A machine learning approach coupling immediate and lagged drivers to predict sand and dust storms in Northern China

Under global climate change, sand and dust storms (SDS) in Northern China have exhibited new spatiotemporal evolution characteristics. To investigate SDS driving mechanisms, this study integrated multi-source remote sensing and reanalysis data (2005-2024), employing SHapley Additive exPlanations (SHAP), receiver operating characteristic curve analysis method, and the extreme gradient boosting machine learning model. The results indicated that SDS frequency increased significantly after 2015, with a 47.6% rise in annual events (from 4 to 6 events/year), while high-intensity events declined from 23.8% to 14.5%, revealing increasing frequency but decreasing intensity. SDS occurrence is primarily triggered by the synergy between strong wind hours and the concurrent VHI. The lag period of eco-hydrological factors such as vegetation, temperature, and moisture on SDS was approximate 5 months. SHAP analysis revealed VHI is dominant driver in hyper-arid and arid transition zones, while soil moisture and evapotranspiration became the primary controls in semi-arid regions. Critical thresholds with narrow 95% confidence intervals for dust occurrence were identified, including VHI below 0.35, strong wind hours exceeding 12.7 h, NDVI below 0.10, and evapotranspiration below 10.8 mm. When VHI exceeds 0.35, the relative SDS risk is reduced by 88.4%, underscoring its potent protective effect. The dynamic thresholds and high-precision prediction model achieved robust performance with an out-of-sample PR-AUC of 0.843, together with strong predictive accuracy as evidenced by a test AUC of 0.935 and annual-scale R² greater than 0.7, providing a scientific basis for accurate dust forecasting and risk management.