Improved Evapotranspiration Estimation using the Penman-Monteith Equation with a Deep Learning (DNN) Model over the Dry Southwestern US: Comparison with ECOSTRESS, MODIS, and OpenET
Muhammad Jawad1, Ali Behrangi1, Mohammad Ali Farmani1, Yuan Qiu2,1, Hossein Yousefi Sohi1, Aniket Gupta1, Guo-Yue Niu1
1Department of Hydrology and Atmospheric Sciences, University of Arizona, Tucson, AZ, USA
2Center for Hydrologic Innovations, School of Sustainability, Arizona State University, Tempe, AZ, USA
As one of the major components of the water cycle, accurate estimation of evapotranspiration (ET) at regional scales is challenging, especially over drylands due to the strong soil water constraint. Recently developed remote sensing-based ET products, especially the widely used Priestley Taylor–Jet Propulsion Laboratory (PT-JPL) product (e.g., ECO3ETPTJPL), tend to overestimate ET in the US southwest drylands. In contrast, the Moderate Resolution Imaging Spectroradiometer (MODIS) based product (MOD16A2) underestimates ET. This study presents a hybrid approach that integrates physical modeling and machine learning to estimate daily actual ET at 500 m over the state of Arizona. We develop an efficient Penman-Monteith (PM) based model to compute three ET components including canopy interception loss, direct evaporation, and transpiration from three buckets of water including the canopy-intercepted, surface soil, and subsurface soil water, respectively, based on energy balance. The PM model generated ET is then post-processed with a sequential Deep Neural Network (DNN) to improve ET estimates. PM with the Deep Learning (PMDL) model is trained and tested using independent sets of eddy covariance measurements at 114 CONUS-wide AmeriFlux sites with various land cover types and a range of aridity index. We then applied the trained PMDL model to the state of Arizona using remotely sensed surface data from MODIS, near-surface atmospheric data from the Analysis Of Record for Calibration (AORC), and surface albedo from ERA5-Land. The model shows significantly better results than other remote sensing-based products, e.g., MOD16A2, ECO3ETPTJPL, ECO3ETALEXI, and OpenET with reference to the AmeriFlux observations. It shows reasonably improved performance metrics (KGEss > ~0.78 and R2 > ~0.85) at daily and monthly scales over various sites and the state of Arizona.