Accurate machine learning reconstruction and forecasting of global wetland extent

Accurately projecting global wetland dynamics is currently constrained by the inability of traditional process-based models (e.g. TOPMODEL) to resolve anthropogenic disturbances, leading to systematic biases in human-modified landscapes. Here, we present a machine learning framework integrating multi-model CMIP6 climate projections with fifteen anthropogenic indicators, benchmarked against terrain based TOPMODEL simulations. Using a Random Forest algorithm, our framework reproduced the historical wetland distribution (2000–2014) with a Pearson correlation (r) of 0.89 (±0.01), corresponding to a 43.55% improvement over TOPMODEL (r = 0.62) and a 98.5% increase in correlation within Human Activity Zones (HAZs). Feature importance analysis reveals that biophysical characteristics and anthropogenic indicators accounted for 73.63% and 26.37% of the model’s predictive power, respectively, underscoring the necessity of incorporating anthropogenic information in spatial modelling. Under four Shared Socioeconomic Pathways (SSPs), the global annual maximum wetland extent is projected to expand from a historical baseline of 8.02 million km2 to between 8.37 and 9.12 million km2 by the end of the century (2086–2100), representing a net increase of 4.35%–13.67%. However, this aggregate growth masks profound spatial heterogeneity: severe degradation hotspots emerge at high northern latitudes, notably Western Siberia, where projected area contraction exceeds 143,000 km2, while expansion dominates tropical basins, with Africa’s net gains reaching 403,537 km2 under SSP5-8.5. These findings demonstrate that integrating anthropogenic features substantially refines wetland forecasting, providing robust scientific support for global wetland conservation strategies and assessments of the United Nations Sustainable Development Goals (SDGs).