Research map / Vision-language and multimodal / Human and embodied modeling
SatlasPretrain: A Large-Scale Dataset for Remote Sensing Image Understanding
Favyen Bastani, Piper Wolters, Ritwik Gupta, Joe Ferdinando, Aniruddha Kembhavi
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 jointly trained SatlasNet on each label type instead of per-type training (ablation) | SatlasNet · SatlasPretrain test, all label types and image modes | average performance: 7.1% relative improvement | Improved |
| use multi-image time series input instead of single image (ablation) | SatlasNet · SatlasPretrain test, all label types and image modes | average performance: 5.6% relative improvement | Improved |
| joint training only without per-type fine-tuning (ablation) | SatlasNet (single-image) · SatlasPretrain test, low-res Sentinel-2, segmentation | F1: 64.8 -> 55.8 | Hurt |
| multi-image input without joint training (ablation) | SatlasNet, high-res NAIP · SatlasPretrain test, high-res NAIP, properties | accuracy: 86.6 -> 77.3 | Hurt |
| pre-train on SatlasPretrain instead of ImageNet | Swin-Base downstream fine-tuning · Seven downstream tasks, 50 examples (average) | average accuracy/F1: 0.38 -> 0.56 | Improved |
| pre-train on SatlasPretrain instead of ImageNet | Swin-Base downstream fine-tuning · Seven downstream tasks, full dataset (average) | average accuracy/F1: 0.75 -> 0.83 | Improved |
| pre-train on SatlasPretrain instead of DOTA | Swin-Base, Airbus Ships · Airbus Ships, 50 examples | F1: 0.62 -> 0.56 | Hurt |
| use MoCo self-supervised pre-training on SatlasPretrain images instead of supervised pre-training | Swin-Base downstream fine-tuning · Seven downstream tasks, 50 examples (average) | average accuracy/F1: 0.56 -> 0.21 | Hurt |
| use remote-sensing-specific self-supervised SeCo instead of ImageNet pre-training | Swin-Base downstream fine-tuning · Seven downstream tasks, 50 examples (average) | average accuracy/F1: 0.38 -> 0.38 | No clear effect |
Closest papers by meaning
- Rethinking Transformers Pre-training for Multi-Spectral Satellite Imagery (CVPR 2024 · similarity 0.89)
- InstructSAM: A Training-free Framework for Instruction-Oriented Remote Sensing Object Recognition (NeurIPS 2025 · similarity 0.89)
- VRSBench: A Versatile Vision-Language Benchmark Dataset for Remote Sensing Image Understanding (NeurIPS 2024 · similarity 0.88)
- SkyScript: A Large and Semantically Diverse Vision-Language Dataset for Remote Sensing (AAAI 2024 · similarity 0.88)
- GeoLLaVA-8K: Scaling Remote-Sensing Multimodal Large Language Models to 8K Resolution (NeurIPS 2025 · similarity 0.87)
- SkySense: A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery (CVPR 2024 · similarity 0.87)
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.
