A location encoder that reads one satellite image now reads radar and optical: LoCA wins 68.8% of 17-task comparisons
Synopsis
Geospatial location encoders could previously align only to a single frozen image encoder, leaving no shared target when radar and optical sensors are combined. LoCA first jointly trains Sentinel-1 and Sentinel-2 encoders with a contrastive loss to produce a co-aligned, high-rank target, then freezes both and fits a location encoder and a metadata encoder to it. Across 17 downstream tasks it takes the most task wins and wins 68.8% of all per-task comparisons.
Interpretation
A location encoder can now align to radar and optical satellite imagery at once, not just a single optical source. Earlier methods such as SatCLIP and GeoCLIP freeze one pretrained image encoder and fit only the inverse map from coordinates to that fixed point cloud; two independently pretrained encoders are not co-aligned, so one location function cannot fit both. On 17 tasks spanning land cover, wildfire forecasting, geophysical and socio-economic structure, biodiversity and climate, LoCA ranks first on 8 and among the top two on 11, and is the only encoder to beat every other model on a majority of tasks.
The approach builds the target space first and then lets coordinates index it. Stage 1 jointly trains Sentinel-1 and Sentinel-2 encoders with the DHEL contrastive loss, so paired inputs meet at the loss optimum while the space keeps a high effective rank; Stage 2 freezes both encoders, attaches a projection adapter to each, and trains a location encoder and a metadata encoder over 14 tabular variables with the multi-view MV-DHEL loss. Across 343,928 land locations on a global grid, LoCA populates 62% of its 158 dimensions, about three times the 20.6% of Copernicus-FM, while a SatCLIP trained jointly from scratch reaches an effective rank of 8.
Removing Stage 1 breaks the multimodal target outright. Replacing the jointly trained S1 and S2 encoders with two independently MoCo-pretrained encoders leaves Stage 2 without a co-aligned multi-view target. reBEN micro-F1 then falls from 58.3 to 53.6, and pairwise accuracy on the five climate-ranking tasks lands between 49.6% and 50.6%, close to random ordering.
Perspective
The demonstrated setting is coordinate-indexed global location embeddings: pretraining uses 100,000 Sentinel-1 and Sentinel-2 image pairs plus 14 static metadata variables, and downstream evaluation trains linear probes, MLPs or a LightGBM ranker on frozen embeddings. The authors propose extending it with temporal information, meteorological variables and broader land-cover metadata.
The model has no temporal dimension, so seasonal and shorter-term dynamics are captured only indirectly; Stage 1 uses only Sentinel-1 and Sentinel-2; the authors note that coordinate-based benchmarks are sensitive to spatial autocorrelation and that blocked or leave-region-out evaluation remains future work.