Research map / Label-efficient and robust learning / Self-supervised and contrastive
Learning Representations of Satellite Images From Metadata Supervision
Jules Bourcier, Gohar Dashyan, Karteek Alahari, Jocelyn Chanussot
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 |
|---|---|---|---|
| add SimCLR image-image loss to metadata-image contrastive pretraining (SatMIPS vs SatMIP) | SatMIP · fMoW kNN | macro F1: 55.2 -> 62.3 | Improved |
| add metadata supervision to SimCLR (SatMIPS vs SimCLR) | SimCLR · fMoW kNN | macro F1: 61.1 -> 62.3 | Improved |
| metadata-image contrast (SatMIP) instead of SimCLR image-only on fMoW | SatMIP · fMoW kNN | macro F1: 61.1 -> 55.2 | Hurt |
| concatenate metadata features with image features at downstream linear probing (ablation) | SatMIP · fMoW linear probe | macro F1: 59.3 -> 63.1 | Improved |
| concatenate metadata features with image features on EuroSAT (ablation) | SatMIPS · EuroSAT linear probe | accuracy: 95.8 -> 96.4 | Improved |
| add hierarchical pretraining from SLIP-YFCC | SatMIP · fMoW kNN | macro F1: 55.2 -> 61.2 | Improved |
| add hierarchical pretraining from SLIP-YFCC | SimCLR · FGSC23 kNN | accuracy: 57.3 -> 55.1 | Hurt |
| use location instead of GSD as the only metadata field (ablation) | SatMIP · fMoW linear probe | macro F1: 37.4 -> 47.1 | Improved |
| add timestamp and 12 other fields to GSD and location (ablation) | SatMIP · fMoW linear probe | macro F1: 51.3 -> 50.7 | No clear effect |
| use strong SimCLR augmentation instead of light random resized crop (ablation) | SatMIP · fMoW kNN | macro F1: 44.6 -> 45.6 | Improved |
| couple one image view between SimCLR and SatMIP objectives (ablation) | SatMIPS · fMoW kNN | training time (rel.): 1.53 -> 1.05 | Improved |
| train with SatMIPS instead of SimCLR for faster convergence | SimCLR · all downstream datasets | kNN performance | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Scale-MAE — Scale-MAE: A Scale-Aware Masked Autoencoder for Multiscale Geospatial Representation Learning (ICCV 2023) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Rethinking Transformers Pre-training for Multi-Spectral Satellite Imagery (CVPR 2024 · similarity 0.85)
- InstructSAM: A Training-free Framework for Instruction-Oriented Remote Sensing Object Recognition (NeurIPS 2025 · similarity 0.85)
- Galileo: Learning Global & Local Features of Many Remote Sensing Modalities (ICML 2025 · similarity 0.84)
- SatCLIP: Global, General-Purpose Location Embeddings with Satellite Imagery (AAAI 2025 · similarity 0.84)
- DiffusionSat: A Generative Foundation Model for Satellite Imagery (ICLR 2024 · similarity 0.84)
- SkySense-O: Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling (CVPR 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.
