A data fusion approach for monitoring remotely sensed seasonal evapotranspiration

Landsat is widely applied for monitoring remotely sensed evapotranspiration (ET) because of the four-decade archive of satellite images that records visible, near-, mid- and thermal-infrared information of the Earth surface at moderate spatial resolution (30 to 100 m). However, the 16-day gap between subsequent Landsat images limits its ability to quantify seasonal ET-particularly in cloud-prone areas. Hence, we developed an ET fusion model that integrates the coarser, more frequently available moderate resolution imaging spectroradiometer (MODIS) images with Landsat images using simple linear regression models. Inputs for the Landsat-MODIS fusion model include MODIS land surface temperature and normalized difference vegetation index (NDVI) data, Landsat-based evaporative fraction maps generated using the mapping evapotranspiration at high resolution with internalized calibration (METRIC) model, and land cover information. The Landsat-MODIS ET fusion model generates ET maps with MODIS temporal and Landsat spatial resolution. Eight Landsat and 31 MODIS images from 2008 were utilized to derive watershed-scale annual ET for the Fish River Watershed in AL using the Landsat-MODIS ET fusion model. Mean annual ET for the watershed was estimated within 4% of annual ET estimates from a water balance method. Results showed that the mean annual ET estimates were improved by 25% and 11%, when compared to those from a non-fusion Landsat-only approach and MOD 16 ET products, respectively, with annual ET reference data coming from a water balance method. In addition, pixel level evaluation using measured ET data from a United States Geological Survey (USGS) station in FL showed significant improvement in daily and seasonal ET estimates, when results were compared to those from the non-fusion Landsat-only approach. Mean absolute error for seasonal ET was improved by 7% (11% to 4%), while daily ET estimates were improved by 38% (0.77 to 0.48 mm/day) 124% (0.33 to 0.74) and 32% (0.57 to 0.75), respectively, using root mean squared error, Nash-Sutcliffe efficiency and coefficient of determination as model-evaluation criteria. Hence, the fusion model has the potential to be an operational moderate resolution ET monitoring tool, particularly in regions where Landsat availability is influenced by persistent cloud cover.