher-dimensional embedding sphere; under alignment, the resulting contrastive loss must remain above the unconstrained minimum.
This is a theoretical result under stated dimensionality and alignment assumptions. It motivates seeking a richer target for the coordinate encoder, rather than assuming that a larger output vector solves the geometry problem.
- Constructing a shared target
LoCA builds that target in two stages. It first jointly trains Sentinel-1 radar and Sentinel-2 optical image encoders so paired images share a high-rank representation space.
It then freezes those encoders and trains coordinate and metadata encoders as adapters to that space. At inference, the coordinate encoder can produce an embedding from latitude and longitude alone.
- Broad results, real exceptions
The authors compare fixed embeddings on 17 tasks, using small task-specific predictors. LoCA beats each of five competing models on a majority of the tasks shared with that model, with a 68.8% pooled win rate across pairwise model-task comparisons.
This indicates broad strength across the tested tasks, but not universal leadership: AlphaEarth leads on SatBird and wildfire forecasting, while GT-Loc and GeoCLIP stand out on California housing.
- What the control supports
A control tests the role of joint image training in the pipeline. On the reBEN land-cover classification task, the full model scores 54.7 micro-F1 points, versus 1.1 when Stage 1 is replaced with two independently pretrained image encoders.
The result supports the importance of the tested Stage 1 setup. Since that control also changes the encoders’ starting states, it does not isolate joint alignment as the only explanation.
Deep
- Original paper authors
- Panagiotis Koromilas, Eleanna Vali, Diogenis Antonopoulos, Eleni Saka, Dimitris Christofis
- Submitted to arXiv
- License
- arXiv non-exclusive