Process-informed ensemble machine learning reveals environmental controls on active dust source areas over the Mongolian Plateau

Dust storms from the Mongolian Plateau threaten air quality, ecosystems, and human health across Northeast Asia. Although satellite-based dust footprints and hotspot-frequency products have advanced dust-source identification, they may not fully distinguish persistent emission-prone areas from transport and deposition regions, and the controls on dust-emission susceptibility remain insufficiently resolved. We developed a process-informed framework that combined multi-year spring (2000−2023) Weather Research and Forecasting model coupled with Chemistry (WRF-Chem) dust-emission diagnostics with ensemble machine learning to map dust-emission susceptibility. Source labels were defined using a two-out-of-three rule based on mean emission, emission extremes, and modelled emission-occurrence frequency. Twelve physiographic, hydroclimatic, and vegetation predictors were used to train random forest, XGBoost, LightGBM, explainable boosting machine, NGBoost, and a weighted ensemble. Models were evaluated using Spatial+ cross-validation, which constrains geographic proximity and feature similarity, and the ensemble map was spatially compared with Dust Storm Detection Index (DSDI)-derived spring dust-storm footprints for 2000–2021. The ensemble achieved the highest and most stable performance. It revealed a persistent southwest-to-south-central core of high susceptibility, surrounded by less stable transition zones. Wind speed was the strongest positive predictor, whereas soil moisture, normalized difference vegetation index, and precipitation reduced susceptibility; temperature, topography, and soil texture had secondary effects. Recurrent dust-storm footprints showed spatial correspondence with high-susceptibility areas, supporting the map as a long-term potential-source layer rather than an event-scale map of active sources. This transferable framework can provide a process-informed susceptibility layer for dust modelling, early warning, and land-restoration planning when combined with event-day meteorological information.