Research map / Image generation / Autoregressive and latent generation
EarthLoc: Astronaut Photography Localization by Indexing Earth from Space
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.
Reframes astronaut photography localization as image retrieval against a worldwide Sentinel-2 satellite database, training a MixVPR-style ResNet50 model (EarthLoc) with clustered batches, year-wise augmentation and a neutral-aware multi-similarity loss.
Problem
Automatically localizing astronaut photographs of Earth, whose only coarse location is the ISS nadir point and whose true extent can be thousands of kilometers away, is slow with existing pairwise matching methods.
Approach
Treat astronaut photo localization as retrieval against a precomputed global satellite database, and train with clustered batches, year-wise augmentation (same augmentation per year across regions), and a neutral-aware multi-similarity loss that treats partially overlapping regions as neither positive nor negative.
Key results
- EarthLoc reaches 57.1% Recall@1 averaged over the six evaluation sets, outperforming all compared methods.
- AnyLoc (DINOv2 + NetVLAD) is the strongest prior baseline at roughly 50% Recall@1 but has about 50x slower feature extraction and 10x larger features than EarthLoc.
- Ablation: adding clustered batches, year-wise augmentation and neutral-aware loss to the multi-similarity baseline raises average Recall@1 from 26.2 to 57.1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use regional quadruplets instead of ordinary image batches (ablation) | EarthLoc training · Six evaluation sets (average) | R@1: 26.2 -> 45.4 | Helped |
| Add clustered batches (ablation) | EarthLoc with regional quadruplets · Six evaluation sets (average) | R@1: 45.4 -> 49.8 | Helped |
| Add year-wise data augmentation (ablation) | EarthLoc with clustered batches · Six evaluation sets (average) | R@1: 49.8 -> 54.1 | Helped |
| Add neutral-aware multi-similarity loss (ablation) | EarthLoc · Six evaluation sets (average) | R@1: 54.1 -> 57.1 | Helped |
| Use EarthLoc instead of AnyLoc | Astronaut photography retrieval · Six evaluation sets (average) | R@1: 57.1 | Helped |
| Use clustered batches rather than random-region batches (ablation) | EarthLoc training · Astronaut photography localization | Recall: roughly 15% | Helped |
Surprising findings
- Self-supervised satellite-pretrained TorchGeo models performed no better than urban visual place recognition methods on this task.
- Previous claims by NASA scientists that retrieval-based APL is infeasible because the reference database cannot be precomputed were contradicted by the proposed pipeline.
Limitations
- Focuses only on daytime, nadir-like queries; panoramic and nighttime imagery not investigated.
- Database zoom levels limit localizable query areas to roughly 5000 to 900,000 sq km.
Experimental setup
Datasets: Sentinel-2 cloudless satellite imagery (s2maps.eu) Gateway to Astronaut Photography of Earth (ISS photos) FMAP ground truth
Benchmarks: Texas (validation) Alps California Gobi Desert Amazon Toshka Lakes
Metrics: Recall@1 Recall@10 Recall@100
Baselines: NetVLAD SFRS Conv-AP CosPlace MixVPR EigenPlaces AnyLoc TorchGeo SeCo TorchGeo GASSL TorchGeo MOCO OGCL MBEG
Builds on: MixVPR Multi-Similarity loss AnyLoc Find My Astronaut Photo
Closest papers by meaning
- GeoRanker: Distance-Aware Ranking for Worldwide Image Geolocalization (NeurIPS 2025 · similarity 0.85)
- PIGEON: Predicting Image Geolocations (CVPR 2024 · similarity 0.84)
- Around the World in 80 Timesteps: A Generative Approach to Global Visual Geolocation (CVPR 2025 · similarity 0.83)
- LocDiff: Identifying Locations on Earth by Diffusing in the Hilbert Space (NeurIPS 2025 · similarity 0.82)
- Scaling Image Geo-Localization to Continent Level (NeurIPS 2025 · similarity 0.82)
- GeoRouter: Dynamic Paradigm Routing for Worldwide Image Geolocalization (NeurIPS 2026 · similarity 0.82)
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.