Three-dimensional canopy morphology and wind dynamics govern global rainfall interception

Accurate simulation of rainfall interception is critical for understanding the global water cycle, yet most earth system models rely on simplified, leaf area index (LAI)-centric vegetation representations that neglect the three-dimensional (3D) structure of plant canopies, introducing substantial structural uncertainty. Here we introduce Feature-Constrained Deep Symbolic Regression (FC-DSR), an interpretable artificial intelligence framework that derives explicit parameterizations for canopy interception. Global analyses show that 3D canopy morphology (canopy depth and width) and wind dynamics exert stronger control on interception than LAI. The resulting morphology-based parameterizations substantially outperform conventional LAI-based approaches, increasing site-level interception loss Kling–Gupta efficiency by 0.27–0.39. Budyko-based water-balance analyses confirm the improved interception scheme, with enhanced evapotranspiration at nearly 70% of global flux sites and stronger runoff dynamics across six major river basins. Implementation of the new scheme in the CoLM land surface model improves runoff simulations for 71.5% of U.S. catchments and better reproduces discharge variability in representative sites. Global applications further reveal highly nonlinear responses of interception to changes in rainfall intensity and vegetation structure. Together, these results highlight fundamental limitations of LAI-centered models and demonstrate that explicit representation of 3D canopy architecture is critical for reliable prediction of the terrestrial water balance.