An empirical approach to retrieve monthly evapotranspiration over Amazonia
Robinson I. Negrón Juárez1, Rong Fu1, Ranga B. Myneni2, Robert E. Dickinson1, Sergio Bernardes3, Huilin Gao1, Carlos Nobre4,Michael Goulden5, Steven C. Wofsy6
1 School of Earth and Atmospheric Sciences, Georgia Institute of Technology, GA, USA. E-mail: , , ,
2Department of Geography, Boston University, MA, USA.
E-mail:
3 Department of Geography, University of Georgia, GA, USA.
E-mail:
4Weather Forecast and Climate Studies Center (CPTEC), National Institute for Space Research (INPE), Cachoeira Paulista, Sao Paulo, Brazil
E-mail:
5Earth System Science and Ecology and Evolutionary Biology, University of California, Irvine, CA, USA.
E-mail:
6Division of Engineering and Applied Science/Department of Earth and Planetary Science, Harvard University.
E-mail:
Intended for Submission to Journal of Geophysical Research-Atmosphere or International Journal of Remote Sensing
Date: January XX, 2007
Author for correspondence
Robinson I. Negrón Juárez
School of Earth and Atmospheric Sciences
Georgia Institute of Technology
Ford ES&T Bldg., 311 Ferst Drive
Atlanta, GA 30332-0340
Phone: (404)894-1624
e-mail:
Abstract.
The regional evapotranspiration (ET) over the Brazilian Amazonia remains uncertain since in situ sites do not cover the entire domain, and its fetch is ~1km. Based upon an improved physical understanding of what controls ET over the Amazonia rainforests from analysis of recentin situ observations, this investigation developed an empirical method to estimate ET over the Brazilian Legal Amazon. Satellite data used in this study include the Enhanced Vegetation Index (EVI) from the Moderate Resolution Imagining Spectroradiometer (MODIS) and surface radiation budget from the International Satellite Cloud Climatology Project (ISCCP) for the period 2000-2004.An empirical model was validated by measurements at four sites located in the Brazilian states of Pará, Amazonas and Rondônia. The observed and calculated values had the same mean and variance. Our results show that the large-scale ET pattern is dominated by changes of surface solar radiation and biome type. At seasonal scale, the regional ET peaks during austral spring (SON). Possible decrease of ET due to deforestation was also observed.
1. INTRODUCTION
Evapotranspiration (ET) is a key component that links climate and terrestrial ecosystem. Over the Amazon forest, ET contributes toabout 50% of total precipitation as calculated by water balance methods and eddy correlation systems [Salati, 1987; Shuttleworth, 1988], however, these studies were limited to smaller areas of the basin and it is not known if there are geographical variations of this rate. Observations provided by the Anglo-Brazilian Amazonian Climate Observation Study (ABRACOS) [Gash et al., 1996] and the more recent Large-Scale Biosphere-Atmosphere experiment in Amazonia (LBA) [LBA, 1996; Avissar et al., 2002; Keller et al., 2004] have improved understanding on the controls of ET at seasonal and interannual time scales. These field studies have shown not only a higher ET in dry season than in wet season, but also a higher ET over areas with less rainfall during dry season in eastern and central Amazonia [e.g., Shuttleworth, 1988; Nepstad et al., 1994; Malhi et al., 2002; Sommer et al., 2002; Souza Filho et al., 2005; Negrón Juárez et al., 2006]. Such a high rate of ET maintained by the rainforest during the dry season plays a central role in determining the date of the subsequent wet season onset [Li and Fu, 2004; Fu and Li, 2004]. A higher ET as a result of forests responding to increased solar radiation can also mitigate the impact of moderately dry anomalies on the surface climate condition [e.g., Negrón Juárez et al., 2006]. Therefore, estimation of basin wide ET at seasonal to interannual scales is essential in determining seasonal and interannual climate variability in the Amazonian region.
Current techniques (e.g., the Eddy correlation technique) used to estimate ET give only point measurements which probably represent ET within 1-2 km in radius as long as the surface characteristics are the same. Due to heterogeneities in land surface and soil characteristics, extrapolation of point data beyond that distance would be inaccurate because of the dynamic and regional variability of ET. Also, observational measurements are not available in remote areas such as the interior western Amazon forests, and numerical model estimates of ET [Potter et al., 2001] are limited by subjective assumptions which are yet to be validated. Satellite data offer an alternativefor estimation of ET over large areas by complementing previous observed measurements and numerical simulations of the ET. A common approach is to relate ET to remotely sensed surface temperature [e.g., Jackson et al., 1977; Jackson, 1988;Hatfield, 1983;Moran et al., 1989;Kustas, 1990] or solar irradiance [Jackson et al., 1983], and several statistical/semi-empirical [e.g. Su et al., 2005; Loukas et al., 2005; Di Bella et al., 2000] or physical methods, based on the Penman-Monteith equation forced by meteorological variables and vegetation indexes [Monteith, 1981; Nishida et al., 2003;Nagler et al., 2005]. While these approaches aim to provide the best possible estimate of ET globally, their application to Amazon forests remains unknown. Although the temperature can be used to estimate ET over some areas in the Amazon the relation between surface temperature and ET varies in sign across the Amazonia. Thus, at central-eastern Amazonia the ET is in phase with temperature [Hutyra et al., 2005; da Rocha et al. [2004], but out of phase in eastern Amazonia [Sommer et al., 2002], and of opposite phase during the wet season over southern Amazonia[Vourlitis et al., 2002]. These observational results clearly suggest that a simple relationship between surface temperature and ET does not exist at the basin scale.
Most of the flux tower observations over Amazonian forests during the last two decades have suggested that ET is primarily controlled by surface radiation and vegetation condition, consistent with that originally reported byShuttlworth [1988]. Therefore, we propose an empirical approach for estimation of ET based on vegetation condition inferred from EVI data [Justice et al., 1998; Huete et al., 2002], and surface net radiation from the International Satellite Cloud Climatology Project (ISCCP) [Zhang et al., 1995, 2004]. Compared to the Normalized Differential Vegetation Index (NDVI), the EVI can better capture canopy structural variation, seasonal vegetation variation, land cover variation, and biophysical variation for high biomass vegetation such as that in the Amazonia [Gao et al., 2000; Huete et al., 2002]. It also performs well under the heavy aerosol and biomass burning conditions [Miura et al., 1998]. The empirical model we used is trained and validated using data collected at eight differentuplandforest sites over the Brazilian Legal Amazon (BLA).
2. Methodology
2.1. Study area and model construction
The study area encompasses the nine states of the BLA (Figure 1) covering an extension of ~5x106 km2 (Brazilian Institute of Geography and Statistics, Latent heat flux (E) data from eightupland sites (listed in Table 1) from the LBA experiment wereused for model construction and validation. Data from K83 and RJA sites were used for training the empirical model,and data from K67 and K34 sites were used for model validation. The sites CRS, DCK, SIN, and BRG were used for model comparison.
Observational data shows that E has a strong linear relationship with measurements of net radiation (Rn). For example, Figure 2 shows the relationship between daily E and Rn at sites K34, K83, and RJA. Latent heat flux represented ~70% of Rn at K34 and K83, but only 40% at the RJA site, showing that for the Amazonia most of Rn is used to sustain ET but at different rates from site to site. Over the dense vegetation of the Amazonian rainforests, ET is primarily contributed by leaf transpiration and evaporation of the intercepted water on canopy. Stomatal closure, hence transpiration rate, is determined primarily by solar radiation at the leaf surface and the availability of water to plants, although vapor pressure deficit and wind speed can also influence ET [Shuttleworth, 1988; da Rocha et al., 2004, Negrón Juárez et al., 2006].
Based on Huete et al. [2002], we considered EVI as a driving factor that characterizesthe forest phenology. Consequently, our empirical model is based on the assumption that ET over the Amazonian rainforests is primarily a function of net radiation and EVI:
ET=f(Rn, EVI) (1)
Different functions were tested independently at each site but the best fit between observed and calculated ET was obtained when
(2)
where EVI is the MODIS EVI data, RnISCCP (Wm-2) is the net radiation from the ISCCP data (see below), and C1, C2, C3, and C4 are constants. For the ET model over the BLA (ETgrl), constants in Equation 2 were adjusted until get simultaneously the maximum determination coefficient (r2) between observed and modeled ET over K83 and RJA. The error between observed and calculated ET was obtained as
Error=(3)
where Y and Y΄ are the observed and calculated ET values, respectively, n is the number of elements in the series, and is the standard deviation of errors.
2.2. MODIS EVI data
The MODIS Vegetation Indices products use surface reflectance from MODIS- Terra corrected for molecular scattering, ozone absorption, and aerosols[Vermote et al., 2002; Miura et al., 2001]. The EVIis an optimized vegetation index developed to enhance the vegetation signal with improved sensitivity in high biomass regions and improved vegetation monitoring through a de-coupling of the canopy background signal and a reduction in atmospheric influences [Justice et al., 1998]. The MODIS EVI is calculated using the red, NIR, and blue reflectances (red , NIR and blue, respectively)
EVI=G(NIR-red)/( NIR+ared-bblue+L) (4)
The coefficients adopted in the EVI algorithm are L=1, a=6, b=7.5, and G=2.5. a and b are coefficients of the aerosol resistance term, which uses the blue band to correct for aerosol influence in the red band, a concept that originated from the Atmosphere Resistant Vegetation Index (ARVI) [Kaufman and Tanré, 1992]. The concept of the soil correction, L, was derived from the soil-adjusted vegetation index (SAVI) [Huete, 1988].
Sixteen-day MODIS EVI images with resolution of 1km x 1km were downloaded from from 2000 to 2005. Monthly data were composed using the number of 16-day average images that overlap the calendar month weighted by the number of actual days that overlap that month. For model construction and validation we used a 3-pixel box centered in the tower location. For regional estimation, the 1km resolution data was aggregated to 0.25°. The aggregation consists of calculating the weighted average of all pixels from the 1 km resolution image that spatially overlaps the output pixel (0.25°) and contributes to the final digital value of this pixel. Figure 3 shows the EVI time series of monthly values at K67, K83, K34, and RJA sites at 1km and 0.25° horizontal resolutions for the period 2000-2004. The T-student test (not shown) pointed out that the average values were the same and the F-test (no shown) was satisfactory suggesting that the spatial pattern at regional scale was maintained after aggregation.
2.3. ISCCP Rn data
Monthly Rn (Wm-2) at the surface was obtained from the upwelling and downwelling shortwave (0.2-5.0 µm) and longwave (5.0-200 µm) radiative fluxes at the surface from ISCCP. This global radiative flux profile data cover the time period from July 1983 to December 2004, has a time step of three hours with horizontal resolution of 2.5°2.5°. Shortwave and longwave radiative flux profiles use data that specify the properties of the earth’s atmosphere and surface. From the balance of full-sky radiative fluxes at the land surface, we calculated the net radiation (RnISCCP). Full-sky fluxes are weighted from fluxes calculated for 15 types of cloud conditions and the clear-sky fluxes by their areas for each cell. Zhang et al. [1995, 2004] present a complete description of this data set.
For model construction and validation we used the RnISCCP encompassing the tower coordinates.For regional estimation of ET, the 2.5° resolution RnISCCP data was interpolated to 0.25° using the Kriging interpolation method [Oliver and Webster, 1990]. Figure 4 shows that RnISCCP at 0.25°0.25° agrees well RnISCCP at 2.5°2.5°, but with some offset respect toobserved Rn, that would be easily counted in the calibration equations. The F-test (an its probability) between observed Rn and RnISCCP at 0.25° at sites K67, K34, K83, and RJA was 1.074 (0.839), 1.569 (0.392), 1.245 (0.559), and 2.667 (0.005), respectively. Excepted by the RJA site, RnISCCP at 0.25° had the same variance respect to observations. A detailed analysis for observed Rn and RnISCCP at 0.25 at RJA showed that the F-statistics for the period January 2000 to December 2001 was 1.9, and its probability was 0.133; therefore observed and satellite series had the same variance. However, for the period 2002 they had different variances that could be associated to bias in satellite measurements or unreported bias in the in situ measurements.
3. Results
3.1. Model training and validation
Figure 5 shows the observed (ETobs) and calculated ET at K83, RJA, K67, and K34 sites. The empirical formula trained for best fit at each specific site (ETsite) generally provide close match of ETobs at seasonal and interannual scales, especially at K83 and K67 sites. However, at RJA site, ETsite tends to overestimate annual maximum ET. The correlation coefficients between ETsite and ETobs are 0.75 at K83 site, 0.73 at K67 site, 0.8 at K34 site, and 0.32at RJA site (Table 2). The apparent higher correlation at K34 site is explained by the shorter time series.
A general empirical model (ETgrl) was obtained with C1=2.1, C2=0.052, C3=1.85, C4=140. Although, some discrepancies are observed between ETgrl and ETobs, the seasonal and interannual variations are in good agreement. The coefficient of determination between ETgrl and ETobsis 0.74, 0.53, 0.39 and 0.31 at K83, K67, K34 and RJA, respectively.The F-test used to determine the variance of observed and calculated values is shown in Table 2. The errors between ETsite (ETgrl) and ETobs at sites K67, K83, K34, and RJA were 9%(17%), 12%(14%), 12%(25%), and 12%(20%), respectively. Using the RnISCCP and EVI at 0.25° spatial resolution, the ETgrlerror respect to ETobs at K67, K83, K34, and RJA were 18%, 19%, 16%, and 16%, respectively. This shows that the 0.25° resolution of satellite inputs would not significantly diminish the quality of the calculated ET.
Figure 6compares the seasonal mean of ETgrl at different spatial resolution respect to ETobs. ETobs includesdata used for model training as well as those reported previously by other scientists at different sites (CRS, DCK, BRG, SIN). Seasonal ETobs correspond to data available listed in Table 1, whereas the ETgrl seasonal averages correspond to the period 2000-2004. At K34, K83, K67, and RJA, the averaged differences at 1km of spatial resolution between ETobs and ETgrl were 8% during the wet season and 10% during dry season, and at 0.25° resolution were similar, 8% and 9% for wet and dry season, respectively. At CRS, DCK, BRG, and SIN, the average discrepancies between in situ ET and satellite estimated ETgrl during the wet season were 26% and 18% at 1 km and 0.25˚ resolution, respectively, and during the dry season are 18% and 19%. The best agreement (~11% discrepancy) occurs at CRS in the central equatorial Amazonia, near to the K34, and the worst at BRG (~30% discrepancy) near the month of the Amazon River where climate is strongly influenced by wind coming from the ocean.
Several factors can contribute to the discrepancies shown in Figure 6. For example, the satellite inputs of the surface solar radiation and EVI may be different from those at the flux towers, in part due to the lack of information on sub-scale structures within the satellite footprints. Theobserved values also have significant uncertainty due to instrument errors as well as the methods used for ET calculations. For example, at the DCK site the ET values were obtained from a combination of observed and calculated data using the Penman-Monteith; at the BRG site ET was calculated using both the Bowen ration energy balance and the Penman-Monteith equation; at the SIN site the reported ET values were obtained using the eddy correlation system and the Priestley-Taylor model.
3.2. Seasonal evapotranspiration
Seasonal ETand precipitation for the period 2000-2004 over BLA are shown in Figure 7. For precipitation we use the Tropical Rainfall Measuring Mission (TRMM) 3B43 which use TRMM data merged with other satellite data and rain gauge available at This dataset has proved to be consistent with observations [Negrón Juárez et al., 2007] and has the advantage of its high spatial resolution. During DJF the maximum values of ET (~3mm day -1, Figure 7a) appear in the north of the Amazonas state, whereas precipitation was higher in southern BLA (>8mm day -1, Figure 7b). In MAM, ET is almost homogeneous over BLA with an average value of 2.6 mm day-1, but some areas in Maranhão State exhibited values of 2.9 mm day-1. During this period the precipitation is concentrated over the northern BLA with values higher than 6 mm day-1 except in the north of the Roraima state where was lower. In JJA a precipitation gradient is observed from south (~1 mm day-1) to north (~ 7 mm day-1). Similar to precipitation, the calculated ET exhibited a gradient with low values (2 mm day-1) in the south and high values in the north, with the highest values (3 mm day-1) observed in the north of the Pará state. In SON the ET had values ranging between 3 and 3.3 mm day-1 except in southern BLA where ranged between 2.2-2.5 mm day-1 and coincided with the domain of Cerrado biome (see Figure 11). In the eastern Amazonia ET was higher than 3.1 mm day-1, but precipitation was lower than 3 mm day-1. This finding is consistent with in situ observations that have proved the importance of root uptake of deep soil moisture to keep the high ET rates [Nepstad et al. 1994; da Rocha et al. 2004; Oliveira et al. 2005]. The increase of ET in eastern Amazonia from September to November suggests an enhanced forest activity that was also observed by Huete et al. [2006]. Contrasting this forest enhancement, over the adjacent deforested areas, the ET had lower value during dry season probably due to shallower rooting depth as also suggested by Huete et al. [2006].
4. Discussion
4.1. Sensitivity of EVI to climate variability
Forest in Amazonia is subjected to severe droughts during El Niño events. Nepstad et al. [2002] conducted an experiment to study the effects of these droughts in an east-central Amazon forest at the Tapajos National Forest, in the Brazilian state of Para. At the canopy level, the authors did no observe substantial drought stress of leaf canopy but a thinning of the leaf canopy. Although the NDVI can be used to monitor the effects on Amazon forest phenology during the El Niño events [Asner et al., 2000], it is sensitive to chlorophyll [Gao et al., 2000] and is asymptotic at high leaf biomass conditions [e.g. Justice et al., 1998; Gao et al., 2002]. On the contrary, the use of EVI seems to be more appropriated since it is sensitive to canopy structure and also performs well even at higher LAI values [Justice et al., 1998; Gao et al., 2002; Huete et al., 2002]. To investigate the response of EVI to climate variability we selected two areas affected differently during the 2005 drought event observed over Amazonia [Marengo et al. 2006]. Figure 8 shows the precipitation (TRMM) and EVI over A1(2°S-5°S,53°W-56°W) and A2 (5°W-7.5°W,65°W-70°W) for year 2004 and 2005. The climatological precipitation (1961-1990) is also showed and was obtained from the Global Precipitation Climatoly Centre. The climatological precipitation at A1 1997 mm. This area encompass the Tapajos National Forest where Nepstad et al. [2002] conducted their experiment. During most of the 2004 year the precipitation was close to climatology, however in 2005 the precipitation was lower than climatology since April, coinciding with a decrease of EVI (Figure 8a). If, as reported by Nespstad et al., the canopy experienced changes in its structure during drought events, this result shows the capability of EVI to detect those changes. However, data as litterfall and LAI are necessaries to be conclusive. On the other hand, the effect of 2005 drought, although was more sever over southern Amazon, appears to have had an opposite effect over A2, since the EVI increased during this year respect to 2004 (Figure 8b). Although in 2005 the precipitation was about 400 mm less than climatology (2578 mm), the decrease was not enough to produce stress on vegetation. On the contrary, with sufficient water supply and more incoming solar radiation a positive response of vegetation is expected. It has to be mentioned that in this area the precipitation have decreased progressively since 2003, with values of 2694 mm, 2364 mm and 2197 mm in 2003, 2004 and 2005, respectively. If this reduction continues the vegetation response is uncertain, since the west part of Amazon does not experience regular dry season and, probably, this forest have not developed strategies to handle droughts [Stokstad, 2005].