Methods

DisALEXI was recently ported to Google Earth Engine as part of the OpenET framework, and refinement is still ongoing. The baseline ALEXI/DisALEXI model structure is described by Anderson et al. (2012, 2018). DisALEXI uses Landsat thermal and optical imagery to spatially disaggregate continental-scale, 4-km resolution ET estimates from ALEXI (Anderson et al. 2007)—an energy balance algorithm based on morning land surface temperature rise data obtained over the U.S. from the GOES geostationary satellite network.

Results from Phase I of the OpenET Accuracy Assessment and Intercomparison study suggested a few modifications to the baseline ALEXI/DisALEXI modeling system that were implemented in Phase II (a description of the OpenET Accuracy Assessment and Intercomparison study is available here). First, a wet bias in arid grasslands in the western U.S. led to modification of the soil resistance formulation in the Two-Source Energy Balance (TSEB) land-surface representation (Norman et al., 1995) used in ALEXI, building on the findings of Kustas et al., (2016).

However, Phase II results indicate that this adjustment must be tempered in some regions; for example, in arid grassland pixels containing small fractions of irrigated pasture or croplands. Further refinement has been implemented in the current version, leading to more reasonable fluxes in these ecoregions

Time-varying misregistration of GOES-based inputs to ALEXI between GOES series had resulted in in spatial artifacts in DisALEXI in the previous collection, particularly along strong moisture discontinuities (e.g. borders between irrigated land area and dry grasslands).  Registration has been improved and regularized in time, reducing these artifacts in the current version.

Additional future refinements will include the slope and aspect corrections to solar radiation load used in eeMETRIC, which should improve disaggregation in regions of high topographic variability. In addition, a Penman-Monteith version of the TSEB is under evaluation and will be implemented in future collections pending satisfactory performance.

eeMETRIC applies the traditional METRIC calibration process of Allen et al. (2007; 2015) and Irmak (Kilic) et al. (2013), where a singular relationship between the near surface air temperature difference (dT) and delapsed land surface temperature (TsDEM) is used to estimate sensible heat flux (H) and is applied to each Landsat scene. Automated selection of the hot and cold pixels for an image generally follows a statistical isolation procedure described by Allen et al. (2013a). The calibration of H in eeMETRIC utilizes alfalfa reference ET calculated from the NLDAS gridded weather dataset using a fixed 15% reduction in computed reference ET to account for known biases in the gridded data set (Blankenau et al., 2020). The fixed reduction does not impact the calibration accuracy of eeMETRIC and mostly reduces impacts of boundary layer buoyancy correction.

The identification of candidates for pools of hot and cold pixels has evolved in the eeMETRIC implementation of METRIC. The new automated calibration process incorporates the combination of methodologies and approaches that stem from two development branches of EEFlux (Allen et al., 2015). The first branch focused on improving the automated pixel selection process using standard lapse rates for land surface temperature (LST) without any further spatial delapsing. The second branch incorporated a secondary spatial delapsing of LST as well as changes to the pixel selection process. The final, combined approach is described by ReVelle, Kilic and Allen (2021a, b). 

Application of the METRIC algorithm near coasts and in mountainous terrain can benefit from special treatment of LST where cool ocean temperatures may influence the relationship between LST and sensible heat flux with distance from the coast, and where lapsing of LST in mountains can vary with location or time of year. Recent efforts by ReVelle et al. (2021a) developed a two-dimensional planar delapsing surface over Landsat images to remove spatial trends in LST that may be related to distance from an ocean, impacts of steep terrain, or frontal weather activity that can cause spatial variation in equilibrium LST. Application of the secondary planar delapsing tends to help equalize LST to represent only the effects that are associated strictly with sensible heat fluxes.

eeMETRIC employs the aerodynamic-related functions in complex terrain (mountains) developed by Allen et al. (2013b) to improve estimates for aerodynamic roughness, wind speed and boundary layer stability as related to estimated terrain roughness, position on a slope and wind direction. These functions tend to increase estimates for H (and reduce ET) on windward slopes and may reduce H (and increase ET) on leeward slopes.

Other METRIC functions employed in eeMETRIC that have been added since the descriptions provided in Allen et al. (2007 and 2011) include reduction in soil heat flux (G) in the presence of organic mulch on the ground surface, use of an excess aerodynamic resistance for shrublands, use of the Perrier function for trees identified as forest (Allen et al., 2018; Santos et al., 2012) and aerodynamic estimation of evaporation from open water rather than using energy balance (Jensen and Allen 2016; Allen et al., 2018). These functions and other enhancements to the original METRIC model are described in the most current METRIC users manual (Allen et al., 2018). eeMETRIC uses the atmospherically corrected surface reflectance and LST from Landsat Collection 2, with fallback to Collection 1 when needed for near real-time estimates. 

Implementation of geeSEBAL was recently completed within the OpenET framework. An overview of the current geeSEBAL version can be found in Laipelt et al. (2021), which is based on the original algorithms developed by Bastiaanssen et al. (1998). The OpenET geeSEBAL implementation uses LST data from Landsat Collection 2, in addition to NLDAS and gridMET datasets as instantaneous and daily meteorological inputs, respectively. The automated statistical algorithm to select the hot and cold endmembers is based on a simplified version of the Calibration using Inverse Modeling at Extreme Conditions (CIMEC) algorithm proposed by Allen et al. (2013), where quantiles of LST and the normalized difference vegetation index (NDVI) values are used to select endmember candidates in the Landsat domain area. The cold and wet endmember candidates are selected in well vegetated areas, while the hot and dry endmember candidates are selected in the least vegetated cropland areas. Based on the selected endmembers, geeSEBAL assumes that in the cold and wet endmember all available energy is converted to latent heat (with high rates of transpiration), while in the hot and dry endmember all available energy is converted to sensible heat. Finally, estimates of daily evapotranspiration are upscaled from instantaneous estimates based on the evaporative fraction, assuming it is constant during the daytime without significant changes in soil moisture and advection.

Based on the results from the OpenET Accuracy Assessment and Intercomparison study, the OpenET geeSEBAL algorithm was modified as follows: (i) the simplified version of CIMEC was improved by using additional filters to select the endmembers, including the use of the USDA Cropland Data Layer (CDL) and filters for NDVI, LST and albedo; (ii) corrections to LST for endmembers based on antecedent precipitation; (iii) definition of NLDAS wind speed thresholds to reduce model instability during the atmospheric correction; and, (iv) improvements to estimate daily net radiation, using FAO-56 as reference (Allen et al., 1998). 

Overall, geeSEBAL performance is dependent on topographic, climate, and meteorological conditions, with higher sensitivity and uncertainty related to hot and cold endmember selections for the CIMEC automated calibration, and lower sensitivity and uncertainty related to meteorological inputs (Laipelt et al., 2021 and Kayser et al., 2022). To reduce uncertainties related to complex terrain, we added some improvements to correct LST and global (incident) radiation on the surface (including the environmental lapse rate, elevation slope and aspect) to represent the effects of topographic features on the model’s endmember selection algorithm and ET estimates.

The geeSEBAL team is currently working to reduce these uncertainties and increase model accuracy for multiple climate conditions and complex topography. Some of the future improvements to geeSEBAL will include: (i) corrections of LST based on slope, aspect, environmental lapse rate to improve global (incident) radiation on complex terrain; (ii) improvements to the hot and cold endmembers selection to estimate the near surface and air temperature difference (dT) in arid and temperate climates with dry summers, where uncertainties in ET estimates are related to elevated LST in bare soil and sparsely vegetated areas; (iii) separate solutions for open water evaporation, including a better representation of the heat transferred to the water column.

The core formulation of the PT-JPL model within the OpenET framework has not changed from the original formulation detailed in Fisher et al. (2008). However, enhancements and updates to model inputs and time integration for PT-JPL were made to take advantage of contemporary gridded weather datasets, provide consistency with other models, improve open water evaporation estimates, and account for advection over crop and wetland areas in semiarid and arid environments. These changes include the use of Landsat surface reflectance and thermal radiation for calculating net radiation, photosynthetically active radiation, plant canopy and moisture variables, and use of NLDAS, Spatial CIMIS, and gridMET weather data for estimating insolation and ASCE reference ET. Similar to the implementation of other OpenET models, estimation of daily and monthly time integrated ET is based on the fraction of ASCE reference ET. Open water evaporation is estimated following a surface energy balance approach of Abdelrady et al. (2016) that is specific for water bodies by accounting for water heat flux as opposed to soil heat flux.  

Priestley-Taylor (PT) evapotranspiration was originally developed to represent wet environment ET (ETw) or advection-free potential ET where the vapor pressure deficit (VPD) between the surface and atmosphere approaches equilibrium (Priestley and Taylor, 1972). True advection-free, equilibrium conditions rarely occur in the natural environment, therefore the PT alpha coefficient was developed to account for additional nonequilibrium, advection-based vapor fluxes. On average, alpha was shown to equal approximately 1.26 for wet environments; however, values can fall well above or below the original 1.26 alpha value depending on environmental conditions such as soil moisture, VPD, and vegetation cover (Jensen et al., 1990; Agam et al, 2010; Engstrom et. al., 2002).

To improve PT-JPL ET estimates for croplands, wetlands, and riparian areas in arid and semiarid environments where advection is prevalent, a PT alpha adjustment layer was developed based on the ratio of ASCE reference ET and Priestley-Taylor ETw, following principles of the complementary relationship of evaporation (Kahler and Brutsaert, 2006; Szilagyi, 2007; Huntington et al., 2011). ASCE reference ET and ETw estimates were developed using the gridMET dataset. The alpha adjustment layer was calculated as the ratio of the growing average bias corrected ASCE reference ET to the growing season average ETw. Growing seasons were defined for each gridMET grid cell based on cumulative growing degree days of daily average air temperature from January 1 and killing frost thresholds of 300 C and -2 C, respectively. Adjustment factors were limited to a minimum of 0.79 and maximum of 2 to account for uncertainty in weather data and model performance in locations with extreme climate and weather (e.g. Death Valley, coastal regions). Adjusted alpha values range from 1 to 2.5, and fall within the range of alpha values previously reported across arid and humid settings (Priestley and Taylor, 1972; McAneney and Itier, 1996; Weiß and Menzel, 2008; Yang et al., 2009; Tabari and Talaee, 2011). Application of adjusted alpha values were limited to cropland, wetland, and riparian areas as defined by the USDA cropland data layer (CDL). 

Results from Phase II of the OpenET Accuracy Assessment and Intercomparison study indicated that application of the PT alpha adjustment layer increased overall accuracy of PT-JPL for cropland, wetland, and riparian areas, however, ET estimates generally remain near or below the ensemble average for these areas. Future enhancements to PT-JPL will include the slope and aspect corrections to short and longwave radiation and LST, similar to eeMETRIC, and further improvements to the PT alpha adjustment layer.

The primary change from the most recent SIMS publication (Pereira et al., 2020) for implementation in OpenET is the integration of a gridded soil water balance model to account for soil evaporation following precipitation events. Results of the Phase I intercomparison and accuracy assessment showed that SIMS generally performed well for cropland sites during the growing season, but had a persistent low bias during the winter months or other time periods with frequent precipitation. This result was anticipated, since the reflectance-based approach used by SIMS is not sensitive to soil evaporation. To correct for this underestimation, a soil water balance model based on FAO-56 (Allen et al., 1998) was implemented on Google Earth Engine and driven with gridded precipitation data from gridMET to estimate soil evaporation coefficients. These coefficients were then combined with the basal crop coefficients calculated by SIMS to estimate total crop evapotranspiration. In addition, a modest positive bias was frequently observed in the SIMS data for periods with low or sparse vegetative cover. To correct for this bias, updates were made to the equations that calculate the minimum basal crop coefficient, to allow lower minimum basal crop coefficient values to be achieved. Full documentation of the SIMS model, current algorithms, and details and equations used in the soil water balance model are included in the SIMS user manual. 

Results from Phase II of the OpenET Accuracy Assessment and Intercomparison study indicated that these changes improved SIMS estimates of total actual ET and increased the overall accuracy of SIMS for cropland sites. However, the reflectance-based approach used by SIMS assumes well-irrigated conditions, and SIMS will still have a positive bias for deficit irrigated crops and croplands with short-term or intermittent crop water stress. At present, SIMS is only implemented for croplands, and future research will extend the vegetation density-crop coefficient approach used within SIMS to other land cover types.  

The Operational Simplified Surface Energy Balance (SSEBop) model by Senay et al. (2013, 2017) is a thermal-based simplified surface energy model for estimating actual ET based on the principles of satellite psychrometry (Senay 2018). The OpenET SSEBop implementation uses land surface temperature (Ts) from Landsat (Collection 2 Level-2 Science Products) with key model parameters (cold reference, Tc, and surface psychrometric constant, 1/dT) derived from a combination of climatological average (1980-2017) daily maximum air temperature (Ta, 1-km) from Daymet, and net radiation data from ERA-5. This model implementation uses the Google Earth Engine processing framework for connecting key SSEBop ET functions and algorithms together when generating both intermediate and aggregated ET results. A detailed study and evaluation of the SSEBop model across CONUS (Senay et al., 2022a) informs both cloud implementation and assessment for water balance applications at broad scales. Notable model (v0.2.6) enhancements and performance against previous versions include additional compatibility with Landsat 9 (launched Sep 2021), global model extensibility, and improved parameterization of SSEBop using the Forcing and Normalizing Operation (Senay et al., 2022b) to better estimate ET in all landscapes and all seasons regardless of vegetation cover density, thereby improving model accuracy by avoiding extrapolation of Tc to non-calibration regions.