In semi-arid regions it is crucial to conserve the scarce water resources as much as possible, especially considering the large water demand for crop irrigation. The irrigation applied needs to be optimized, without negatively affecting the food production. Since a large part of the irrigation water is lost through evapotranspiration, it is a must to produce accurate evapotranspiration estimates on the scale of individual fields. The proposed project aims to provide these estimates using a new, innovative approach based on advanced machine learning techniques, which combines local observations, remote sensing data, and detailed simulations.

The goal of this project is to develop a computationally fast method that calculates the surface evapotranspiration and water transport at field-scale (~100m) within a large area of irrigated agriculture, using the Yaqui Valley in Mexico (~225.000 hectares) as a representative case study.

Regional weather models have difficulty in irrigated areas because their spatial resolution is too coarse for the large local differences in surface conditions. Furthermore, a proper theoretical framework is still lacking to calculate the land-atmosphere exchange of heat and moisture under such heterogeneous conditions. To reach the goal of this project, we therefore first aim to simulate the atmospheric flow above irrigated areas with highresolution large-eddy simulations (LESs), which can explicitly resolve all relevant turbulent motions and land-atmosphere interactions but are computationally expensive.

Subsequently, we aim to achieve the same performance with computationally cheaper coarse resolutions by embedding a neural network (NN) within LES. The neural network will be used to detect spatial and temporal patterns from local observations, satellite data, and high-resolution LES, which can then inform coarser resolutions. The outcome of this project will be field-scale evapotranspiration estimates resulting from the NN-enabled LES, which we will compare to existing evapotranspiration products and estimation approaches.