Answers are grounded in this paper and cite their sources.
Bi-temporal change detection using the Differenced Normalized Burn Ratio (ΔNBR) remains the operational standard for satellite-based burned area mapping, yet cloud-free pre-fire imagery is frequently unavailable. While single-date methods bypass this temporal constraint, conventional index thresholding suffers from high false-positive rates and substantial area estimation errors, often exceeding a MdAPE of 63.7%. To address this limitation, we assess single-image approaches against EFFIS reference perimeters across 30 Euro-Mediterranean wildfire sites from 2021 to 2026, and introduce EMBER, an unsupervised framework designed to delineate burned perimeters from a single Sentinel-2 scene. The methodology expands surface reflectance bands into a 20-dimensional feature space using 3×3 local spatial means, applying principal component analysis and a Gaussian Mixture Model governed by a near-infrared maximum-gap rule and spatial connectivity constraints. Equivalence testing via Two One-Sided Tests (TOST) confirmed formal statistical equivalence between EMBER and bi-temporal ΔNBR within a δ=±0.10 tolerance bound (pTOST<0.01) across the pixel-level overlap metrics (F1-score: 83.9% vs. 85.6%; IoU: 72.3% vs. 74.9%; Precision: 94.7% vs. 94.6%; Recall: 78.5% vs. 80.5%). For boundary displacement (HD95=792.5m vs. 758.5m), no statistically significant difference was detected relative to ΔNBR (padj=0.393). Processing each scene in an average of 10.2s, EMBER does not aim to replace bi-temporal analysis, but demonstrates that local spatial-spectral clustering can serve as a reliable alternative when pre-fire baseline data is missing, bypassing data retrieval delays without the computational overhead of deep learning architectures.
No takes yet. Share an insight, caveat, or question.
Mercatini et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: