Spatiotemporal downscaling approaches for monitoring 8-day 30 m actual evapotranspiration

Abstract

Continuous monitoring of actual evapotranspiration (ET) is critical for water resources management at both regional and local scales. Although the MODIS ET product (MOD16A2) provides viable sources for ET monitoring at 8-day intervals, the spatial resolution (1km) is too coarse for local scale applications. In this study, we propose a machine learning and spatial temporal fusion (STF)-integrated approach in order to generate 8-day 30m ET based on both MOD16A2 and Landsat 8 data with three schemes. Random forest machine learning was used to downscale MODIS 1km ET to 30m resolution based on nine Landsat-derived indicators including vegetation indices (VIs) and land surface temperature (LST). STF-based models including Spatial and Temporal Adaptive Reflectance Fusion Model and Spatio-Temporal Image Fusion Model were used to derive synthetic Landsat surface reflectance (scheme 1)/VIs (scheme 2)/ET (scheme 3) on Landsat-unavailable dates. The approach was tested over two study sites in the United States. The results showed that fusion of Landsat VIs produced the best accuracy of predicted ET (R2=0.52–0.97, RMSE=0.47–3.0mm/8days and rRMSE=6.4–37%). High density of cloud-clear Landsat image acquisitions and low spatial heterogeneity of Landsat VIs benefit the ET prediction. The downscaled 30m ET had good agreement with MODIS ET (RMSE=0.42–3.4mm/8days, rRMSE=3.2–26%). Comparison with the in situ ET measurements showed that the downscaled ET had higher accuracy than MODIS ET.