Research map / Label-efficient and robust learning / Recognition and long-tail
Measuring the Intrinsic Dimension of Earth Representations
Arjun Rao, Marc Rußwurm, Konstantin Klemmer, Esther Rolf
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 |
|---|---|---|---|
| increase SatCLIP spherical-harmonic order from L=10 to L=40 (ablation) | SatCLIP location encoder · Land sampling (100k points) | global ID (FisherS): 5.00 -> 8.08 | Improved |
| increase SatCLIP embedding size D from 64 to 512 (ablation) | SatCLIP-L10 · Embedding-size study | global ID (MLE): 1.958 -> 1.956 | No clear effect |
| increase SatCLIP embedding size D from 64 to 512 (ablation) | SatCLIP-L10 MLP head · Air temperature | R2: 95.9 -> 95.9 | No clear effect |
| increase RCF embedding size D from 64 to 512 (ablation) | RCF image encoder · S2-100K | global ID (MLE): 5.56 -> 6.32 | No clear effect |
| increase GeoCLIP max RFF frequency and hierarchy levels M (ablation) | GeoCLIP location encoder · YFCC fine-tuning | global ID (FisherS) | Improved |
| increase number of multi-scale components S (ablation) | Space2Vec / Sphere2Vec · MOSAIKS nightlights | global ID | Improved |
| pre-train SatCLIP on more modalities (S2 to S1+S2 to all MMEarth) (ablation) | SatCLIP · MMEarth air temp, elevation, population | global ID and R2 | Improved |
| higher ID of frozen pretrained embeddings | geographic INRs · five TorchSpatial tasks | test R2 / top-1 | Improved |
| lower TwoNN ID of head activations | MLP head on frozen INR · five TorchSpatial tasks | test R2 / top-1 | Improved |
| end-to-end supervised training instead of pretraining | Sphere2Vec and Space2Vec · SustainBench image-location regression | R2 | Improved |
| switch estimator from distance-based MLE to angle-based FisherS | SatCLIP-L40 · Land sampling (100k points) | global ID: 2.03 -> 8.08 | Mixed |
| increase neighbors k from 5 to 200 with ESS (ablation) | GeoCLIP · Fibonacci sampling | global ID (ESS): 12.031 -> 98.897 | Hurt |
Closest papers by meaning
- No Location Left Behind: Measuring and Improving the Fairness of Implicit Representations for Earth Data (ICLR 2025 · similarity 0.91)
- GeoCore-9B: Towards Geo-Aware Generative Foundation Models in Earth Observation (NeurIPS 2026 · similarity 0.86)
- Grids Often Outperform Implicit Neural Representation at Compressing Dense Signals (NeurIPS 2025 · similarity 0.85)
- Towards Faithful Reasoning in Remote Sensing: A Perceptually-Grounded GeoSpatial Chain-of-Thought for Vision-Language Models (ICLR 2026 · similarity 0.85)
- Geographic Location Encoding with Spherical Harmonics and Sinusoidal Representation Networks (ICLR 2024 · similarity 0.85)
- Poly2Vec: Polymorphic Fourier-Based Encoding of Geospatial Objects for GeoAI Applications (ICML 2025 · 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.
