Wildfire Smoke Density Classification via Evidential Deep Learning
There's a new probabilistic model that sorts wildfire smoke density captured in satellite images into three levels: Light, Moderate, and Heavy. It distinguishes between different types of uncertainty all in one go. This model uses a pre-trained EfficientNet-B3 backbone paired with a CBAM module and has an evidential deep learning head that estimates Dirichlet concentration parameters without needing Monte Carlo sampling. When it was tested on 16,298 satellite patches from the Wildfire Detection dataset, it scored an impressive 93.8% in weighted test accuracy (91.1% unweighted). This approach provides reliable confidence measures for evaluating smoke intensity, which is essential for emergency responses, air quality assessments, and managing health risks.
Key facts
- Framework categorizes smoke density into Light, Moderate, Heavy classes
- Provides epistemic and aleatoric uncertainty in single forward pass
- Uses EfficientNet-B3 backbone with CBAM module
- Evidential deep learning head predicts Dirichlet concentration parameters
- No Monte Carlo sampling required
- Evaluated on 16,298 satellite patches from Wildfire Detection dataset
- Achieves 93.8% weighted test accuracy, 91.1% unweighted
- Aims to improve emergency response, air quality modeling, health risk management
Entities
—