Research map / Large language models / AI and society
Position: Mission Critical – Satellite Data is a Distinct Modality in Machine 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.
Position paper arguing satellite data is a distinct ML modality, outlining its unique characteristics, deployment and evaluation challenges, ethical concerns, and discussion topics for the SatML community.
Problem
Machine learning for satellite data (SatML) mainly applies methods designed for natural images or text via "lift and shift," which is ill-suited to satellite data's scale, spectral, spatiotemporal, label, and deployment characteristics.
Approach
Argues that satellite data should be recognized as a distinct data modality in ML, with dedicated research agendas for methods, evaluation, ethics, and community practices, rather than treating SatML as a mere application area.
Key results
- Reports that Copernicus Data Space Ecosystem held 66 petabytes of public Sentinel data as of January 2024, while ImageNet is about 150 GB.
- Notes that a 2024 review found 19 SatML-promoting workshops at CS/ML conferences since 2018 (Table in appendix).
- Cites a rotation-equivariant small network outperforming a larger CNN on SatML land cover mapping, and an architecture nearly 100x smaller than standard segmentation models achieving state-of-the-art land cover segmentation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Pre-train supervised models on satellite datasets instead of ImageNet | SatML models · SatML tasks | performance: 18% | Helped |
| Use ImageNet pre-training instead of random initialization | SatML models · Some SatML tasks | performance | Hurt |
| Use rotation-equivariant models instead of standard convolutional neural networks | Land cover mapping models · SatML land cover mapping tasks | performance | Helped |
| Use structured masking instead of conventional random masking | Masked autoencoder models · Downstream SatML tasks | performance | Helped |
| Fuse complementary sensor data instead of training on a single sensor | SatML models · SatML tasks | performance | Helped |
| Evaluate a post-disaster building-damage model at a new disaster site | Building-damage detection models · New disaster site | accuracy: 30% | Hurt |
Surprising findings
- ImageNet pre-trained models are much worse than random initialization for some SatML tasks.
- Random convolutional features and random pixel sets are strong baselines for many SatML tasks, but generally not for natural images.
Limitations
- No new experiments or empirical results; argument is based on cited literature and perspective.
- Scope is focused on SatML and largely excludes aerial and other geospatial modalities beyond brief mention.
Closest papers by meaning
- Satformer: Accurate and Robust Traffic Data Estimation for Satellite Networks (NeurIPS 2024 · similarity 0.81)
- DiffusionSat: A Generative Foundation Model for Satellite Imagery (ICLR 2024 · similarity 0.80)
- Data Augmentation Approaches for Satellite Imagery (AAAI 2025 · similarity 0.79)
- The State of Data Curation at NeurIPS: An Assessment of Dataset Development Practices in the Datasets and Benchmarks Track (NeurIPS 2024 · similarity 0.79)
- GeoLLM: Extracting Geospatial Knowledge from Large Language Models (ICLR 2024 · similarity 0.79)
- Prediction-Augmented Trees for Reliable Statistical Inference (NeurIPS 2026 · similarity 0.79)
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.