Research map / Vision-language and multimodal / CLIP and zero-shot
WildSAT: Learning Satellite Image Representations from Wildlife Observations
Rangel Daroya, Elijah Cole, Oisin Mac Aodha, Grant Van Horn, Subhransu Maji
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 |
|---|---|---|---|
| fine-tune with WildSAT on species observations | base models (avg across 15 architectures, no random) · UCM linear probing | accuracy: 81.8 -> 87.9 | Improved |
| fine-tune with WildSAT on species observations | base models (avg incl. random) · BEN20k linear probing | micro F1: 38.5 -> 53.1 | Improved |
| fine-tune with WildSAT | randomly initialized ViT-B/16 · BEN20k linear probing | micro F1: 0.0 -> 51.2 | Improved |
| fine-tune with WildSAT | CLIP ViT-B/16 · 7 satellite datasets (avg) | accuracy: 71.6% -> 76.6% | Improved |
| fine-tune CLIP with cross-modal methods (TaxaBind, GRAFT) instead of leaving unchanged | CLIP ViT-B/16 · 7 satellite datasets (avg) | accuracy: 71.6% -> 59.8% | Hurt |
| add WildSAT fine-tuning | SatlasNet, SatMAE, TaxaBind · SatBird USA Summer | top-k accuracy: 47.66 -> 49.53 | Improved |
| add WildSAT fine-tuning | SatlasNet, SatMAE, TaxaBind · SatBird Kenya | top-k accuracy: 23.80 -> 24.12 | Improved |
| add WildSAT fine-tuning | ResNet50 encoders · Cashew1k segmentation | IoU: 59.9% -> 72.2% | Improved |
| add WildSAT fine-tuning | ResNet50 encoders · SAcrop3k segmentation | IoU: 21.4% -> 22.7% | Improved |
| add location (loc) loss (ablation) | WildSAT (random ResNet50 + ImageNet ViT-B/16) · UCM/AID/RESISC45/So2Sat20k avg | accuracy: 48.6% -> 57.5% | Improved |
| add environmental covariates to location encoder (ablation) | WildSAT with loc · UCM/AID/RESISC45/So2Sat20k avg | accuracy: 57.5% -> 62.0% | Improved |
| add text loss (ablation) | WildSAT with loc+env · UCM/AID/RESISC45/So2Sat20k avg | accuracy: 62.0% -> 66.1% | Improved |
Closest papers by meaning
- SatCLIP: Global, General-Purpose Location Embeddings with Satellite Imagery (AAAI 2025 · similarity 0.87)
- DiffusionSat: A Generative Foundation Model for Satellite Imagery (ICLR 2024 · similarity 0.86)
- AnySat: One Earth Observation Model for Many Resolutions, Scales, and Modalities (CVPR 2025 · similarity 0.86)
- SkySense-O: Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling (CVPR 2025 · similarity 0.84)
- Landsat30-AU: A Vision-Language Dataset for Australian Landsat Imagery (AAAI 2026 · similarity 0.84)
- SkyScript: A Large and Semantically Diverse Vision-Language Dataset for Remote Sensing (AAAI 2024 · 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.
