Research map / Label-efficient and robust learning / Recognition and long-tail
TorchSpatial: A Location Encoding Framework and Benchmark for Spatial Representation Learning
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.
Introduces TorchSpatial, a unified framework implementing 15 location encoders, the LocBench benchmark (7 geo-aware classification and 10 regression datasets), and the Geo-Bias Score metric based on spatial self-information.
Problem
Spatial representation learning lacks a standardized framework and benchmark for location encoders, and no universal measure exists for the geographic bias introduced by geo-aware models.
Approach
A unified PyTorch framework consolidating 15 2D and 3D location encoders within a shared Enc(x)=NN(PE(x)) design, evaluated on a curated benchmark, plus a geographic bias metric (Geo-Bias Score) built on spatial self-information that is numerically comparable across models.
Key results
- Sphere2Vec-sphereC+ achieves the best Top-1 accuracy on 4 of 7 classification datasets (e.g., 75.38 average over species datasets), while location encoders generally boost accuracy over the No Prior image model (67.94 average).
- Adding location encoders raises geo-bias scores on most classification datasets (e.g., wrap on NABirds† unmarked 57.37 vs 39.71 for No Prior), whereas tile has little effect.
- On regression, Space2Vec-theory, Sphere2Vec variants, and wrap improve R^2 on population density, forest cover and elevation; GPT-4V zero-shot reached only 17% on fMoW and 18% on iNat2018.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add a location encoder to the image-only classifier | Geo-aware image classification models · LocBench image classification datasets | Top-1 accuracy | Helped |
| Use location information for regression | Geo-aware image regression models · Population density, forest cover, elevation, and most other tasks | R² | Helped |
| Use location information | Geo-aware image regression models · Nightlight luminosity | R² | No clear effect |
| Add a location encoder to the image classifier | Geo-aware image classification models · Datasets with geographically biased spatial distributions | Geo-Bias Score | Hurt |
| Use a location encoder | Geo-aware image regression models · Globally uniform MOSAIKS datasets | Geo-Bias Score | No clear effect |
| Replace wrap's residual-block neural module with the feed-forward neural module (ablation) | wrap location encoder · iNat2017 Val | Top-1 accuracy: 68.16 -> 69.52 | No clear effect |
Surprising findings
- Adding a location encoder generally increases geographic bias scores even while improving overall accuracy, except for the tile encoder.
- Location information did not significantly help nightlight luminosity prediction.
Limitations
- Framework currently supports only point (location) encoders, not polylines, polygons, or networks.
- No error bars reported across random seeds due to cost.
- GPT-4V evaluated only on small subsets of data due to budget constraints, so its geo-bias scores are not directly comparable.
Experimental setup
Datasets: BirdSnap BirdSnap† NABirds† iNat2017 iNat2018 YFCC100M-GEO100 fMoW MOSAIKS SustainBench
Benchmarks: LocBench BirdSnap NABirds iNat2017 iNat2018 YFCC fMoW MOSAIKS SustainBench
Metrics: Top-1 accuracy Top-3 accuracy MRR R^2 MAE RMSE unmarked SSI geo-bias score marked SSI geo-bias score
Baselines: No Prior tile wrap wrap + ffn rbf rff Space2Vec-grid Space2Vec-theory xyz NeRF Sphere2Vec variants Siren (SH)
Builds on: Space2Vec Sphere2Vec TorchGeo NeRF Random Fourier Features
Closest papers by meaning
- Poly2Vec: Polymorphic Fourier-Based Encoding of Geospatial Objects for GeoAI Applications (ICML 2025 · similarity 0.88)
- Vision-Language Reasoning for Geolocalization: A Reinforcement Learning Approach (AAAI 2026 · similarity 0.87)
- Textual Supervision Enhances Geospatial Representations in Vision-Language Models (ICML 2026 · similarity 0.87)
- SatCLIP: Global, General-Purpose Location Embeddings with Satellite Imagery (AAAI 2025 · similarity 0.87)
- GRE Suite: Geo-localization Inference via Fine-Tuned Vision-Language Models and Enhanced Reasoning Chains (NeurIPS 2025 · similarity 0.86)
- SpatialRGPT: Grounded Spatial Reasoning in Vision-Language Models (NeurIPS 2024 · 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.