Evaluation of Sentinel-1, Sentinel-2, and environmental variables for the retrieval of roughness length for momentum (z0m) over agricultural areas

Evapotranspiration (ET) is the transfer of water from the Earth’s surface to the atmosphere. Its estimation at high spatial and temporal resolution is key to monitoring plant water use and ecosystem stress. Remote sensing can provide data at such resolution; however, it cannot measure ET directly. Consequently, models are used to estimate ET from remote sensing data. An important parameter in ET models is the aerodynamic resistance (r_ah), which quantifies the resistance to heat transfer between the land surface and the atmosphere. r_ah depends on roughness length for momentum (z_0m), which determines how much turbulence is generated by wind at the land surface. In this study, we evaluate the capacity of Sentinel-1, Sentinel-2, and several environmental variables (meteorological, climatological, and soil moisture data) to estimate z_0m using machine learning. Models with different variable combinations are trained on 30 AmeriFlux and ICOS agricultural sites and evaluated using both site-based and spatial block cross-validation. In site-based cross-validation, the model including all the variable classes performs best (R² = 0.67; RMSE = 0.032 m), while the SAR-only model performs worst (R² = 0.05; RMSE = 0.054 m). In spatial block cross-validation, the performance of the model with all the features (R² = 0.57; RMSE = 0.036 m) suggests relatively good transferability to new regions. Overall, while individual predictions have a high relative error, the method captures broad patterns of roughness variation.