Research map / Vision-language and multimodal / Cross-modal learning
Beyond What's Shared: Recovering Lost Unique Information from Intermediate Layers to Boost Multimodal Geo-Foundation Models
JangHyeon Lee, Philipe Ambrozio Dias, Yao-Yi Chiang, Dalton Lunga
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fuse intermediate and final layer embeddings (BWS) | SatCLIP · Biome | top-1 accuracy: 0.848 -> 0.911 | Improved |
| fuse intermediate and final layer embeddings (BWS) | CSP · Elevation | R2: 0.390 -> 0.659 | Improved |
| fuse intermediate and final layer embeddings (BWS) | GeoCLIP · Elevation | R2: 0.608 -> 0.809 | Improved |
| add BWS on top of RANGE | RANGE · Ecoregions | top-1 accuracy: 0.894 -> 0.920 | Improved |
| add BWS on top of RANGE | RANGE · Temperature | R2: 0.894 -> 0.897 | No clear effect |
| add BWS on top of RANGE | RANGE · Housing | R2: 0.449 -> 0.457 | No clear effect |
| use BWS layer fusion instead of dimension-matched concatenation with GeoCLIP (ablation) | SatCLIP · Seven-task average | average performance: 0.748 -> 0.788 | Improved |
| use BWS layer fusion instead of random projection of same dimension (ablation) | SatCLIP · Seven-task average | average performance: 0.750 -> 0.788 | Improved |
| apply BWS to a contrastively trained (multimodal) image encoder rather than unimodal (ablation) | MoCo-ResNet50 image encoder · Seven-task average | average performance: 0.247 -> 0.301 unimodal vs 0.523 -> 0.700 multimodal | Improved |
| use uniform layer weights (alpha=0) instead of emphasizing early or late layers (ablation) | BWS · Seven-task average | overall average: 0.788 vs 0.781 (alpha -2.0) and 0.776 (alpha 2.0) | Improved |
| use middle layers instead of late layers only (ablation) | SatCLIP · Seven-task average | overall average: 0.755 -> 0.780 | Improved |
| increase encoder depth when using BWS (ablation) | SatCLIP with BWS · Seven-task average | overall average: 0.790 -> 0.826 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- RANGE — RANGE: Retrieval Augmented Neural Fields for Multi-Resolution Geo-Embeddings (CVPR 2025) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- GeoCore-9B: Towards Geo-Aware Generative Foundation Models in Earth Observation (NeurIPS 2026 · similarity 0.88)
- SkyMoE: A Vision-Language Foundation Model for Enhancing Geospatial Interpretation with Mixture of Experts (AAAI 2026 · similarity 0.87)
- Textual Supervision Enhances Geospatial Representations in Vision-Language Models (ICML 2026 · similarity 0.86)
- SkySense: A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery (CVPR 2024 · similarity 0.86)
- Galileo: Learning Global & Local Features of Many Remote Sensing Modalities (ICML 2025 · similarity 0.86)
- GeoLink: Empowering Remote Sensing Foundation Model with OpenStreetMap Data (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
