Machine learning quantifies dew formation mechanisms across China’s climate gradients

Study region 11 representative flux sites spanning China's major climate gradients (24˚– 43˚N, 91˚–128˚E), from Tibetan Plateau's alpine grasslands to Xishuangbanna's tropical rainforests. Study focus Accurately quantifying dewfall remains a central ecohydrological challenge. Previous data-driven estimations predominantly relied on coarse monthly data, which fundamentally obscures the sub-diurnal physical thresholds of condensation. To address this gap, we developed interpretable machine learning models—Random Forest, Support Vector Regression, and Artificial Neural Networks—using continuous multi-site data to simulate dew formation at daily and monthly scales. We then applied Explainable AI (SHAP, lag, and partial correlations) to systematically decouple the underlying environmental drivers. New hydrological insights for the region Daily-scale modeling is essential for capturing threshold-driven condensation. It successfully avoids the severe overfitting inherent in traditional monthly aggregations, which falsely learn seasonal trends rather than physical mechanics. Furthermore, dew acts as a ubiquitous water subsidy, contributing 0.98–6.72% of annual precipitation. This buffering capacity is disproportionately critical for mitigating short-term drought stress in water-limited and alpine environments. Finally, Explainable AI reveals that RE (ecosystem respiration) /NEE (net ecosystem carbon exchange) a relatively high contribution within the model are not direct causal drivers of condensation, but act as integrated proxies for nocturnal microclimatic conditions.