LoCA: Enabling Multi-Modality in Geospatial Location Encoders

LoCA. Enabling Multi-Modality in Geospatial Location Encoders. By Panagiotis Koromilas and nine others, from the University of Athens, the Cyprus Institute, and eight other institutions. Published on the twenty-third of October. Its code repository is a placeholder: a readme, and nothing to run yet.

Location encoders map geographic coordinates to high-dimensional descriptors, learned through contrastive alignment with paired Earth Observation imagery. The authors prove that this construction is structurally limited. Any Lipschitz location encoder produces an output of intrinsic dimension at most two, regardless of its architecture or bandwidth. That ceiling forces the contrastive loss to strictly exceed its asymptotic optimum. Empirically the picture matches: joint training also collapses the encoder's embedding covariance to low rank. However the encoder is built, its output cannot fill the high-dimensional descriptor space it is mapped into, and the loss pays for that shortfall.

Current location encoders sidestep this spectral collapse by training against a frozen image encoder. The authors reinterpret that frozen encoder as a covariance regularizer. It supplies a high-rank target for the location encoder to fold onto. The trouble arrives with more than one modality. A location encoder should then map a single location to several descriptors at once, and those descriptors may be incompatible with one another.

From this the authors identify two requirements a regularizer must satisfy for location encoding in multimodal regimes. The first is co-alignment across modalities, because adapters alone cannot align encoders that were pretrained independently. The second is high effective rank, because the location encoder inherits the spectrum of its target. Both follow from the diagnosis. The frozen image encoder worked as a single high-rank target. With several modalities, the target has to stay high-rank and also has to agree with itself across modalities. LoCA, short for Location Contrastive Alignment, is a two-stage pipeline designed to satisfy both.

Stage one jointly trains Sentinel-1 and Sentinel-2 encoders with contrastive learning. The result is a multimodal target that is co-aligned and high-rank, which are exactly the two requirements. Stage two freezes those image encoders. Against them it trains two encoders as projection adapters, using multi-view contrastive learning: one for location, and one for tabular metadata. The order of the stages follows from the second requirement: since the location encoder inherits its target's spectrum, the target has to be fixed and high-rank before the location encoder trains against it. The frozen target plays the role the single image encoder played before, now with more than one modality behind it.

LoCA is compared against five state-of-the-art encoders on seventeen tasks. They range from land cover and wildfire forecasting to geophysical, socio-economic, biodiversity and climate applications. LoCA achieves the most task wins. It is the only model to outperform every other on a majority of the tasks. And it wins sixty-eight point eight percent of all per-task comparisons.

Sixty-eight point eight percent of per-task comparisons won is where the work lands, but it starts from a ceiling: an intrinsic dimension of at most two for any Lipschitz location encoder. LoCA does not try to break that ceiling. Rather than asking more of the location encoder, it gives the encoder a better target to fold onto, built from Sentinel-1 and Sentinel-2 encoders that are aligned with each other and high in rank. Seventeen tasks and five rival encoders are the test of that choice, and the location encoder that results can draw on more than one view of the ground beneath its coordinates.