Prithvi-EO-2.0-300M-BurnScars
| Property | Value |
|---|---|
| Author | IBM-NASA Geospatial |
| Performance | 87.52% IoU for Burned Areas |
| Architecture | UNetDecoder-based |
| Paper | arXiv:2412.02732 |
What is Prithvi-EO-2.0-300M-BurnScars?
Prithvi-EO-2.0-300M-BurnScars is a specialized Earth observation model fine-tuned for detecting and segmenting burned areas in satellite imagery. Built upon the Prithvi-EO-2.0-300M foundation model, it processes multi-spectral HLS (Harmonized Landsat Sentinel) imagery to identify burn scars with high accuracy.
Implementation Details
The model utilizes six spectral bands (Blue, Green, Red, Narrow NIR, SWIR, SWIR 2) and has been trained on approximately 800 labeled 512x512 image chips from the continental US. It employs a UNetDecoder architecture and achieves state-of-the-art performance with 87.52% IoU for burned area detection.
- Training conducted using TerraTorch framework
- Supports single-timestamp segmentation
- Validated through 5 repeated training runs
- Non-overlapping train/validation/test splits
Core Capabilities
- Precise burn scar segmentation in satellite imagery
- Three-class classification: burned areas, non-burned areas, and no data/clouds
- Flexible deployment with single-timestamp analysis
- High accuracy with 93.00% mean IoU on test data
Frequently Asked Questions
Q: What makes this model unique?
This model specializes in burn scar detection using satellite imagery, featuring high accuracy and practical deployment capabilities for real-world applications. It builds upon the proven Prithvi-EO-2.0 architecture while implementing specific optimizations for burn detection.
Q: What are the recommended use cases?
The model is ideal for monitoring wildfire impacts, forest management, ecological assessment, and disaster response applications where accurate mapping of burned areas is crucial. It works with HLS satellite imagery and can process single-timestamp data effectively.
