🌌 Surya – Active Region Segmentation
📖 Model Overview
This repository hosts fine-tuned weights of Surya – a heliophysics foundation model – for the task of solar Active Region (AR) segmentation.
Solar Active Regions are magnetically complex structures associated with flares and coronal mass ejections (CMEs). Within ARs, the Polarity Inversion Line (PIL) serves as a critical precursor of eruptions. Accurate segmentation of ARs containing PILs is essential for space weather forecasting and understanding solar magnetic complexity.
📊 Results
We benchmarked Surya against a standard UNet baseline on the ARPIL dataset.
| Model | Params | IoU | Dice Coeff |
|---|---|---|---|
| UNet | 9.2 M | 0.688 | 0.801 |
| Surya (LoRA) | 4.1 M | 0.768 | 0.853 |
Surya achieves higher segmentation quality with fewer parameters, highlighting the benefits of foundation model pretraining and parameter-efficient adaptation.
🖼 Example
- Top Row: Input SDO/HMI data (Date: 2014-02-01, Time: 08:12)
- Middle Row: Surya segmentation output
- Bottom Row: Ground Truth
⚡ Usage
Follow the instructions at Surya/downstream_examples/ar_segmentation
🤝 Acknowledgements
- NASA IMPACT and IBM for developing the Surya foundation model