Towards Consistent Geospatial Representations: Aligning One Encoder to Many Foundation Models
We introduce Pangaea-Latents, an extension of the Pangaea benchmark providing frozen embeddings of $5$ GFMs over its $11$ datasets, along with reproduced baselines, and find GFM embeddings to be neither redundant nor disjoint.
ProofPaper ↗
Key points
- Geospatial foundation models (GFMs) differ widely in pretraining objective, modality and resolution, and none leads across benchmarks.
- We ask whether their representations are complementary, and whether a single encoder can inherit that complementarity.
- We then propose GeoMixture, a preliminary framework to align a ViT to several GFMs per dataset through a multi-representation alignment (mREPA) objective.
- Our encoder ties the best GFM in average rank across nine Pangaea datasets, trading peak performance for cross-task consistency.
Sources (1)
- [1]Towards Consistent Geospatial Representations: Aligning One Encoder to Many Foundation ModelsarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 20, 02:52 PM
We introduce Pangaea-Latents, an extension of the Pangaea benchmark providing frozen embeddings of $5$ GFMs over its $11$ datasets, along with reproduced baselines, and find GFM embeddings to be neither redundant nor disjoint.
Geospatial foundation models (GFMs) differ widely in pretraining objective, modality and resolution, and none leads across benchmarks.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 20, 2026Convergent Performance, Divergent Mechanisms in LoRA-Adapted Vision Foundation Models for Brain MRI
- Oct 19, 2026LSA-PAST: Layer-Specific Adaptation and Prototype Transport for Compositional Incremental Learning
- Oct 19, 2026Arrow detection with ConvNext pretrained on DINOv3 [P]
- Oct 19, 2026Corrected Block Sparse Spectral Optimization: Exact Models, Finite Identification, and Local Dynamics
- Oct 18, 2026The Explainability Illusion: Must ViTs Exhibit Patch-Decoding Interpretability?
- Oct 18, 2026A discrete-continuous-time stochastic framework for attention-deficit/hyperactivity disorder-related symptom modeling and pharmacological intervention via methylphenidate