Machine Learning-Based Analysis of Atmospheric Process Sensitivity to Land Surface Properties Over Complex Terrain

This study investigates the sensitivity of numerical simulations of atmospheric processes over complex terrain to land surface model (LSM) parameters, focusing on thermally driven circulations in an idealized valley. The sensitivity analysis is performed using the Weather Research and Forecasting model coupled with the Noah-MP LSM, for forest and grassland land cover classes. An ensemble of 400 simulations for mixed forest and grassland is emulated with a suite of machine-learning regressors, and variance-based global sensitivity indices are computed from the surrogates. Sensitivity is evaluated for near-surface wind, potential temperature, turbulent kinetic energy (TKE), and surface heat fluxes at representative valley locations. Results clearly discriminate influential and non-influential parameters, providing guidance in identifying those that most affect model results and therefore should be treated with the most care. Over forests, stomatal minimum resistance (RS) is the most influential parameter overall, followed by canopy height (HVT), with importances intensifying during the night. Near-infrared leaf reflectivity (RHOL_NIR), leaf area index, and optimal transpiration temperature (TOPT) also result as important parameters. Over grasslands, by contrast, both leaf and stem (SAI) area indices dominate, followed by RHOL_NIR. The roughness length (Z0MVT) is the most important parameter affecting turbulent kinetic energy during daytime, while RS and HVT are largely non-influential. The results highlight that the proposed approach, integrating variance-based global sensitivity analysis with machine learning, represents a reliable framework to perform sensitivity analyses for complex numerical models, providing valuable insights into the dynamic interactions between land surface parameters and atmospheric conditions over complex terrain.