From global burned-area products to higher-resolution regional forest fire monitoring: a Landsat-based machine-learning reconstruction for northern Moroccan forests.

Wildfires are a major disturbance in Mediterranean forests, yet long-term burned-area accounting remains uncertain in fragmented landscapes where coarse-resolution global products may smooth burn perimeters and miss small scars. This study presents a forest-constrained regional reconstruction of annual forest burned area at 30 m resolution for northern Morocco over 1984-2025, excluding 2012, using the Landsat archive, spectral change metrics, and supervised machine learning. Annual mapping was constrained by Landsat-derived forest-domain masks anchored to an official 2018 forest inventory. Burned/unburned discrimination used paired pre- and post-fire composites and three predictors: ΔNBR, ΔMIRBI, and ΔBAI. Five classifiers were evaluated under a strict chronological design, with 2001-2018 for training, 2019-2021 for validation, and 2022-2024 for independent testing. A calibrated radial-basis support vector machine performed best on the independent sample-based test period, with PR-AUC = 0.988, ROC-AUC = 0.989, F1 = 0.948, Brier score = 0.035, and a fixed operating threshold of τ = 0.50. The reconstructed record showed strong interannual variability, with a median of 609 ha and a range of 135-19,451 ha, and closely matched official annual burned-area statistics (r = 0.998, NRMSE = 16.97%, bias = + 4.17%). Remaining uncertainties relate mainly to very small fires, limited clear-sky Landsat observations, and point-based validation records. The resulting dataset provides a spatially explicit baseline for annual forest burned-area accounting and regional fire monitoring in northern Moroccan Mediterranean forests.