OpenEnergy-Engine: OlmoEarth Energy Infrastructure Segmentation
Fine-tuned OlmoEarth (ViT-BASE) for segmenting energy infrastructure from 12-band Sentinel-2 imagery, using EIA Form 860 power plant data as ground truth labels.
Model Details
- Base model: OlmoEarth V1 BASE (89M encoder params)
- Decoder: UNetDecoder
- Task: Semantic segmentation (11 classes)
- Input: 12-band Sentinel-2 L2A, 10m resolution
- Training: 30 epochs, encoder frozen first 10, RTX 5090
- val_loss: 0.231
Classes
| ID | Class | EIA Source Codes |
|---|---|---|
| 0 | background | - |
| 1 | solar | SUN |
| 2 | wind | WND |
| 3 | gas | NG, BFG, SGC, SGP, OG |
| 4 | coal | COL, PC, WC |
| 5 | nuclear | NUC |
| 6 | hydro | WAT |
| 7 | oil | OIL, DFO, RFO, TDF |
| 8 | biomass | WDS, LFG, OBG, OBL, OBS, AB, MSW |
| 9 | geothermal | GEO |
| 10 | storage | MWH |
Training Data
- Labels: EIA Form 860 (2024) - 13,000+ US power plants
- Imagery: Sentinel-2 L2A via Microsoft Planetary Computer (May-Aug 2023)
- Windows: 8,113 train / 1,158 val (128x128 @ 10m)
Model tree for 2imi9/openenergyengine-olmoearth-base
Base model
allenai/OlmoEarth-v1-Base