We present a model to retrieve actual evapotranspiration (ET) from
satellites' vegetation indices (Parameterization of Vegetation Indices for
ET estimation model, or PaVI-E) for the eastern Mediterranean (EM)
at a spatial resolution of 250 m. The model is based on the empirical
relationship between satellites' vegetation indices (normalized difference vegetation index (NDVI) and enhanced vegetation index (EVI) from
MODIS) and total annual ET (ET
Actual evapotranspiration (ET) is a primary component of the global water cycle. Its assessment at global and regional scales is essential for forecasting future atmospheric feedback (Jung et al., 2010; Oki and Kanae, 2006; Zemp et al., 2014). Estimating ET at such scales, however, is not straightforward and requires the use of models (Chen et al., 2014; Hu et al., 2015; Jung et al., 2009; Trambauer et al., 2014). Data-driven models using satellite information benefit from a continuous spatio-temporal direct observation of the land surface (Ma et al., 2014; Shi and Liang, 2014).
Satellite-based ET models are classified into two categories: (1) empirical, using the relationship between in situ ET and satellite-derived vegetation indices (VIs) (Glenn et al., 2011; Nagler et al., 2012; Tillman et al., 2012), and (2) physical, using surface temperature from satellites to solve energy balance equations (Anderson et al., 2008; Colaizzi et al., 2012). Some models combine the two approaches (Tsarouchi et al., 2014).
Although physically based models are much more common their performance is comparable to that of the empirically based models (Glenn et al., 2010). The accuracy of both approaches is within that of the eddy covariance measurements (70–90 %) used for their calibration or validation (Kalma et al., 2008). Yet the empirical approach is simpler than the physically based model and requires less additional information.
The basis for the empirical model is the resource optimization theory. This theory suggests that plants adjust their foliage density to the environmental capacity to support photosynthetic activity and transpiration (Glenn et al., 2010). Accordingly, changes in vegetation foliage cover (and VIs) would affect ET, resulting in high ET–VI correlations. Then, the empirical equation could be used to retrieve ET in space and time.
This approach is mostly used in vegetation systems with an annual cycle of growth and drying where VIs define well the phenological stages (Glenn et al., 2011; Senay et al., 2011). However, in complex systems comprised of annual (i.e. herbaceous) and perennial (i.e. woody) vegetation the model must be adjusted with additional meteorological data (Maselli et al., 2014).
The main drawback of the empirically based approach is that it is limited to a specific site and vegetation type (Glenn et al., 2010; Maselli et al., 2014; Nagler et al., 2012). No common relationship was found between ET and VIs for different sites and climatic conditions.
Here we used MODIS VIs and land surface temperature (LST) products and eddy covariance ET from 16 FLUXNET sites with different plant functional types to establish empirical relationships between VIs (and/or LST) and ET in Mediterranean vegetation systems. We first examined those relationships in annual vegetation systems and complex systems comprising both annual and perennial vegetation. Three empirical models were examined with 16-day and mean annual data: (1) simple and (2) multiple-variable regressions, and (3) a modified version of the Temperature and Greenness (TG) model . We used a performance-simplicity criterion to choose the best model to retrieve ET for the EM. Models' estimates were compared with two ET operational products: MODIS (MOD16) and the Land Surface Analysis Satellite Applications Facility product based on MSG satellites' data (LSA-SAF MSG ETa, hereafter MSG). Finally, we evaluated our model against ET calculated from water catchment balances in the EM.
In situ ET was derived from eddy covariance towers that constitute the
international flux towers net (FLUXNET). Two open FLUXNET sources were used
to acquire the data sets: the Oak Ridge National Laboratory Distributed
Active Archive Centre (available online at
We used 16-day normalized difference vegetation index (NDVI) and enhanced vegetation index (EVI) at a spatial resolution of 250 m (MOD13Q1) and 8-day LST at 1 km spatial resolution (MOD11A2) from MODIS on board the Terra satellite. Although Terra provides LST twice a day (around 10:30 and 22:30 local time) here we used only daytime LST, which is the relevant for ET processes. The 8-day LST was averaged to match the 16-day temporal resolution of the VI product.
The MODIS 16-day VI product is a composite of a single-day value selected
from 16-day period based on a maximum-value criterion (Huete et
al., 2002). It represents the vegetation status of the entire 16-day period
because of the gradual development of the vegetation. This enables
regressing the MODIS VI product against 16-day averages of ET. NDVI is defined
as (Rouse et al., 1974)
For model development, time series of NDVI, EVI and LST at each FLUXNET site
were obtained from MODIS Land Product Subsets
(
Description of the 16 selected FLUXNET sites. Horizontal line
divides between the six FLUXNET sites in PA systems (top) and the nine
FLUXNET sites in AN systems (bottom). Plant functional types (PFTs) include
CSH: closed shrublands; WDL: woodland; SAV: savannah; ENF: evergreen
needle-leaved forest; WSA: woody savannah; CRO: croplands; and GRA:
grasslands. Mean annual precipitation (
Water balances from six catchments along the north-to-south
rainfall gradient in the eastern Mediterranean (Fig. S2). Catchments area is
in 10
Model results were compared with two satellite operational ET products in 2011 at 148 main basins in the
eastern Mediterranean. These operational MODIS and MSG ET products are based on different
physical models and have different spatial and temporal resolutions
(1 km/8 day for MOD16, and 3.1 km/daily for MSG)
(Hu et al., 2015). The annual MODIS (MOD16A3)
and daily MSG (LSA-SAF MSG ETa daily) ET products were downloaded for 2011 for the
EM region. The basins map was taken from HydroSHEDS, a mapping product based
on a high-resolution elevation layer developed by the Conservation Science
Program of the World Wildlife Fund (
We evaluated our model with mean annual ET calculated
from water balances at six catchments along a north–south
rainfall gradient (130–800 mm yr
Precipitation data (
For timescales of several years d
Calculating ET from water balances has some drawbacks like the difficulty to properly estimate the spatial distribution of precipitation over the entire catchment and uncertainties of catchment boundaries (Conradt et al., 2013). However, this is the best existing approach to compare in situ ET with satellite-derived ET at a basin scale.
Perennial and annual vegetation in Mediterranean regions have distinct phenology, contributing differently to the VI signal (Helman et al., 2015; Karnieli, 2003; Lu et al., 2003). Here we examined VI–ET relationships in vegetation systems comprising both annual and perennial vegetation (i.e. forests, woodlands, savannah and shrublands, hereafter PA) separately from those comprising only annual vegetation (i.e. croplands and grasslands, hereafter AN).
We found that annual vegetation in the understory of PA systems might contribute significantly to VIs while having very small contribution to the total ecosystem ET. In some cases, this results in an apparent phase shift between ET and VIs (Fig. 1), leading to a negative correlation or a lack of correlation. Moreover, we found that AN sites exhibit one single ET–VI relationship under a wide range of rainfall conditions whereas similar types of PA systems have significantly different ET–VI relationships (i.e. different slopes) under different climatic regimes (unpublished results).
Sixteen-day eddy covariance ET averages (black dots) and MODIS-derived NDVI (green line) at two vegetation systems: (top) PA, i.e. comprising perennial and annual vegetation (evergreen coniferous forest), and (bottom) AN, i.e. only annual vegetation (corn and soybean cropland). Note: in the cropland site (bottom) is the NDVI during the growing season after the annual minimum was subtracted.
Therefore, AN sites (FLUXNET sites in AN systems) were selected from wide
range of climatic regimes, while PA sites (FLUXNET sites in PA systems) were
selected only from Mediterranean climate regions. Selection of the FLUXNET
sites had to fulfil the following criteria: (1) at least 3 years of
satellite and eddy covariance data in the FLUXNET site; (2) missing data
less than 30 days yr
Three regression models were examined using the satellite-derived NDVI, EVI,
LST and eddy covariance ET data:
Simple regressions of ET against VIs or LST with 16-day and annual data. Multiple-variable regressions using NDVI (or EVI) and LST as independent
variables and ET as the dependent variable. Regressions were conducted
with both 16-day averages and annual data. Modified version of the TG model proposed by
Sims et al. (2008) using LST as a
proxy for radiation and potential ET (Maeda et
al., 2011) with 16-day data alone.
We used all models with 16-day ET averages and 16-day VIs and/or LST data
but only the first two models with total annual ET and mean annual VIs
and/or LST because the TG model was designed to work only with 16-day data
(Sims et al., 2008). In AN, we
subtracted the annual minimum VIs before integrating them over the growing
season instead of using the original 16-day VI data (see
Helman et al., 2014a, b, 2015). The integral over the VIs during the growth
season was used in the two first models against total annual ET. Multiple-variable
regressions were applied only on NDVI and LST data or EVI and LST
data, but not on NDVI with EVI data because NDVI and EVI were highly
correlated (
The original TG model is based on the observed correlations between
MODIS-EVI and gross primary production (GPP) measured at several FLUXNET sites, which were further refined by incorporating LST
data (Sims et al., 2008):
Here, we modified the TG model by using ET instead of GPP in Eq. (4):
Pearson's correlation coefficient (
ET was assessed for the eastern Mediterranean using the best models for AN and PA systems separately. To produce the required land cover map, we classified pixels as AN and PA based on their NDVI during the year. Low NDVI during the dry season (< 0.25) implies absent or dry vegetation typical for AN systems (Lu et al., 2003). Yet some PA systems (e.g. open shrublands) also have low NDVI during this period but differ from AN systems by smaller NDVI change (< 0.4) during the growth season (Lu et al., 2003; Roderick et al., 1999).
Hence, we classified pixels with minimum NDVI < 0.25 as AN only if their NDVI increased by more than 0.4 during the growth season. To account for the high NDVI in agricultural fields of the Nile Delta, pixels with minimum NDVI smaller or equal to 0.35 were also classified as AN only if their NDVI increased by more than 0.35. All remaining pixels were classified as PA (Fig. S5).
Although this classification procedure might be coarse, we preferred it to the MODIS land cover product for two reasons. First, a significant discrepancy was found between the MODIS-based land cover product and actual land cover type distribution in the eastern Mediterranean (Sprintsin et al., 2009a). Second, this procedure produces the required mask layer at the spatial resolution of the model (250 m), while the MODIS-derived land cover product is available at a coarser resolution (500 m).
The produced AN/PA land cover map showed the general pattern known for this
region (Fig. S5). Moreover, the total AN area estimated for Israel not
considering the Golan Heights grasslands (i.e. considering mainly Israel's
croplands) was 255
On average, the absolute correlation coefficient (
Relationships between annual ET (mm yr
Correlation coefficients (
LST was negatively correlated with 16-day and total annual ET in all PA
FLUXNET sites. This implies the role of transpiration in attenuating thermal
load (Rotem-Mindali et al.,
2015). Mean annual LST was highly correlated with total annual ET (
Correlation coefficients from the cross-site comparisons were as high as
those from site-specific regressions when using annual data in PA sites
(Fig. 3). Correlations were equally high for both linear and exponential
functions (
Same as Fig. 2 but for all PA sites together. The linear (dashed
line) and exponential (solid line) functions are presented for the ET–VI
relationships, and the
Although a linear regression function is usually preferred to explain simple
relationships between two parameters, the exponential relationship is more
realistic in the case of ET–VIs. This is because VIs exhibit exponential
relationships with the leaf area index (LAI; Baret et al., 1989; Duchemin et al., 2006), which is directly related to water
consumption and ET. Also, ET is usually greater than 0 in places with low
vegetation cover (VIs
The contribution of annual and perennial vegetation to VIs at the sub-pixel level is most difficult to distinguish in PA systems. In some cases, one of those components might have a dominant contribution to VIs, albeit insignificant for the ecosystem flux exchange (Fig. 1). This is probably one of the reasons that VIs could not be used to assess ET at a seasonal timescale (i.e. using 16-day data) in such systems. However, at interannual timescales (i.e. using the annual mean) relationships between ET and VIs were strong and might be used to retrieve total annual ET in PA systems.
In AN, correlation coefficients from the cross-site regressions of ET
against VIs (i.e. the integrals over the growing season period) using the
annual data were comparable to those achieved when using the 16-day data
(Table 4). The
Correlations coefficients (
Correlations did not significantly improve (
The mean absolute error (MAE) for Table 4. The 16-day MAE is in millimetres per day (mm d
In PA, correlation coefficients from the multiple-variable regressions were
substantially higher (
Mean annual ET (2000–2014) from PaVI-E for the eastern Mediterranean.
The modified TG model resulted in significantly higher
NDVI and EVI explained most of the interannual changes in ET in both AN and
PA systems (Table 4). This means that a single ET–VI regression function
could be used to estimate total annual ET in those systems. Multiple-variable
regressions and the modified TG model had higher
Finally, we named this model the Parameterization of Vegetation Indices for ET estimation model (PaVI-E). The mean relative error of PaVI-E was 13 and 12 % for AN and PA, respectively. This is within the accuracy of the eddy covariance measurements that were used for calibration and much lower than those reported for more complex models (Glenn et al., 2010; Kalma et al., 2008). PaVI-E was used to assess total annual ET at a spatial resolution of 250 m for the EM after using the land cover map created for AN and PA as a mask layer (Sect. 3.3 and Fig. S5).
Figure 4 shows the mean annual ET at the EM for the period of 2000–2014. The
annual products of PaVI-E is available by request at 1 km spatial
resolution for the entire EM and at 250 m for Israel
(
ET estimates from PaVI-E were compared with two operational remote sensing
ET products in 148 large basins (> 10 km
Total annual ET for the eastern Mediterranean from PaVI-E, MODIS (MOD16) and LSA-SAF MSG ETa (MSG) for 2011. Grey colouring in MOD16 and MSG indicates missing data.
However, some discrepancies also exist. MOD16 estimates were lower along the
EM coast compared to PaVI-E and MSG. ET estimates from MSG were higher along
the eastern coast especially to the east of the Galilee Sea (mean ET of
Total annual ET at 148 eastern Mediterranean basins (Fig. S1) from
MODIS (MOD16) and LSA-SAF MSG ETa (MSG) vs. PaVI-E. The slope (a),
intersection (b), Pearson's (
(Left) Scatter plot of the mean annual ET (2000–2013) retrieved from PaVI-E and calculated using the water balance equation at six catchments along the EM north–south rainfall gradient (Fig. S2). (Right) Comparison between mean annual ET estimates from PaVI-E, MOD16, MSG and the water balances in those six water catchments. MA-N, MA-CS and MA-S stand for the northern, central-southern, and southern parts of the Mountain Aquifer of Israel, respectively.
Comparing the three models at a basin scale resulted in good agreement
between them (
The relatively higher (lower) MOD16 estimates in xeric (mesic) Mediterranean areas (Fig. 6) was already pointed out by Trambauer et al. (2014), who compared this product with several independent ET models. Furthermore, comparison of MOD16 and MSG ET products in Europe showed that correlations with in situ ET (from 15-eddy covariance sites) were better for MSG (Hu et al., 2015), and that MOD16 underestimates ET in Mediterranean dry regions similarly to the observation in this study (Fig. 5).
ET estimates for PaVI-E were evaluated against ET calculated from six water
catchments along rainfall gradient in the EM. PaVI-E
estimates were highly correlated with the ET calculated from water balances
(
As shown in Fig. 5, ET estimates derived from PaVI-E are significantly higher than those from MOD16 and MSG in the dry areas of the EM. This is due to the exponential functions used in PaVI-E (Eqs. 7 and 8). It derived a comparable ET to the calculated from the water balance equation at the dry catchment of Mamashit with a slight overestimation of 15 mm (< 15 %, Fig. 7). MSG largely underestimated the calculated ET in Mamashit (by more than 85 %), while MOD16 had no data for this area.
Three empirical VI-based ET models using only eddy covariance ET and MODIS
vegetation indices and land surface temperature data for Mediterranean
vegetation systems were tested. VIs in vegetation systems comprising mostly annual
vegetation (i.e. grasslands and croplands) had strong relationships
with intra-annual (16-day ET averages) and interannual (total annual ET) ET
estimates. The mean relative error was larger for intra-annual relationships
compared to interannual relationships (32 compared to 12 %). In complex systems
with annual and perennial vegetation (i.e. forests, woodlands, savannah and
shrublands) ET–VI relationships were strong only at interannual timescales
(i.e. using annual data). Intra-annual relationships were poor probably due
to the mixed VI signal contributed by annual and perennial vegetation that
constitute different vertical layers in those systems.
While annual vegetation (mostly herbaceous vegetation in the understory) is
the main contributor to the intra-annual VI change, it constitutes only a
minor contributor to the total ecosystem ET in complex Mediterranean systems.
Multiple-variable regression and a modified version of the TG model with VIs and LST were
not significantly better than the simple ET–VI model for both PA and AN
vegetation systems (
The empirical ET–VI model, PaVI-E, had comparable estimates to MOD16 and LSA-SAF MSG ET models in the eastern Mediterranean. PaVI-E also agreed well with ET calculated using the water balance equation at six catchments along the south–north EM rainfall gradient. PaVI-E is the first ET model with such high resolution (250 m) for this region. Its advantage is in its simplicity and spatial resolution compared to the coarser resolutions of MOD16 and LSA-SAF MSG ETa products (1 and 3.1 km, respectively). We are confident that using PaVI-E will enhance the hydrological study in this region, where ET plays a major role in the hydrological cycle.
This research was supported by the Israel Hydrological Service (IHS) Water Authority (Grant no. 4500962964). David Helman greatly acknowledges personal grants provided by the JNF-Rieger Foundation, USA, and the IHS. The authors thank two anonymous referees and Janne Rinne for their comments and suggestions that improved the previous version of this manuscript. This study used eddy covariance data acquired by the FLUXNET community and in particular by the following networks: AmeriFlux (US Department of Energy, Biological and Environmental Research, Terrestrial Carbon Program), CarboItaly and CarboEuropeIP. MODIS subsets and tile land products were acquired from the Oak Ridge National Laboratory Distributed Active Archive Centre (ORNL DAAC) and the US Geological Survey (USGS). Edited by: J. Rinne