Research map / Vision-language and multimodal / Human and embodied modeling
RANGE: Retrieval Augmented Neural Fields for Multi-Resolution Geo-Embeddings
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes RANGE, a retrieval-augmented neural field that builds location embeddings by concatenating a geolocation embedding with a soft-weighted average of SatMAE features retrieved from a database of similar-looking satellite images, plus a spatially smoothed variant RANGE+.
Problem
Contrastive location-image methods (SatCLIP, GeoCLIP) produce geo-embeddings that keep only information shared with images, discarding unique visual detail useful for downstream geospatial tasks.
Approach
Approximate a location's unique high-resolution visual features by softmax-weighted retrieval over a compact database of low-resolution (projected SatCLIP) and high-resolution (SatMAE) image embeddings keyed by location embedding similarity, then concatenate with the original geo-embedding; RANGE+ adds a geodesic-retrieval term weighted by beta for spatial smoothness.
Key results
- RANGE reaches 83.3% biome accuracy vs 68.9% for SatCLIP, a 13.1% gain over the prior state of the art.
- RANGE+ reaches 0.851 R2 on elevation and 0.811 on population density, versus 0.666 and 0.684 for SatCLIP.
- On iNat-2018 with a pretrained classifier, Img+RANGE+ achieves 75.1% top-1 accuracy vs 66.1% for the image classifier alone and 75.1% for Img+SatCLIP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add semantic retrieval-augmented visual features to SatCLIP embeddings | SatCLIP · Biome | accuracy: 68.9 -> 83.3 | Helped |
| Add semantic retrieval-augmented visual features to SatCLIP embeddings | SatCLIP · Ecoregions | accuracy: 69.3 -> 75.7 | Helped |
| Add semantic retrieval-augmented visual features to SatCLIP embeddings | SatCLIP · Country | accuracy: 82.8 -> 93.7 | Helped |
| Add semantic retrieval-augmented visual features to SatCLIP embeddings | SatCLIP · Temperature | R2: 0.825 -> 0.895 | Helped |
| Add semantic retrieval-augmented visual features to SatCLIP embeddings | SatCLIP · Elevation | R2: 0.666 -> 0.844 | Helped |
| Add semantic retrieval-augmented visual features to SatCLIP embeddings | SatCLIP · Population | R2: 0.684 -> 0.799 | Helped |
| Add semantic retrieval-augmented visual features to SatCLIP embeddings | SatCLIP · Cali-Housing | R2: 0.400 -> 0.422 | Hurt |
| Add spatial retrieval to RANGE (ablation) | RANGE · Country | accuracy: 93.7 -> 94.7 | Helped |
| Add spatial retrieval to RANGE (ablation) | RANGE · Ecoregions | accuracy: 75.7 -> 75.3 | Hurt |
| Use top-k retrieval instead of soft selection (ablation) | RANGE · Biome | accuracy: 82.8 -> 83.3 | Hurt |
| Use top-1 retrieval instead of soft selection (ablation) | RANGE · Biome | accuracy: 75.6 -> 83.3 | Hurt |
| Tune temperature per task rather than use a fixed temperature (ablation) | RANGE · Biome | accuracy: 83.3 -> 83.5 | Helped |
Surprising findings
- RANGE underperforms on California housing price prediction (0.422 R2 vs 0.400 for SatCLIP is higher, but RANGE+ drops to 0.336), which the authors attribute to 1990 census labels versus 2020 Sentinel imagery.
- Retrieval from only 10% of the database retains similar performance, suggesting low variance in satellite semantics allows a compact database.
Limitations
- Underperforms on California housing price prediction, attributed to temporal mismatch between labels and imagery.
- Adds storage and retrieval overhead for the image database compared with purely parametric embeddings.
- The information-theoretic argument is informal and the retrieval approximation is not formally shown to recover unique information.
Experimental setup
Datasets: SatCLIP dataset (82k locations) iNaturalist 2018 Biome/Ecoregion (Dinerstein et al.) ERA5 air temperature elevation population density California housing
Benchmarks: Biome EcoRegions Country Temperature Elevation Population Cali-Housing iNaturalist 2018
Metrics: accuracy R2 top-k accuracy
Baselines: SatCLIP GeoCLIP SINR CSP SphereM SphereC Wrap Theory
Builds on: SatCLIP GeoCLIP SatMAE CLIP
Closest papers by meaning
- Scaling Image Geo-Localization to Continent Level (NeurIPS 2025 · similarity 0.87)
- SatCLIP: Global, General-Purpose Location Embeddings with Satellite Imagery (AAAI 2025 · similarity 0.87)
- GeoRanker: Distance-Aware Ranking for Worldwide Image Geolocalization (NeurIPS 2025 · similarity 0.86)
- G3: An Effective and Adaptive Framework for Worldwide Geolocalization Using Large Multi-Modality Models (NeurIPS 2024 · similarity 0.86)
- PIGEON: Predicting Image Geolocations (CVPR 2024 · similarity 0.85)
- Vision-Language Reasoning for Geolocalization: A Reinforcement Learning Approach (AAAI 2026 · similarity 0.84)
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.