MODIS Evapotranspiration Downscaling Using a Deep Neural Network Trained Using Landsat 8 Reflectance and Temperature Data

The MODIS 8-day composite evapotranspiration (ET) product (MOD16A2) is widely used to study large-scale hydrological cycle and energy budgets. However, the MOD16A2 spatial resolution (500 m) too coarse for local regional water resource management in agricultural applications. In this study, we propose a Deep Neural Network (DNN)-based downscaling approach generate 30 m ET using Landsat 8 surface reflectance temperature AgERA5 meteorological variables. model was trained at 500 as reference applied resolution. tested on 15 images over three sites United States compared with classical random forest regression that has been often downscaling. All evaluation sample sets DNN had higher R2 lower root-mean-square deviations (RMSD) relative RMSD (rRMSD) (the average values: 0.67, 2.63 mm/8d 14.25%, respectively) than (0.64, 2.76 14.92%, respectively). Spatial improvement visually evident both downscaled maps MOD16A2, while DNN-downscaled appeared more consistent land cover variations. Comparison situ measurements (AmeriFlux) showed better accuracy, of 0.73, 5.99 rRMSD 48.65%, (0.65, 7.18 50.42%,