Zero-shot burned area mapping with the Segment Anything Model (SAM): a label-free framework for post-fire
Fatih Fehmi Şimşek1, Melih Altay2
1TÜBİTAK Space Technologies Research Institute, Ankara, 06800, Turkey. fehmi.simsek@tubitak.gov.tr.
Environmental Monitoring and Assessment
|July 6, 2026
Summary
This study introduces a zero-shot approach using the Segment Anything Model (SAM) for accurate burned area mapping. The method achieves high performance comparable to supervised models without requiring labeled data, offering a scalable solution for forest fire monitoring.
Area of Science:
- Remote Sensing
- Geospatial Analysis
- Ecology
Background:
- Accurate burned area mapping is crucial for understanding fire impacts on ecosystems and carbon cycles.
- Traditional methods face limitations in accuracy and scalability due to spectral confusion and reliance on thresholds.
- Deep learning models show promise but require extensive labeled data and retraining for new regions.
Purpose of the Study:
- To propose and evaluate a novel zero-shot burned area mapping approach using the Segment Anything Model (SAM) with Sentinel-2 data.
- To assess the performance of SAM in achieving accurate and scalable burned area detection without task-specific training.
- To investigate the impact of different input data combinations and model configurations on mapping accuracy.
Main Methods:
- Utilized the Segment Anything Model (SAM), a foundation model, for zero-shot segmentation of burned areas.
- Generated composite input images from Sentinel-2 derived indices: Normalized Burn Ratio (NBR), NBR2, and Normalized Difference Vegetation Index (NDVI).
- Systematically investigated the effects of pre-processing, post-processing, and hyperparameter tuning (e.g., multi-scale configurations) on SAM's performance.
Main Results:
- Achieved high Intersection over Union (IoU) scores (0.89 for Bursa, 0.87 for Çanakkale) and F1 scores (0.94 for Bursa, 0.92 for Çanakkale).
- Demonstrated that SAM performance is comparable to supervised deep learning models, despite requiring no labeled training data.
- Found that multi-scale configurations and composite index inputs significantly improved geometric integrity, boundary accuracy, and discrimination between burned and unburned areas.
Conclusions:
- The proposed zero-shot SAM approach offers a fast, scalable, and cost-effective alternative for burned area detection, reducing reliance on manual labeling.
- SAM provides a flexible and generalizable framework for environmental monitoring, adaptable to various ecosystems and sensor conditions.
- This study pioneers the systematic application of SAM for burned forest area detection, highlighting its potential for data-scarce or time-critical scenarios.