Spatially explicit machine learning reveals province-specific drivers of deforestation in the Hyrcanian forests

Rapid deforestation is occurring worldwide, making it crucial to understand the spatial factors driving it to develop targeted, region-specific conservation strategies. In this study, we used three machine learning models, i.e., boosted regression trees (BRT), random forest (RF), and classification and regression trees (CART), to map deforestation susceptibility across the Hyrcanian forests in northern Iran, focusing on Gilan, Mazandaran, and Golestan provinces. We identified historical deforestation sites through field surveys and archival records, and derived fifteen biophysical and human-related variables from high-resolution digital elevation models, satellite imagery, climate data, and local archives. Among the models, BRT with an average AUC of 0.997 provided better predictive performance compared to RF (AUC = 0.993) and CART (AUC = 0.924). The BRT results revealed distinct regional drivers of deforestation. In Golestan, susceptibility was strongly shaped by vegetation dynamics (NDVI 2023, 20.5 %; NDVI 2018, 12.2 %) and wind effect (18.7 %), highlighting the vulnerability of fragmented foothill forests to canopy loss and physical disturbances. In Mazandaran, elevation (31.8 %) was the dominant predictor, reflecting how accessibility gradients govern land conversion, with secondary influences from agricultural proximity (17 %) and NDVI change. In Gilan, deforestation was primarily driven by proximity to agricultural lands (30.7 %) and elevation (26 %), alongside road access (10.6 %) and wind effect (7.9 %), underscoring the combined effects of land-use pressure and natural disturbance in this densely populated coastal province. These results demonstrate that deforestation in the Hyrcanian forests is not governed by a universal set of drivers but by region-specific interactions between vegetation dynamics, topography, accessibility, and human pressures. Conservation strategies should therefore be tailored: in Golestan, by monitoring canopy decline and protecting wind-exposed forests; in Mazandaran, by implementing elevation-based land-use zoning and regulating expansion in lowland zones; and in Gilan, by integrating agricultural buffer zones, road management, and ecological restoration. • Deforestation susceptibility was modeled in three provinces of northern Iran. • Vegetation dynamics and wind were dominant in Golestan province. • Elevation was the top predictor in Mazandaran province. • In Gilan, human pressure and accessibility were the main drivers. • Province-specific conservation strategies are needed for deforestation control.