LoCA aligns location encoders with multimodal satellite imagery
Original title: LoCA: Enabling Multi-Modality in Geospatial Location Encoders
On 17 tasks LoCA has the most wins and is the only model ahead on a majority of tasks.
The paper argues Lipschitz location encoders are limited to intrinsic dimension at most 2, so contrastive loss cannot reach its asymptotic optimum, and joint training collapses embedding covariance. LoCA first contrastively trains Sentinel-1 and Sentinel-2 encoders into a co-aligned high-rank target, then freezes them and trains location and tabular metadata encoders as multi-view adapters. Against five state-of-the-art encoders it wins 68.8% of per-task comparisons.
- Any Lipschitz location encoder has intrinsic dimension at most 2.
- Stage 1 jointly trains Sentinel-1 and Sentinel-2 encoders.
- Stage 2 freezes image encoders and trains projection adapters.
- Evaluation covers 17 tasks against five state-of-the-art encoders.
- LoCA wins 68.8% of all per-task comparisons.
- Products
- Sentinel-1Sentinel-2
- AI models
- LoCA
From the original article
1 min read at arxiv.orgLocation encoders map geographic coordinates to high-dimensional descriptors through contrastive alignment with paired Earth Observation imagery. We prove that this construction is structurally limited: any Lipschitz location encoder produces an output of intrinsic dimension at most 2, regardless of architecture or bandwidth, which forces the contrastive loss to strictly exceed its asymptotic optimum.
By Panagiotis Koromilas · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.