Reproduce “Scale-based Approach for Active Wildfire Segmentation on Satellite Imagery” on OpenResearch

About this paper

Matheus F. Kovaleski, Cristiano Premebida, João Ruivo Paulo

Researchers at the University of Coimbra developed a scale-based evaluation protocol for active wildfire segmentation, revealing U-Net's robust performance across fire-size distribution shifts and emphasizing the critical role of the SWIR2 spectral band for effective detection, in contrast to DeepLabV3+'s generalization limitations.

Key insights

• U-Net demonstrated the most robust performance across all evaluation conditions, particularly maintaining high recall, attributed to its encoder-decoder structure and skip connections that preserve fine spatial details.
• The SWIR2 spectral band consistently provided more discriminative information for active fire segmentation than SWIR1 across all architectures, with Dual-SWIR and SWIR2-only configurations yielding superior results.
• The proposed scale-based evaluation protocol effectively exposed generalization limitations in models like DeepLabV3+, which showed significantly lower recall on large-fire test sets despite often correct predictions, highlighting the need for more challenging benchmarks.

Summary by alphaXiv.