Research map / Label-efficient and robust learning / Self-supervised and contrastive
GEO-Bench: Toward Foundation Models for Earth Monitoring
Alexandre Lacoste, Nils Lehmann, Pau Rodriguez, Evan David Sherwin, Hannah Kerner, Björn Lütjens, Jeremy Andrew Irvin, David Dao, Hamed Alemohammad, Alexandre Drouin, Mehmet Gunturkun, Gabriel Huang and 5 more
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 |
|---|---|---|---|
| Use ConvNeXt or SwinV2 instead of the other evaluated RGB baselines | GEO-Bench classification · GEO-Bench classification (six tasks, RGB) | normalized accuracy | Improved |
| Use ImageNet-pretrained ResNet18 instead of randomly initialized ResNet18 | ResNet18 · GEO-Bench classification | accuracy | Improved |
| Use remote-sensing-pretrained ResNet18 or ResNet50 instead of timm-pretrained weights | ResNet18 and ResNet50 · GEO-Bench classification (RGB) | normalized accuracy | No clear effect |
| Pretrain ResNet50 on Sentinel-2 multispectral data (ablation) | ResNet50 · Sentinel-2 classification tasks | normalized accuracy | Improved |
| Add randomly initialized first-layer weights for missing multispectral channels (ablation) | RGB-pretrained ResNet50 and ViT-S · Sentinel-2 classification tasks | normalized accuracy | No clear effect |
| Use multispectral channels with ViT-S (ablation) | ViT-S · Sentinel-2 classification tasks | normalized accuracy | Hurt |
| Reduce the classification training-set size (ablation) | GEO-Bench classification models · GEO-Bench classification | normalized accuracy | Mixed |
| Use ConvNeXt-B-timm with 2% of the training data instead of ResNet18-Rnd with the full training set (ablation) | GEO-Bench classification · GEO-Bench classification (aggregated) | normalized accuracy: 50x | Improved |
| Use U-Net rather than DeepLabV3 | ResNet segmentation baselines · GEO-Bench segmentation | normalized IoU | Improved |
| Use ResNet101 rather than ResNet50 as the segmentation backbone (ablation) | U-Net and DeepLabV3 segmentation baselines · GEO-Bench segmentation | normalized IoU | Hurt |
| Reduce the training-set size (ablation) | GEO-Bench classification models · GEO-Bench classification | dataset discriminativity | Improved |
Closest papers by meaning
- GeoCore-9B: Towards Geo-Aware Generative Foundation Models in Earth Observation (NeurIPS 2026 · similarity 0.87)
- Mars-Bench: A Benchmark for Evaluating Foundation Models for Mars Science Tasks (NeurIPS 2025 · similarity 0.87)
- REOBench: Benchmarking Robustness of Earth Observation Foundation Models (NeurIPS 2025 · similarity 0.87)
- Locate Anything on Earth: Advancing Open-Vocabulary Object Detection for Remote Sensing Community (AAAI 2025 · similarity 0.87)
- No One Knows the State-of-the-Art in Geospatial Foundation Models (NeurIPS 2026 · similarity 0.87)
- InstructSAM: A Training-free Framework for Instruction-Oriented Remote Sensing Object Recognition (NeurIPS 2025 · 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.
