Evapotranspiration from Remote Sensing to Improve the Swat Model in Eastern Amazonia

In this study, we estimated the evapotranspiration from orbital images MODIS (Moderate Resolution Imaging Spectroradiometer) for assimilation in the hydrological modeling of the SWAT (Soil Water Assessment Tools) model. The data used include the period between October 2003 and December 2006 of the sub-basin of the Lajeado River, located in the Tocantins-Araguaia River basin in Tocantins state. Overall, the results of the use of heat flows estimated by remote sensors in the SWAT model can be considered satisfactory. The values of the COE (coefficient of efficiency of Nash-Sutcliffe) ranged from -0.40 to 0.91 in the comparison with the daily flow data and from 0.17 to 0.77 with the monthly flow data, with the assimilation of evapotranspiration from orbital images. These results indicate benefit to the model adjustment due to improvement in the data assimilated of approximately 0.91 in the COE on daily scale and 0.60 in the CEO on monthly scale.


INTRODUCTION
Improving the accuracy of forecasts of ungauged basins is one of the most challenging tasks in hydrology (Lévesque et al., 2014;Oliveira, 2014;Franks et al., 2005;Goswami et al., 2007;Sivapalan et al., 2003;Zhang et al., 2009).Several researchers have used remote sensing data to better represent the spatial variations of the surface, and there are several that we can highlight.Andersen et al. (2002) used data on leaf area index (LAI) derived from the Advanced Very High Resolution Radiometer (AVHRR) in a distributed hydrologic model and found that LAI can better represent the spatial heterogeneity of plant surface in the simulations of the model in the river basin of Senegal.Garcia-Quijano & Barros (2005) and Gebremichael & Barros (2006) used satellite images to estimate LAI in order to assimilate into a distributed hydrological model, which showed that the daily variation of evaporation and photosynthesis can be captured by the interaction of biophysical and hydrological processes.Zhang & Wegehenkel (2006) added data from remote sensing of LAI in a model of water balance and found a strong control of the LAI seasonal evapotranspiration and runoff in a watershed in Germany.McMichael et al. (2006) showed that the LAI using remote sensing in distributed hydrological modeling can reduce the uncertainty of estimates of flow in a watershed in the semi-arid region of California-USA.Yildiz & Barros (2007) used data of LAI estimated from orbital images in hydrologic models and found that vegetation plays a key role in control of hydrological processes in a catchment area of mid-latitudes.
More recently, Zhang et al. (2009) showed that the use of evapotranspiration estimations from remote sensing applied to the calibration of hydrologic model SIMHYD can improve the modeling daily and monthly ungauged catchment.This demonstrates that the calibration using the evapotranspiration estimated and observed flow improves the results of hydrologic modeling.
In this context, the objective of this study is to estimate evapotranspiration (ETrs) for the tropical region, specifically in the eastern Amazon, from orbital images for use in hydrologic modeling.For this we used the SEBAL model (Surface Energy Balance Algorithms for Land - Bastiaanssena et al., 1998) to generate input information from the hydrologic model SWAT (Soil Water Assessment Tools) to simulate the water balance in the Lajeado River basin in the state of Tocantins, Brazil.It is worth to be mentioned that only few published works address sensing data assimilated in SWAT model, such as Immerzeel et al. (2008) and Immerzeel & Droogers (2008).

Study area
This study was developed in eastern Amazonia in the years 2004 and 2005 (presented).The area of interest for the study is located in the Lajeado basin Figure 1.The data used in this work are from the experimental site "Bananal Island Floodplain" (BIF); these were collected at the north side of the island near the boundaries with the Cantão State Park, with the installation of a 40-meter tall micrometeorological tower, located approximately 2 km east of Javaezinho river, at coordinates 9.824º S, 50.158 º W, and altitude 120 m Figure 1.

Observational meteorological data
The rainfall and Stream flow data used in this work are the stations (code): Piranhas River (949000), Marianopolis (949003), Caseara (950000) and river Javaés (950002) and Ponte Piranha River (27380000) for the period 2000 to 2008 (observational data) measured in the watershed area and made available by the National Water Agency (ANA, 2013).
The observational data are part of the project of large-scale Biosphere-atmosphere in Amazon (LBA).Micrometeorological measurements (hourly and daily) of flows of energy (latent and sensible heat), concentration of water vapor and the basic meteorological variables (air temperature, radiation balance, wind speed and soil heat flux), were obtained through conventional sensors and eddy correlation, respectively.

MODIS time series data
To construct the daily series of ETrs, images were used of MODIS of Aqua satellite (Kaufman et al., 1998)

Evapotranspiration and streamflow models
The evapotranspiration estimation by SEBAL needs little information from the surface to estimate the components of energy balance through remote sensing.The ETrs is obtained as the residue of the equation of energy balance.Equation 1presents the premise of the energy balance model, and SEBAL details may be seen in (Bastiaanssen et al., 1998).
Where: λET -is the evapotranspiration (W m -2 ), Rn -is net radiation at the surface (W m -2 ), H -is the sensible heat flux (W m -2 ) and G -is the flow of heat to the soil (W m -2 ).
Equation 2 becomes latent heat λET instant estimated by SEBAL at the time of passage of the satellite in W/m -2 , ET time ( ET hourly ) in mm / hr.
where; ET hourly is the hourly ET (mm/hr), 3600 is the time conversion from seconds to hours, and λ is the latent heat of vaporization or the heat absorbed when a kilogram of water evaporates (J/kg).
Later the Equation 3, the Reference ET Fraction (ET rF ) is defined as the ratio of the computed hourly ET ( ET hourly ) for each pixel to the reference ET (ET r_hourly ) computed from weather data: where; ET hourly is from Equation 2 (mm/hr) and ET r_hourly is the reference ET at the time of the image from the REF-ET software (mm/hr).ET rF is similar to the well-known crop coefficient, Kc.ET rF is used to extrapolate ET from the image time to 24-hour or longer periods.The transformation to the daily range follows the equation (Equation 4): where; ET r-24 is the cumulative 24-hour ETr for the day of the image.This is calculated by adding the hourly ETr values over the day of the image.
The hydrological cycle simulated by SWAT is composed of four control volumes: (i) surface reservoir, (ii) subsurface reservoir, (iii) underground storage tank -free or shallow aquifer and (iv) underground storage tank -deep aquifer.The contributions from these reservoirs to runoff come from the lateral flow from the soil profile and the return flow of shallow aquifer.The volume of the reservoir that percolates through the subsurface soil profile represents the shallow aquifer recharge.The water that percolates to the deep aquifer does not return to the system.In the SWAT model, the determination of the production of surface water basin is based on the water balance, equation (Equation 5) details may be seen in (Arnold et al., 1996).
( ) Where SWt -is the amount of water in the ground on t, SW -is the initial amount of soil water in mm, R -is the daily rainfall in mm, Q -is runoff in mm, ET -is evapotranspiration in mm, P -is the percolation in mm, q lat -is the lateral flow in mm, q r -is the return flow, and mm is the time in days.

Modeling experiments
Conceptually SWAT subdivides the catchment into sub-basins and a river network based on a digital elevation model (DEM).Based on unique combinations of soil and land use, the sub-basins were further detailed into hydrological response units (HRUs), which were the fundamental units of calculation.A total of 23 sub-basins and 525 HRUs were delineated in the study area.
The simulation scenario consisted of two numerical experiments in order to verify the improved efficiency and applicability of the SWAT model to assimilate the ETrs extracted from remote sensing (in daily scale).It is worth mentioning, that all meteorological data inserted into SWAT followed the daily time scale.

Experiment 1 -contours initial calibration
Experiment 1 was used to select the "default" model, where the weather data were used for the initial condition required to represent physically the location of study.This experiment was the basis for the others, due to its simplicity and effectiveness in representing the initial conditions of the study area.After the first experiment, it was how the model was able to reproduce the observed flow using the statistical tools.

Experiment 2 -manual calibration with evapotranspiration (ETrs)
For the second numerical experiment, the tuning parameters of the model were no longer modified, and therefore, preceded to the assimilation of data evapotranspiration estimated by the model SEBAL from orbital images.

Model assessment
The verification phase of the SWAT model is essential for assessing the performance of the simulation, and therefore the operational use in the management of water resources.The performance of the model was obtained from the observed and simulated values, calculated using the following statistical methods: the coefficient of efficiency of Nash-Sutcliffe (COE) and the correlation between predicted and observed streamflow (CORR), with 10% confidence interval around the absolute error.
The COE can vary from less infinity to 1 (Nash & Sutcliffe, 1970), with 1 indicating the perfect simulation of the simulated event (Equation 6): Where E cal and E ob -are the simulated runoff and observed stream flow respectively, E m -is the arithmetic mean of the observed runoff, and N -is the number of samples.

RESULTS AND DISCUSSION
Having calibrated the model with data for the period 2000 to 2003, the first simulation of the monthly and daily streamflow was tested for the years 2004 and 2005, using the calibration with the initial contours.In general, the SWAT model overestimated the peak streamflow rates, possibly because the model is generating very of base flow and low evaporation.

Experiment 1 -contours initial calibration
After viewing the initial results, we can highlight the difficulty of the SWAT model in the peak discharge and flow basis.For daily values, the peak flows were overestimated, while in the dry season, the simulated

Experiment 2 -manual calibration with evapotranspiration (ETrs)
In general, we can highlight a considered and significant improvement in total flow (about 10%) after the assimilation of the ETrs.Moreover, to emphasize that this flow simulation underestimates the observed stream flow.
In Figure 3, one can observe that there is a decrease in overestimation of the maximum streamflow.The minimum flow simulated reaching close to zero, obtained a significant improvement compared with the observed streamflow.By comparison of results between experiments, it appears that the results with the uptake of ETrs are relatively acceptable, where the simulated flow underestimates the observed flow in percentages lower than 10%.

Assessment of the efficiency experiments
The development of results in the performance verification of the model in the experiments, show that the comparison of daily data resulted in a coefficient of efficiency (COE) from -0.40 in Exp. 1, after the assimilation of the value ETrs of (COE) increased to 0.91 Exp. 2, showing a significant improvement in results of 9.000 m 3 s -1 to deficit of 800 m 3 s -1 , respectively.The monthly results also showed that after the assimilation of the ETrs, the (COE) increased from 0.17 to 0.77, which improved results of deficit 311 m 3 s -1 to less negative of 62 m3 s -1 , respectively.
The results indicate that the use of ETrs Exp. 2, can improve the estimated daily and monthly runoff in ungauged catchments.They also showed good agreement between the modelled mean runoff and ( Setegn et al., 2010;Liew & Garbrecht, 2003;Machado et al., 2003;Baldissera, 2005;Paiva & Paiva, 2006;Lubitz, 2009).Although the general features of the results obtained in this work are common to other cases discussed in the literature, there are peculiarities of each case.

CONCLUSION
In general, the estimated ETrs data from remote sensors and integrated into the SWAT model significantly improved the numerical simulations of streamflow in the ungauged catchments thus demonstrating the wide applicability of remote sensing data for the prediction of flow in a distributed model composed of combined physical soil-plant-atmosphere led to more realistic results of the study area.Even so, the findings of this study show that the use of spectral information from remote sensors of moderate spatial resolution can provide estimates of heat fluxes at the surface and spatial evapotranspiration consistent with use and land cover for assimilation into hydrology models.
The COE values, used as main index template efficiency assessment, were above 0.85% in comparison of daily data and monthly streamflow, after the assimilation of ETrs, indicating a good adjustment.Comparison of daily data that produced the best results was partly explained by the incoming data (with hourly frequency, flawless records).Other problems such as the inability to reproduce the spatial distribution of rainfall or obtaining data from more meteorological stations around the basin, may have contributed to the decrease in the accuracy of the modeling.
, passing at 13:30 local time, with spatial resolution of 250 m (surface reflectance), 500 m (NDVI -Normalized difference vegetation index) and surface temperature at 1000 m for 2004 and 2005 in grid 2° × 2° in the limits of 50° to 48° W and 10 ° to 08 ° S.

Figure 1 .
Figure 1.Map of sub-basins and rain gauge stations and gauged the study area.Figura 1. Mapa das sub bacias e estações de precipitação na área de estudo.
values were very close to those observed.There was a slight delay in the responses of the model, especially at the end of the rainy season.This may be associated with delayed observation of rain in the region with respect to the measurement section Figure 2a.The results in the monthly basis show that the model simulated the seasonal variation, although overestimating the volume of runoff compared with observed values for the runoff model could answer a few rain events Figure 2b.