Year: 2026
Type: preprint
Abstract: The August 2026 wildfires affected several provinces in northeastern Algeria, including Jijel, one of the country's most forested regions. This study maps and quantifies the resulting forest loss using Sentinel-2 imagery and deep learning. Median composites from before (12-22 August 2026) and after (1-10 September 2026) the fires are compared using NDVI and NBR. A U-Net model, trained mainly on Sentinel-2 data from El Tarf with ESA WorldCover labels and supplemented by additional private data, is then applied unchanged to both composites, and forest loss is defined as pixels that are forest before the fires but not after. Across the 2,384.4 km² analyzed, the model maps 943.7 km² of pre-fire forest, of which 468.4 km² (46,840 ha; 49.6%) is lost. The loss is highly concentrated: 17 contiguous patches larger than 100 ha account for 87.7% of it, and the five largest alone for 81.2%. Spectral analysis supports these results, as loss pixels show far stronger declines than stable forest (mean ΔNDVI = -0.413 vs. -0.013; ΔNBR = -0.631 vs. -0.016). Against ESA WorldCover 2021, the pre-fire map has high precision (0.917) but moderate recall (0.603), so the estimated loss is likely conservative.
DOI: 10.31223/x5gr70
Author: Raid Zella
Language: English
Topic: Fire effects on ecosystems
Subfield: Global and Planetary Change
Field: Environmental Science
Domain: Physical Sciences
SDGs: Life on land, Climate action
Open Access status: green
OpenAlex ID: W7235546748
