You are on page 1of 15

Journal of Arid Environments 191 (2021) 104513

Contents lists available at ScienceDirect

Journal of Arid Environments


journal homepage: www.elsevier.com/locate/jaridenv

Towards a remote sensing data based evapotranspiration estimation in


Northern Australia using a simple random forest approach
V. Douna a, *, V. Barraza a, F. Grings a, A. Huete b, N. Restrepo-Coupe b, c, J. Beringer d
a
Instituto de Astronomía y Física del Espacio (IAFE, UBA-CONICET), Ciudad Universitaria, Ciudad Autónoma de Buenos Aires, Argentina
b
School of Life Sciences, University of Technology of Sydney, Sydney, Australia
c
Department of Ecology and Evolutionary Biology, University of Arizona, Tucson, USA
d
School of Agriculture and Environment, University of Western Australia, Crawley, Western Australia, Australia

A R T I C L E I N F O A B S T R A C T

Keywords: In this work we have developed a random forest regressor to predict daily evapotranspiration in three eddy-
Evapotranspiration covariance sites in Northern Australia from in-situ meteorological data and fluxes, and satellite leaf area index
Remote sensing and land surface temperature data. The variable analysis for the random forest regressor suggests that leaf area
Random forest
index is the most important one at this temporal scale. A development sample corresponding to the period
Australia
2010–2013 was used, while the year 2014 has been separated for testing. Using this approach, we have obtained
satisfactory performances in the three sites, with RMSE errors around 1 mm/day (and relative RMSEs ∼ 0.3) in
comparison to the measured values. With the final aim of testing the predictive capability of a model that uses
remote sensing data as input, regressors that predict evapotranspiration exclusively from leaf area index and land
surface temperature, have been trained resulting in reasonable performances. The RMSEs over the test set are ∼
1.1 − 1.2 mm/day. In all cases, the errors are comparable to those obtained in previous works that use similar
locations and different methods. When compared to the measured values, the random forest predictions of
evapotranspiration are more accurate than the global MODIS ET 8-day 1 km product, which was used as
benchmark for the performance evaluation of our models, in the three selected locations.

1. Introduction limitation in Australia has been considered as a driver of the decrease in


global evapotranspiration (Jung et al., 2010). Australia has large sea­
Evapotranspiration (ET) is one of the most relevant variables in sonal variability in precipitations and soil moisture, a low population
terrestrial ecosystems. It is a key factor on the water cycle at regional density, and its savannas have been known to be well preserved for
and global scales (e.g. Huntington, 2006; Ukkola et al., 2013; Fisher years, providing a very convenient environment to test models for
et al., 2017, and references therein), as well as in the energy and carbon estimating ET using multi-wavelength remotely-sensed variables. As
cycles among others (e.g. Zhang et al., 2020; Halladay and Good, 2017; evapotranspiration (or the latent heat flux (LE) which is its equivalent in
Mystakidis et al., 2016, and references therein). Its importance has energy) depends on the availability of water in the soil, surface, and/or
triggered great efforts to quantify and predict ET at different spatial and vegetation and also of the necessary energy for driving the change in the
temporal scales. The processes that regulate the evapotranspiration rate, water state, it cannot be remotely sensed directly. However, remote
which depend on the complex structure of the ecosystem, are highly sensing (RS) does provide continuous data on surface properties and
nonlinear and strongly influenced by the environment and the human biophysical variables with various temporal and spatial resolutions, and
activity. In this context, the understanding of the relevant factors that thus it has been an invaluable tool for estimating ET at different spatial
affect ET and its precise quantification are critical for sustainable agri­ scales (e.g. Zhang et al., 2016; Verstraeten et al., 2008; Kalma et al.,
culture, forest conservation, and also for water resources and environ­ 2008; Glenn et al., 2007). Several models have been used in Australia,
mental management. from those relying on vegetation indices (or even microwave indices)
Monitoring evapotranspiration in Australia has been of great interest combined with water and surface energy balance models such as the
during the last two decades (Glenn et al., 2011). In fact, the soil moisture Penman-Monteith (PM) or Priestly-Taylor (PT) equations (e.g. Barraza

* Corresponding author.
E-mail address: vdouna@iafe.uba.ar (V. Douna).

https://doi.org/10.1016/j.jaridenv.2021.104513
Received 18 February 2020; Received in revised form 16 March 2021; Accepted 18 April 2021
Available online 25 May 2021
0140-1963/© 2021 Elsevier Ltd. All rights reserved.
V. Douna et al. Journal of Arid Environments 191 (2021) 104513

et al., 2017; Guerschman et al., 2009; Cleugh et al., 2007), to more National Flux Network (OzFlux, www.ozflux.org.au, Beringer et al.,
complicated approaches such as variational data assimilation (VDA) 2016), which is the Australian and New Zealand regional network of the
schemes (combined source or dual-source), which include both the dy­ FLUXNET international micrometeorological network (Baldocchi et al.,
namics of the surface energy balance (SEB) models and that of land 2001). From the available in-situ variables, as a first approach we have
surface temperature (LST) to estimate LE (e.g. Barraza et al., 2019). selected those that are commonly used as predictors in other machine
Remote sensing global ET products, as the MODerate Resolution learning or physically motivated models to estimate ET. In views of
Imaging Spectroradiometer (MODIS) MOD16 (e.g. Mu et al., 2007, developing an operational model, among all the existing machine
2011), are available providing decades of data. However, distinct per­ learning techniques we have implemented a random forest because it is
formances have been informed for existing global ET models when versatile, robust to outliers and overfitting, provides reliable feature
applied to different regions (e.g. Majozi et al., 2017; Ramoelo et al., importance estimates and requires low computational time. The specific
2014, and citations therein), which still opens space to the development goals of our work are: i) to evaluate the importance of the selected input
of new tools of local or regional application. As the global products are variables in the final ET estimation; ii) to explore the possibility of
aimed to provide extensive estimations of evapotranspiration, it is ex­ training a simple and low-cost machine learning model to predict local
pected that their uncertainties at the local and regional scales are evapotranspiration from remotely-sensed LAI and LST, and meteoro­
considerable. Barraza et al., (2017) evaluated the possibility of using logical data, iii) to test the accuracy of a model trained only with remote
global meteorological data, and optical and microwave indices to esti­ sensing data, in order to develop an operative product that could
mate LE at regional scale in Northern Australia, obtaining better per­ eventually be of use where in-situ data is not available.
formances than the global product. They point out that the fact that the With these aims, in Section 2 we present the in-situ and remotely-
land surface model and the reanalysis data are validated using eddy sensed data used as input and the selected study areas, we describe
covariance data enhances the performance. Although many models of the models and present the error metrics that we use to quantify their
different complexities for ET estimation have been successfully devel­ performances. We also discuss the selection of the datasets and param­
oped and tested at the local scale, most of them (and also their local eters and the importance of the input variables. In Section 3 we present
validation) rely on the availability of in-situ data. The lack of continuous the errors of the predictions of ET and their comparison to in-situ
in-situ data in some regions reinforces the need of developing tools to measured ET data and also with other existing ET products. In Section 4
predict ET at local and regional scales that rely only on remote sensing we compare the performances of our models to previous works and we
data, that could be of use where local data is not (or no longer) available. discuss the strengths and limitations of our approach. In Section 5 we
Also assessing the importance of each of the predicting variables in summarize the main conclusions of the work and the future prospects.
different ecosystems and its uncertainties may be invaluable for this aim.
Given the lack of a comprehensive understanding of the underlying 2. Materials and methods
processes that regulate ET at regional scale (e.g. Mallick et al., 2018;
Ershadi et al., 2015; Bhattarai et al., 2019, and references therein), 2.1. Site description, in-situ meteorological and eddy covariance
machine learning techniques (as well as multivariate statistics) may measurements
constitute an alternative approach.
In the last decades, machine learning (ML) techniques have been To train the ET algorithm we relied on water fluxes and meteoro­
increasingly used in hydrological and ecological applications (see e.g. logical measurements from three eddy covariance (EC) sites located
Crisci et al., 2012; Raghavendra and Deka, 2014; Park et al., 2016; along the North Australian Tropical Transect ((NATT) Koch et al., 1995).
Kousari et al., 2017; Lary et al., 2018; Cai et al., 2019, and references This ecological and rainfall gradient of more than 1100 km in length
therein), as they allow to mimic the result of complex processes by covers an area of 1.38 million km2 between 12∘ S to 23∘ S and 128∘ E to
automatically reproducing the relations between the input and output 138∘ E. A classic monsoon climate characterizes the region, with reversal
variables. In particular, only in the last few years machine learning of flow from low-level continental-origin easterlies during winter to
techniques were applied to predict ET (or reference ET) in different low-level westerly flow during summer (December–March) (Suppiah,
ecosystems and crops and characterize the influence of distinct relevant 1992). Mean annual precipitation shows a gradual decrease from 1700
variables (Dou and Yang, 2018; Mehdizadeh et al., 2017; Gocić et al., mm in the northern mesic tropics to around 300 mm in the xeric
2015; Shrestha and Shukla, 2015; Abdullah et al., 2015; Petković et al., southern region (Eamus et al., 2013; Hutley et al., 2011; Kanniah et al.,
2015; Chen et al., 2014; Tabari et al., 2013; Huo et al., 2012; Kumar 2011). Rainfall varies inter-annually from less than 30% of the mean
et al., 2011; Jain et al., 2008; Yang et al., 2006). The most explored (coefficient of variation, CV) to more than 60% CV, from north to south,
algorithms are artificial neural networks (ANNs) and support vector respectively (Hutley et al., 2011).
machines (SVMs), but also regression trees (RTs), extreme learning We selected these EC locations that represent different structural/
machine (ELM) and adaptive neuro-fuzzy inference systems (ANFISs) functional forest types (woody savannah vegetation classes) and pre­
have been used to this aim, among others. Relative humidity, radiation, cipitation regimes (see Table 1). From north to south, these are: Howard
wind speed and the maximum and minimum air temperature have been Springs (AU-How), Daly River Savannah (AU-DaS) and Dry River (AU-
considered as ET predictors (e.g. Huo et al., 2012; Jain et al., 2008; Dry). Based on the Australian National Vegetation Information System
Petković et al., 2015). As transpiration has been argued to be a major (NVIS, 2007) land cover types correspond to Eucalyptus woodlands (AU-
contributor to total terrestrial ET (Wang et al., 2014; Wei et al., 2017; How, AU-DaS) and Eucalyptus open forests (AU-Dry) (see Fig. 1). Land
Hoek van Dijke et al., 2019), vegetation characteristics also play a key cover type at each EC site was defined by the dominant class on a 25 km
role on its determination. It is widely known that climate and land cover window with the EC tower at the centre.
have a huge impact on ET, but also radiation and precipitation are Instrumentation across the OzFlux network presents a high degree of
known to affect ET regionally (e.g. Dos Santos et al., 2018; Teuling et al., consistency (see Beringer et al., 2016, for details on the tower configu­
2009). Nevertheless, the relative importance of the regional ET drivers ration and instruments). Flat terrain at the sites (slope ∼ 1∘ or less) and
for a given ecosystem and their influence are still under debate. In this homogeneous vegetation around the towers allow representative EC
sense, machine learning has provided useful tools. measurements (Barraza et al., 2017). More details on the location sites
In this work, we will evaluate a simple machine learning approach, a and instruments are available in the OzFlux website1 and the OzFlux
random forest (RF), to test the potential of this type of algorithms to
predict evapotranspiration in Northern Australia, using LAI and LST
satellite data as well as in-situ meteorological variables and fluxes as 1
http://www.ozflux.org.au/monitoringsites/index.html.
input. The latter were measured in the EC towers of the Australian

2
V. Douna et al. Journal of Arid Environments 191 (2021) 104513

Table 1
Description of eddy covariance flux tower sites used in this study. H: vegetation height of the upper stratum, z: measurement (tower) height, MVG: Australian Major
Vegetation Map (Euc. wood. = Eucalyptus woodlands, Euc. op. for. = Eucalyptus open forests), MP: Measurement period, years of available eddy covariance data, PI:
Principal investigators.
Code Name Lat (◦ S) Lon (◦ E) H (m) z (m) MVG MP PI References

AU-How Howard 12.495 131.150 18.9 23 Euc. wood. 2001–2014 Prof. J. Beringer Eamus et al., (2000)
Springs & Dr. L. Hutley Hutley et al., (2011)
AU-DaS Daly River 14.159 131.388 16.4 23 Euc. wood. 2008–2014 Prof. J. Beringer Hutley et al., (2011)
Cleared & Dr. L. Hutley
AU-Dry Dry River 15.259 132.371 12.3 15 Euc. op. for. 2008–2014 Prof. J. Beringer Sea et al., (2011)
& Dr. L. Hutley

Fig. 1. Land cover map including OzFlux locations (black circles) across the North Australian Tropical Transect (NATT).

special issue.2
Table 2
We used the OzFlux standard processing protocol OzFluxQCv2.7.1
Fluxes and meteorological in-situ measurements used in this study, abbreviations
released under the GNU general public license (2007) by the OzFlux
and units.
community using Python (Enthought Python Distribution version 7.3-1,
Isaac et al., 2017). The data was quality filtered and gap-filled as in In-situ measured variable Abbreviation Units

Restrepo-Coupe et al., (2016). The data processing included removal of Evapotranspiration (daily average) ET mm/day
spikes, rejection of fluxes where more than 1% of 10 Hz observations Precipitation (accumulative daily) Prec mm
Relative humidity (daily average) RH %
were missing from the 30-min average, linear corrections for sensor drift
Wind speed (daily average) WS m/s
and calibration changes, WPL and other corrections, and rejection of Minimum air temperature (daily minimum) Min Ta ∘
C
observations when wind originated from behind the 3D-anemometer Maximum air temperature (daily maximum) Max Ta ∘
C
and tower. The resulting half-hourly EC measurements were Short-wave incoming radiation (daily average) SWin W/m2
post-processed for consistency among sites and to lower uncertainties in Short-wave outcoming radiation (daily average) SWout W/m2
the computed fluxes, which included additional quality control assess­
ment and removal of outliers (Barraza et al., 2018; Restrepo-Coupe
et al., 2016). Only data without precipitation in the last 24 hs was used. 2.2. Remote sensing dataset
We aggregated fluxes and other meteorological data to obtain their
accumulated, mean, minimum or maximum daily values. Land surface temperature (LST) products are provided by Copernicus
Fluxes and meteorological data used in this study and measured at Global Land Service (Copernicus Website). The LST product, based on
each location were: evapotranspiration (mm/day), relative humidity infrared data from sensors onboard different geostationary (GEO) sat­
(%), short-wave incoming radiation (W/m2), short-wave outcoming ellites, consists of instantaneous fields estimated every hour using data
radiation (W/m2), wind speed (m/s), air temperature (∘C) and precipi­ that is representative of such time steps and has a spatial resolution of
tation (mm). In Table 2, the selected measured variables are listed with 0.05∘. The period analyzed was 2010–2014. For the study area the
their corresponding abbreviation and units. Multi-Function Transport Satellite (MTSAT, from 2010 to 2014) was

2
https://bg.copernicus.org/articles/special_issue618.html.

3
V. Douna et al. Journal of Arid Environments 191 (2021) 104513

used to obtain the LST product.3 LST was scaled to 1 km using the decision trees for a set of input variables, has been marginally used in
bi-linear neighbor interpolation. remote sensing applications lately (for example, Otgonbayar et al.,
The MODIS leaf area index (LAI) product MOD15A2 was down­ 2019; Jouybari-Moghaddam and Saradjian, 2019; Zhao et al., 2018;
loaded from the National Aeronautics and Space Administration (NASA) Yang et al., 2017; Wang et al., 2016; Liu et al., 2015; Gleason and Im,
site (http://reverb.echo.nasa.gov/). MOD15A2 presents an 8-day tem­ 2012; Mutanga et al., 2012, among others). RF regression presents many
poral resolution and a spatial resolution of 1 km. To generate the daily advantages, for instance, its simplicity and the few parameters it in­
product we filtered the MOD15A2 product based on QA information and volves in comparison to other ML algorithms, its efficiency over large
used a spline interpolation to obtain daily LAI values. datasets and also its low sensitivity to outliers. A k-fold cross-validation
For both remote sensing products, we extracted a 1 km window scheme was adopted, which consists of using the k-th part of the data as
centred on the location of the flux tower at each site. The mean and the validation sample while the rest is used for training, and repeating
standard deviation of all pixels were assumed to be representative of the the process k times, selecting different subsamples each time. The results
ecosystem. of each evaluation are then averaged to assess the final performance
over the training and validation sets.
2.3. Error metrics The RF regression to predict ET (mm/day) was implemented using
the scikit-learn library on Python language. We used minimum daily air
To evaluate the performance and predictive capability of the RF, and temperature (Min Ta, ∘C), maximum daily air temperature (Max Ta, ∘C),
to compare its predictions to other existing products or models, we have accumulative daily precipitation (Prec, mm), mean daily relative hu­
used the following error metrics, where yi and ŷi are the measured midity (RH, %), mean daily short-wave incoming radiation (SWin, W/
(reference) and predicted values for the i-th sample respectively and N is m2), mean daily short-wave outcoming radiation (SWout, W/m2), mean
the number of samples (i.e. in vectorial notation, y = (y0 , y1 , …, yi , …, daily wind speed (WS, m/s), mean daily LST (LST, K) and daily LAI data
yN− 1 ) and ̂
y = (̂
y0 , ̂
y1 , …, ŷi , …, ̂
y N− 1 )). (LAI, m2 m− 2 ) as input variables. Other remotely-sensed variables could
also be included in the analysis, for instance vegetation indices as NDVI
• Root mean squared error (RMSE): or EVI; considering that these are supposed to be highly correlated with
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ LAI, we decided not to include them. At first, we also discarded the
√ N− 1
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ √ 1 ∑ possibility of using reanalysis data instead of in-situ meteorological data.
y ) = MSE(y, ̂y ) = √
RMSE (y, ̂ (yi − ŷi )2 (1) Although this could in principle support a more operative approach,
N i=0
Barraza et al., (2019) have found large discrepancies between the
ground-based measurements of some meteorological variables and
global meteorological forcing data (GMD) in AU-How, which can
considerably degrade the performance of the ET (or LE) retrieval. As a
• Mean absolute error (MAE):
first approach, in this work we have also eliminated the dates in which
1∑N− 1 any of the input variables was not available, in order not to introduce
MAE (y, ̂y ) = |yi − ŷi | (2) additional hypothesis or biases.
N i=0
We have used the bootstrap sampling method for building the trees
and the MSE criterion to measure the quality of the split. A 5-fold cross
validation was done with the complete dataset corresponding to AU-
• Relative root mean squared error (rRMSE): How in order to select the parameters of the RF regression that mini­
mize the mean RMSE over the validation sets and that is still able to
y) =
rRMSE (y, ̂
RMSE(y, ̂y )
(3) generalize correctly, but always preserving a low computational time.
y For the final RF regressors we used 200 trees and a maximum depth of
30.4 Each RF is trained per site.
with y the mean value of the yi values. As expected, the most important variable to estimate ET in our RF
model is LAI in all the sites (see feature importances in Fig. 2, and e.g.
• Bias: Wang et al., 2014). On the left panel of Fig. 2, the feature importance
1∑N− 1 computed by scikit-learn is shown for each zone. Feature importance is
Bias (y, ̂
y) = (yi − ŷi ) (4) calculated as the decrease in node impurity that each variable in­
N i=0
troduces in the built trees, weighted by the probability of reaching that
The correlations between the measured (reference) and predicted ET node. The node probability can be calculated as the number of samples
values were tested by means of linear fits, whose slope should be close to that reach the node divided by the total number of samples. The higher
unity if the model succeeded to learn the trends of the data. We have also the value, the more important the feature. The rest of the variables are
used both the coefficient of determination R2 , which accounts for the less relevant than LAI by a factor of 2.5–10. In all cases the meteoro­
proportion of the variance in the dependent variable that is predictable logical variables contribute similarly, except for the wind speed in
from the independent variable, and the standard deviation of the AU-How, the incoming and outcoming radiation in AU-Das and the
regression, to evaluate the linearity of the fits. relative humidity in AU-Dry, which appear to be the most important
ones (in increasing order) after LAI. The Pearson correlation between
2.4. Model description and performance LAI and ET was ∼ 0.66; for the rest of the variables the correlation with
ET was not significant (r less than ∼ 0.38). As another test of feature
In order to predict the evapotranspiration given the set of LAI, LST relevance, in AU-How we have also alternatively modified each feature
and meteorological data and fluxes described in the previous sub­ by a list of random numbers in the same range as the originals. The fact
sections, we have applied a Random Forest regression technique (Brei­ that a variable is “important” means that the model relies on it for the
man, 2001) per site. This supervised ensemble learning method that predictions; thus, replacing its values for others and running the model
combines the quantitative predictions of an assembly of independent

4
The results were equivalent to the ones obtained by leaving the maximum
3
For more information about the algorithm see http://land.copernicus.eu/ depth unlimited. The rest of the parameters were left as default, as varying
global/products/lst. them did not show significant improvement in the metrics.

4
V. Douna et al. Journal of Arid Environments 191 (2021) 104513

Fig. 2. Left: Mean feature importances for 50 randomly selected training sets (leaving 20% of the samples for validation in each selection). The period 2010–2013
was used as the development sample. The random selection of the sets was stratified to include 20% of the samples corresponding to each season of each year in the
validation set. Right: Mean increase in the RMSE over the development sample in AU-How for 50 randomly selected training sets when the different features are
replaced by random numbers in the original range one by one (20% of the samples as the validation dataset and the rest for training). Grey dotted line: mean RMSE
over the validation dataset considering all the features for comparison.

over the modified set should increase considerably the model error, as global ET product is to provide ET estimations worldwide, we can expect
the relation between the feature and the predicted variable is altered. its performance to be lower in the selected locations but it somehow
The results of this test are shown on the right panel of Fig. 2. The mean provides an acceptance threshold for the error metrics of our models. In
RMSE over 50 validation sets increases by a factor ∼ 3 when the LAI particular, we have chosen the MODIS ET global product (MODIS Global
data is randomized; while the modification of each of the other variables Terrestrial Evapotranspiration 8-Day Global 1 km, MOD16A2 V105).
implies an increase in the RMSE (on the validation set) that, even in the With the aim to obtain comparable estimations of ET, the ET random
worst cases, is always less than ∼ 40%. forest estimations were averaged in the 8-days MODIS composition
The predominant role of LAI data on the evapotranspiration esti­ period.5 The associated error of the RF results was obtained by propa­
mation suggests the need of evaluating RF models trained using only LAI gation. It is worth to note that other global products or even multi-model
satellite data (MLAI ) as a baseline, in the search for a model that relies algorithms may have been chosen for this aim (see e.g Dirmeyer et al.,
only on remotely-sensed data. On the other side, a model that uses LAI, 2006; Yao et al., 2011, 2014, and references therein). However, as the
LST and all the meteorological variables and fluxes (Mall ) is the one that MODIS product in this work is just used as reference, we did not perform
is supposed to provide the best performance. However, it may not be a deep study of the differences among available products in the study
operative as it uses many in-situ data as input. Thus, we are particularly area.
interested in testing a model that uses LAI and LST remote-sensed data In Fig. 3 we show a diagram of the procedure for the development
(MLAI+LST ) in comparison with the other two. and testing of our models. In Sect.3.1, we present the results of the
Considering the aforementioned parameter selection, we have comparison of our models with the EC-tower measurements for the three
developed RF regressions for each EC site including LAI, LST and all the different study sites. The results of the comparison with the MODIS ET
measured daily variables described in Sects.2.1 and 2.2 as predictors global product are presented in Sect.3.2.
(Mall ), using LAI and daily LST (MLAI+LST ), and also using only LAI data
(MLAI ). The samples corresponding to 2014 were kept as heldout (test), 3. Results
while the period 2010–2013 was used as the development sample, for
training and validation, in each of the zones separately. 20% of the 3.1. Comparison with in situ ET measurements
development sample was chosen as the validation set, while the rest was
used for the RF training. Such an approach reduces the performance In Table 3 the mean error metrics for the Mall model and their de­
considerably with respect to a model that selects random dates from the viations are reported, for the training, validation and heldout sets. The
whole period, both for training and testing. However, we are particu­ RMSE values in the validation sets are around ∼ 0.7 − 0.9 mm/day,
larly interested in exploring the possibility of developing models that while the RMSE is near to ∼ 1 mm/day in the three EC sites in the
may be applied to future temporal extensions of a given dataset, as an heldout set. The rRMSEs for the predicted ET values (RMSE divided by
operational approach. We have also forced the random selection of the the mean ET measured values) for the heldout set (year 2014) are
sets to include 20% of the samples corresponding to each season of each ∼ 30%.
year in the validation set. With this selection constraint we are certain The predicted values of ET as a function of the measured ET values
that the ET values used for training and validation span the whole for each site are shown in Fig. 4 (Time series for the three sites are shown
possible ET range. This procedure was repeated 10 times for each zone, in Fig. S1 in Appendix). For the three EC sites the relation between the
selecting different validation and training sets each time, resulting in 10 predicted and measured ET values is clearly linear for the training and
different RF regressions. The result of each model consists of the mean validation sets (slopes around 0.8 − 0.9 and R2 ∼ 0.9). For the heldout
and standard deviation of ET for the 10 RF regressions. The error metrics sets (2014) the linear fits are less accurate (slopes in the range 0.6–0.8,
were extracted for each model. R2 ∼ 0.4–0.5). In Fig. 5 the monthly mean of the predictions of Mall are
The testing of the models was performed in two levels: i) the final shown in comparison to the monthly mean of the measured ET values.
performance of the models was assessed by comparing the predicted ET The error bars correspond to the standard deviation of the predictions,
values with the tower-measured ones; ii) we have compared the pre­ considering both the fluctuations in the predictions of the 10 RFs and the
dictions with a global ET product for the three sites. As the aim of a daily variations of ET during each month. In the same figure, the relative

5
For the comparison, we have divided the accumulated evapotranspiration
MODIS product by 8 in each of the available dates. Moreover, we have calcu­
lated the 8-days average of the daily ET predicted by the RFs and the 8-days
average of the daily measured ET in the same composition periods as the
MODIS ET product.

5
V. Douna et al. Journal of Arid Environments 191 (2021) 104513

Fig. 3. Workflow of our methodology for the development and testing of the random forest regressor models.

Table 3
Mean error metrics for model that uses all the variables (Mall ) for the training, validation and heldout sets.
RMSE (mm/day) MAE (mm/day) Bias (mm/day) rRMSE

AU-How Training 0.25±0.01 0.18±0.01 − 0.01±0.01 0.08±0.01


(N = 741) Validation 0.67±0.04 0.48±0.03 − 0.01±0.05 0.22±0.02
Heldout 0.99±0.06 0.71±0.05 − 0.02±0.06 0.28±0.02
AU-DaS Training 0.29±0.01 0.16±0.01 − 0.01±0.01 0.14±0.01
(N = 851) Validation 0.79±0.04 0.44±0.03 − 0.03±0.06 0.38±0.05
Heldout 1.03±0.06 0.77±0.05 0.45±0.08 0.31±0.02
AU-Dry Training 0.31±0.01 0.19±0.01 − 0.01±0.01 0.15±0.01
(N = 787) Validation 0.90±0.11 0.54±0.04 0.01±0.08 0.43±0.05
Heldout 1.08±0.04 0.72±0.04 0.09±0.08 0.32±0.01

error in the monthly average estimation of ET for the development and Using MLAI+LST , we obtained RMSEs that are 1.1–1.2 mm/day for the
heldout sets are shown. As it is expected, the predictions follow the heldout set, which represents rRMSEs of 0.3–0.4 (See all error metrics in
seasonality of ET, and also show larger fluctuations during the wet Table S2 in the Appendix). These RMSEs are 10–25% higher than those
season. of the Mall model. All the error metrics corresponding to MLAI+LST reflect
The errors (over the heldout set) for the MLAI+LST and MLAI models in a slightly better agreement of the predictions with the measured values
comparison with the measured ET values are shown in Table 4. For the than those of MLAI . The predicted values of ET as a function of the
MLAI model, which we use as baseline, the RMSE values of the com­ measured ET values for each zone using the MLAI+LST model are shown in
parison with the measured ET values are 1.2–1.4 mm/day over the Fig. 6 (see time series in Fig. S5). The fact that the curves in the heldout
heldout set, while the rRMSEs are around 0.4 for the three sites (see all set, specially in AU-DaS and AU-Dry (for ET ≳ 4 mm/day), deviate
error metrics in Table S1 of the Appendix). The predicted values of ET as considerably from unity is due to the large dispersions and flattening of
a function of the measured ET values for each zone are shown in Fig. S2 the ET values, that are also present and even more evident in the esti­
(see time series for the three areas in Fig. S3). In AU-Dry the estimated mations that use only LAI. The monthly average of the real ET values and
ET values with the model that uses only LAI as input show a constant that of the predictions are shown in Fig. 7, for both development and
plateau. This is due to the fact that in the training and validation sets the heldout sets. Also the relative error of the monthly average of ET is
range of LAI is lower than for the heldout set, and so the response of the shown as an inset panel for both sets. Although Mall shows a better
model to the highest LAI values is not properly reproduced. As a agreement with the observed values during 2014 (heldout) than
consequence for some of the dates in AU-Dry during 2014 the model MLAI+LST , the averaged predictions of MLAI+LST still follow the seasonal
assigns lower ET values that the real ones. The monthly mean of the variations.
predictions are presented in Fig. S4, both for the development and As an example of the possible applications of such a model, we have
heldout sets, in comparison to the monthly mean of the measured ET used the mean of the predictions of MLAI+LST in AU-How and AU-DaS
values. It should be pointed out that if the heldout set (with the same (with a random selection of the heldout set which reduces the errors
number of samples as before) is randomly selected from the complete considerably) to predict ET in a forth FLUXNET site, AU-Ade (Adelaide
dataset, the RMSEs are reduced around 20–30% and the linear trends River, Tropical Eucalyptus woodlands, 2007–2009, Lat: 13.0769, Lon:
between the measured and predicted values are certainly improved in 131.1178, PI: Prof. Jason Beringer, Dr. L. Hutley Sea et al., 2011), which
AU-Dry. is located between the other two in the NATT. Although a quantitative

6
V. Douna et al. Journal of Arid Environments 191 (2021) 104513

Fig. 4. Modelled vs measured ET in each zone (left: AU - How, center: AU - DaS, right: AU - Dry). The points and error bars correspond to the means and standard
deviations of the values of ET predicted by the Mall model. Upper: Training and validation datasets (2010–2013). Lower: Heldout set (2014).

comparison of the predictions with measured ET values is not possible as product also in AU-DaS and AU-Dry, even from MLAI . The lowest rRMSE
there are no ET in-situ measurements available for the period 2010–2014 between the measured values and MODIS product correspond to AU-Dry
in the aforementioned site, in Fig. 8 the monthly averaged ET values for (as the mean of the ET measured values is higher) and the lowest bias to
the four sites are shown. The seasonal behaviour and the relative trend AU-DaS. In the latter location, both the rRMSE and bias between the RF
of ET expected from LAI and LST, with respect to the other sites of the predictions and the MODIS product are the lowest. In the three sites the
transect, are reproduced. underestimation of ET of the MODIS product is non-negligible.

4. Discussion
3.2. Comparison to a global ET product
In this manuscript we have used a machine learning approach to
The resulting temporal series for the development and heldout sets
estimate ET from satellite products (LAI and LST), and in-situ meteoro­
(obtained as described in Sect. 2.4) in comparison with the MODIS ET
logical data, by means of a random forest regression technique.
product for the three selected sites is shown in Fig. 9. As this product is
Australian savannas are largely ecologically intact, with low levels of
an 8-day composite dataset, with the aim of obtaining a comparable
fragmentation, providing an ideal location to study the environmental
estimation of ET, we have divided the accumulated evapotranspiration
determinants of ecosystem growth and functioning along a rainfall
product by 8 in each of the available dates and we have calculated the 8-
gradient (Hutley et al., 2011).
days average of the daily ET predicted by the RFs in the same compo­
In our previous analysis over the NATT area, we evaluated the po­
sition period as the MODIS ET product. The associated errors of the RF
tential to use remote sensing data using a variational data assimilation
estimates were obtained by propagating the standard deviations of the
model and a Penmann-Monteith approach. The approach presented here
daily predictions of the 10 RFs. In Table 5 the error metrics for the
has some key advantages in comparison to our previous methodology for
comparison of the 8-day averaged MODIS ET product and the 8-days
the NATT area (Barraza et al., 2017, 2019): (1) the simplicity of the
averaged predictions of the RF models with respect to the mean
algorithm; (2) the freely available satellite observations; and (3) the
measured ET values (in the same composition periods) are shown. As
high temporal resolution of the available data (daily). The key hypoth­
expected, in the three locations the RF predictions present lower errors
esis is that both satellite products (LAI and LST) provide complementary
than the MODIS ET product in comparison with the measured values.
information, including the thermodynamic characteristics of the surface
The MODIS ET product presents a clear bias with respect to the
and chlorophyll concentration, that improve our ability to determine ET.
measured values in the three sites, that is also seen in Fig. 9. The largest
Random Forest (RF) is very useful for handling high-dimensional data
bias is observed in AU-How, where the ET product is ∼ 1.8 mm/day
and to treat non-linear problems. Given that each tree is built inde­
lower than the real values. In this site, the RMSE of the RF predictions
pendently, the RF method is quite robust to avoid overfitting and to
using all the variables is less than 20% of that of the MODIS ET product
outliers. The fact that the method does not depend on hidden weights or
in comparison with the real values, which corresponds to a rRMSE that is
require regularization as deep learning models, makes it more inter­
∼ 5 times lower. Even the predictions of the MLAI models present an
pretable and user-friendly than many other machine learning methods.
RMSE that is less than 30% of the RMSE of the MODIS product, and a
Other models as multi-layer perceptrons that implied higher computa­
rRMSE that is more than 3.5 times lower. The RF predictions present a
tion times were tested (result not shown) obtaining similar
better agreement with the measured ET values than the MODIS ET

7
V. Douna et al. Journal of Arid Environments 191 (2021) 104513

Fig. 5. Mean modelled (from all input vari­


ables) and measured ET as a function of the
month of the year in each zone (upper: AU -
How, center: AU - DaS, lower: AU - Dry). The
points and error bars correspond to the
means and standard deviations of the values
of ET. Inset: Mean monthly relative error in
ET (difference between measured and pre­
dicted ET, divided by the measured value). In
both panels, the straight line corresponds to
the development set and the slashed lines
correspond to the heldout set (2014).

8
V. Douna et al. Journal of Arid Environments 191 (2021) 104513

Table 4 performance. The variability of Transpiration (T) to ET can result from


Comparison of the mean error metrics and standard deviations of the models biotic-abiotic responses to weather and soil moisture dynamics at daily
that use all the variables (Mall ), LAI and LST (MLAI+LST ), and only LAI (MLAI ) for to weekly scales. However, at the seasonal timescale, such variability
AU-How, AU-DaS and AU-Dry, over the heldout dataset in comparison with the becomes small, and the ratio (T/ET) converges toward a stable mean
measured ET values. state, which is set by LAI (Berkelhammer et al., 2016).
RMSE (mm/ MAE (mm/ Bias (mm/ rRMSE In this framework, the models trained using all the variables as input
day) day) day) (Mall ) reproduce the general trends of the (daily) evapotranspiration for
AU- 0.99±0.06 0.71±0.05 − 0.02±0.06 0.28±0.02 the three EC sites. The second best performance was obtained by the
How MLAI+LST model. As expected, MLAI is considerably less accurate. How­
Mall AU- 1.03±0.06 0.77±0.05 0.45±0.08 0.31±0.02
ever, the 8-days averaged MLAI model (see Table 5) is associated with
DaS
AU- 1.08±0.04 0.72±0.04 0.09±0.08 0.32±0.01 RMSE values lower than 0.5 mm/day, which is of the order of those
Dry reported previously by other models in the NATT area (Barraza et al.,
AU- 1.24±0.04 0.86±0.04 − 0.16±0.06 0.35±0.01 2017, 2019). The accuracy of the predictions is in all cases reduced for
How extreme evapotranspiration values, as the predictions are overestimated
MLAI+LST AU- 1.14±0.03 0.87±0.03 0.35±0.05 0.35±0.01
in the low ET limit and underestimated in the high ET limit. This is
DaS
AU- 1.22±0.06 0.86±0.04 − 0.15±0.04 0.36±0.02 probably related to the low number of samples in the training set for
Dry both the low and high ET limits. This behaviour is very typical in most
AU- 1.40±0.06 1.03±0.05 − 0.23±0.05 0.40±0.02 ML approaches and has been previously reported in the literature (e.g.
How
Dou and Yang, 2018).
MLAI AU- 1.24±0.04 0.95±0.04 0.51±0.06 0.38±0.01
DaS The predicted ET using these three models captures the seasonal
AU- 1.24±0.03 0.90±0.03 − 0.13±0.03 0.37±0.01 variation very well in the transient periods and during stable state of the
Dry growing season. The temporal evolution of the predicted ET values
follows the trend of the observed values. Moreover, in AU-How, the
estimations of MLAI+LST (and to a lesser extent, MLAI ) show a linear
performances. Thus, with the aim of testing an operational and inter­
relation with the measured values. At these EC sites, ET is driven pri­
pretable model, we focused in this manuscript on RF.
marily by surface meteorology during the rainy season and then there is
Our results suggest that LAI can be considered as the first-order
a strong control exerted by the vegetation on transpiration during the
factor affecting ET. We found a strong correlation between LAI and ET
dry season (Barraza et al., 2019). In comparison to previous studies (in
for all the study areas and the permutation importance shows the high
many of which the models show a mismatch during the dry period, e.g.
relevance of LAI for the ET models (which does not necessarily imply a
Yebra et al., (2013); Barraza et al., (2017; 2019)), there is no period
causal relationship). Transpiration is directly related to canopy
when the mean values are out of the range with respect to the in-situ
conductance, which is positively correlated to LAI; therefore, it is not
values. This implies that the RF models could represent the seasonality
surprising that LAI can act as a control over ET partitioning (Transpi­
of the vegetation state.
ration and Evaporation) (Good et al., 2014; Schlesinger and Jasechko,
Annual rainfall has been reported to be above average in Northern
2014; Wang et al., 2014). However, using only the feature with the
Australia during 2014 (Bureau of Meteorology, 2015). For instance, in
highest ranking importance (LAI) was not enough to achieve the highest
2012–2013, annual rainfall at Howard Springs was 1288 mm, whereas

Fig. 6. Modelled vs measured ET in each zone using only LST and LAI as predictor variables (left: AU - How, center: AU - DaS, right: AU - Dry) and best linear fits. The
points and error bars correspond to the means and standard deviations of the values of ET predicted by the MLAI+LST model. The dashed black line corresponds to the
identity. Upper: Training and validation datasets (2010–2013). Lower: Heldout set (2014).

9
V. Douna et al. Journal of Arid Environments 191 (2021) 104513

Fig. 7. Mean modelled (from LST and LAI)


and measured ET as a function of the month
of the year in each zone (upper: AU - How,
center: AU - DaS, lower: AU - Dry). The
points and error bars correspond to the
means and standard deviations of the values
of ET. Inset: Monthly relative error of
monthly ET (difference between measured
and predicted ET, divided by the measured
value). In both panels, the straight line cor­
responds to the development set and the
slashed lines correspond to the heldout set
(2014).

10
V. Douna et al. Journal of Arid Environments 191 (2021) 104513

Fig. 8. Left: Mean monthly predicted ET for AU-How, AU-Ade, AU-DaS and AU-Dry. The error bars in the predicted values were calculated from error propagation of
the standard deviations of the daily ET values predicted by the MLAI+LST model. Right: Mean monthly LAI and LST for the 4 sites for comparison.

Fig. 9. Left: Modelled (with MLAI+LST ) vs measured ET (8-days average) for the complete dataset in AU-How (upper panel), AU-DaS (middle panel) and AU-Dry
(lower panel), and 8-days-averaged ET from the MODIS ET global product (green diamonds). Right: Modelled vs measured ET (8 days average) and 8-days
average ET from the MODIS ET product as a function of time. The error bars in the predicted values were calculated by error propagation of the standard de­
viations of the daily ET values predicted by the models. (For interpretation of the references to colour in this figure legend, the reader is referred to the Web version of
this article.)

in 2013–2014 it was 1948 mm. The higher annual rainfall in approach by which 2014 is chosen as the heldout set while the devel­
2013–2014, increases deep soil moisture storage, which was also higher opment set is 2010–2013, the reported excess could contribute to
for longer during the dry season of 2013–2014 than that for 2012–2013 degrade the performance of the ET predictions considerably. This
(Moore et al., 2016). Considering that we are exploring a temporal mismatch is particularly enhanced in those models that do not use

11
V. Douna et al. Journal of Arid Environments 191 (2021) 104513

Table 5 consistent measurements, lower errors are expected to be found


Mean error metrics of the models over the whole dataset and of the MODIS ET compared to other ecosystems and locations non-represented at the EC
product in comparison with the measured ET values. network. However, further evaluation would be needed to upscale the
RMSE MAE Bias (mm/ rRMSE results of our RF predictions to the regional scale and to estimate the
(mm/day) (mm/day) day) errors introduced by spatial heterogeneities at pixel scale. An alternative
AU- 2.04 1.83 − 1.80 0.64 validation technique as the triple collocation method (Stoffelen, 1998)
How which has been previously used to estimate uncertainties in ET esti­
MODIS global ET AU- 1.22 0.98 − 0.77 0.53 mations at locations with missing ground observations (e.g. Barraza
product DaS
et al., 2018), could be used with this aim in the future.
AU- 1.28 0.94 − 0.88 0.50
Dry In this context, it is relevant to mention that scale inconsistencies
AU- 0.38 0.24 0.02 0.12 may have certain effects on the evaluation and comparison of the sat­
How ellite and surface data: (i) eddy covariance fluxes and meteorological
Mall AU- 0.29 0.18 − 0.04 0.13 sensors (radiation, temperature, humidity) have different footprints,
DaS
AU- 0.37 0.21 − 0.07 0.14
and (ii) the tower flux footprint is only a small fraction (< 1 km) of the
Dry footprint of optical (MODIS; 1 km) and LST (5 km). In particular, in our
AU- 0.50 0.30 0.05 0.16 work there might be non-negligible uncertainties associated with the
How size of MODIS and LST pixels and the tower flux footprint. To address the
MLAI+LST AU- 0.42 0.27 − 0.03 0.18
eddy covariance-remote sensing footprint issues, we evaluated the Land
DaS
AU- 0.57 0.36 − 0.05 0.22 Cover Variability at different windows sizes (1, 3 and 5 km). The
Dry dominant land cover class has a predominance between 90 − 92%, 80 −
AU- 0.56 0.36 0.08 0.18 87% and less than 70 − 60% of the total area in a 1 km, 3 km and 5 km
How window respectively around the towers. Although some variances
MLAI AU- 0.44 0.30 − 0.03 0.19
shown in this study may be explained by scale-inconsistencies and
DaS
AU- 0.65 0.43 − 0.03 0.25 observation uncertainties, our RF methodology was able to capture
Dry subtle variations in ET, with RMSEs commensurate with those reported
previously (Barraza et al., 2017, 2019; Yebra et al., 2013). With respect
to the temporal limitation associated with the availability of high spatial
meteorological variables as input (MLAI+LST and MLAI ). Also the range of
resolution LST data, the new generation of sensors GOES-R and Hama­
LAI values during 2014 in AU-Dry is larger than in previous years and
wani 9 will provide data every 15 min with a spatial resolution up to 2
thus the predictions of this supervised machine learning scheme are
km, improving the access to high-resolution thermal data. In this
inevitably affected. Although this represents one of the inner limitations
context, our models are useful in terms of obtaining an operational daily
of the selected temporal approach, we are particularly interested in
local ET estimation scheme, using multiple LST observations in varying
exploring the possibility of developing models that may be applied to
time windows. Our results also enhance the possibility of obtaining ET
future temporal extensions of a given dataset, as an operational
estimates in different temporal scales in light of the upcoming facilities.
approach. It should be pointed out that a model that uses all the input
variables (as Mall ) may not be operative if in-situ measurements are not
4.1. Comparison with previous ET models
available, and thus within our approach it only constitutes an upper
limit of expected performance. An alternative could be to replace in-situ
By means of a simple data-driven approach, we have achieved in the
meteorological data with Global Meteorological data (GMD), but in our
heldout set mean RMSE daily values of ∼ 1 mm/day using all the vari­
previous studies we found a large disagreement between them (Barraza
ables and between 1.1 − 1.2 mm/day for the models that use only LAI
et al., 2019).
and LST. In particular, for AU-How, the relative RMSE over the heldout
We have explored the potential of extending the predictions of the
set (RMSE divided by the mean daily ET, which corresponds to 3.13
models that use only remote sensing data to other sites where there are
mm/day) is around 0.3–0.35 for the RF regressors that use all the var­
no EC towers, by using the MLAI+LST models trained for AU-How and AU-
iables and both LAI and LST. Our modelled ET and estimated errors were
DaS in AU-Ade. In this location, in-situ ET data is not available in the
comparable to (or even lower than) those reported from more complex
period 2010–2014 and thus it was not possible to make a robust quan­
terrestrial observation models (Dou and Yang (2018), Yang et al.,
titative comparison. However, the monthly estimates of ET follow the
(2006), Leuning et al., (2008), Barraza et al., (2017), Yebra et al.,
expected seasonal behaviour and as a first approximation, they repro­
(2013), Dou and Yang (2018), Yang et al., (2006), Leuning et al., (2008),
duce the expected rainfall gradient across the region (Hutley et al., 2011;
Barraza et al., (2017), Yebra et al., (2013)). However, the comparison
Barraza et al., 2017), following LAI and LST. Thus, as a first approach,
with other models and other ET estimations in the literature is not
the application of the ET estimation scheme that we are presenting could
straightforward, as the procedures and sites are different as well as the
be useful even in other areas, as LAI and LST can be assessed remotely at
input data.
large scales without the need of measuring the fluxes in-situ.
In particular, these errors are comparable to the relative RMSEs
The main goal of using remote sensing products to assess ET is to
obtained for the deciduous broadleaf forest research site in Hainich
obtain estimates over areas where in-situ measurements are not avail­
National Park in Germany (DE-Hai) using more complex machine
able. Several methods for upscaling EC measurements to the regional
learning algorithms by Dou and Yang (2018), trained and tested using
scale have been developed (e.g. Li et al., 2018, and references therein],
air temperature, net radiation, relative humidity and soil temperature as
as ground-based observations are considered to be representative only in
input variables. Although the RMSEs and MAEs they obtain for DE-Hai
the flux-tower-footprint scale (∼ 100m to ∼ 1km, e.g. (Liu et al., 2016)).
over their validation and testing datasets (RMSE∼ 0.20–0.35 mm/day,
Machine learning models, including RFs that use land cover data,
MAE∼ 0.15–0.25 mm/day) are lower than ours, the range in the ET
meteorological variables and solar radiation as explanatory variables,
daily measured values in DE-Hai and its mean value is considerably
have been evaluated with this aim (Xu et al., 2018). Land surface in­
smaller than in AU-How, which results in similar, or even higher relative
homogeneities at the sub-pixel scale, may introduce discrepancies be­
RMSEs (rRMSE ∼ 0.25–0.45) than ours.
tween the in-situ measurements and the ET estimations at the satellite
The mean errors that we obtain are also similar to the ones informed
pixel scale (e.g. Ke et al., 2016; Hu et al., 2015). As the RF models were
as a result of applying other physically motivated models based on
trained over a widely validated and relatively homogeneous region with
surface energy balance schemes. For instance, Mu et al., (2011) provide

12
V. Douna et al. Journal of Arid Environments 191 (2021) 104513

mean RMSE values for daily ET determinations of 0.84 mm/day across error in the evapotranspiration predictions that results from our RF
46 eddy flux towers (Ameriflux) and 0.90 mm/day using GMAO mete­ approach may be useful locally if historical data is available, in terms of
orology. In Freeman Ranch Mesquite Juniper tower which is the woody obtaining a more precise operational ET estimation from local
savanna with the highest mean ET observed (2.08 mm/day) the RMSEs ground-measured or remotely sensed data.
are 0.91 and 0.85 mm/day respectively, which correspond to rRMSEs of
0.44 and 0.41. Despite our RMSEs and MAEs for AU-How on the heldout 5. Conclusions
set are higher than theirs, our relative RMSEs are lower for both RF
regressors trained with all in-situ variables and only LAI data. In this work, we have evaluated the capability of a random forest
On the other hand, Leuning et al., (2008) have developed a simple regressor to predict daily evapotranspiration in three EC sites in
model to predict the daily average evapotranspiration using remotely Australia (AU-How, AU-DaS, AU-Dry), using daily satellite LAI and LST
sensed LAI data and the PM equation. In Virginia Park, classified as a products and mean daily in-situ meteorological data. Despite the
woody savanna as AU-How, they obtain a relative root-mean-square fundamental differences among the existing models (e.g. semi-
error in their estimate of the average evaporation of approximately empirical, variational, machine learning), our RF (daily) models pre­
64%, over a mean ET value of 1.2 mm/day. As expected, this error is sent comparable (and even better) performances than other existing
larger than the ones we obtain in the test sets. global products at the selected locations. In particular, the ET pre­
It has been shown that combining multiple algorithms to predict ET dictions of our RF models (using all the input variables, only LAI and
or LE may improve the accuracy in the predictions. Yao et al., (2014) LST, and only LAI) show better agreement with the measured values
have found a decrease in the RMSE of the LE determination of around than the MODIS ET global product used as benchmark for the three sites,
6W/m2 for forest and savanna sites by applying a Bayesian model where the latter shows a clear bias with respect to the measured values.
averaging (BMA) method that merges five different algorithms driven by Although this product is intended to work globally while our approach is
daily tower-specific meteorology. In particular, a better agreement of local, the fact that our estimations have similar or even better perfor­
their method with the observations than the MODIS LE product is shown mances in each site suggests the utility of training a RF locally and then
for AU-How, especially during the dry months. A similar behaviour using it to predict ET from satellite data in an operational and simple
would be expected when applying a combined method to determine ET. way. This approach could be useful to estimate ET when in-situ data is
However, as we are using the MODIS ET product only as a benchmark not available, in the locations where a thorough validation of the global
for performance evaluation, a straightforward comparison of ET prod­ products has not been performed or even to complement other
ucts in this location is beyond the scope of our work. physically-based models. We have also started exploring the possibility
A bias of ∼ 1.8 between the MODIS ET product and the EC tower of spatially extending the predictions of the model that uses LAI and LST
measured values in AU-How is clearly visible (see Fig. 9). A smaller bias to another site in the NATT. The spatial ET pattern was generally
of ∼ 0.8 − 0.9 mm/day is shown for AU-DaS and AU-Dry. The same reasonable and the model was able to capture significant seasonal var­
biases are observed when comparing the MODIS product with our RF iations. However, the capability of such a RF scheme to map ET varia­
predictions. This bias with respect to the in-situ measurements has not tions will be further explored and evaluated. The final goal of such
been reported before 2010 (see Barraza et al., 2018). The RMSE, MAE exploration would be to develop an operative methodology to predict
and rRMSE errors of the product when compared to the averaged daily ET, precise in the local scale, that relies only on remote sensing
measured values are similar to the ones we obtain from the models data, since such a product could be of use in other regions of the globe
trained using only LAI daily data. This suggests that even a RF model where long-term in-situ data may not be available.
that uses only LAI satellite data, is able to predict ET values which are Overall, these results show that this approach presents several ad­
more accurate than those obtained from the MODIS ET product in the vantages to complement the previous ones applied over the NATT area
three EC sites. This is not surprising as the MODIS product that we are (Yebra et al., 2013; Barraza et al., 2017, 2019). We conclude that in
using as a benchmark for comparison is intended to provide global es­ particular the RF LAI + LST model can be used to obtain reasonable daily
timates, while our scheme is trained locally. But this result indicates that estimations of ET and to provide complementary information to other
by means of this kind of models, LAI and LST could be certainly useful as global products (e.g. MODIS and BESS) in terms of flux dynamics.
proxies of ET in sites where meteorological or in-situ data is not
available. CRediT authorship contribution statement
Finally, we have compared the predictions of our RF methods in a
reduced daily timescale (10–17hs) to the results of the VDA combined- V. Douna: Writing – original draft, preparation, Conceptualization,
source scheme in Barraza et al., (2019) for AU-How (see Appendix 2). Methodology, programming. V. Barraza: Conceptualization, Method­
Although the RMSE and MAEs of these RF models are slightly higher ology, Writing – original draft. F. Grings: Conceptualization, Writing-
than those obtained without limiting the daily window, the relative Reviewing and Editing, Supervision. A. Huete: Supervision. N.
RMSEs are lower due to the fact that the mean ET values are higher. As Restrepo-Coupe: eddy covariance data curation. J. Beringer: PI of the
expected, the VDA results present a big dispersion and less sensitivity to eddy covariance tower and Supervision.
the ET real changes than the RF predictions. Despite the different nature
of both models, it should be pointed out that the accuracy of the ET Declaration of competing interest
predictions from MLAI and Mall RF models is greater than that of the VDA
model, using similar input data. This behaviour can be appreciated in The authors declare that they have no known competing financial
Fig. S6. The VDA predicted ET values present more outliers during all interests or personal relationships that could have appeared to influence
the selected period than those obtained from the RF and in particular, its the work reported in this paper.
predictions are much lower and flatter than the observed values during
2012. The limitations of the VDA model to predict EF during this period Acknowledgments
have been previously discussed in Barraza et al., (2019). This
disagreement is also reflected in the poorer error metrics resulting from This work was funded by PICTO-2014-0099 and PICT 2018-1819 (PI:
the comparison between the VDA predictions and the RF ones. It should Barraza, Veronica). The authors thank Ozflux Network for making the
be pointed out that the comparison of our results with those obtained data freely available as well as the flux towers principal investigators.
from physically-based models is not straightforward. Although the latter
are intended to work globally while our approach is local, the lower

13
V. Douna et al. Journal of Arid Environments 191 (2021) 104513

Appendix A. Supplementary data 10.1016/j.rse.2012.07.006. ISSN 0034-4257. http://www.sciencedirect.com/scienc


e/article/pii/S0034425712002787.
Glenn, E.P., Doody, T.M., Guerschman, J.P., Huete, A.R., King, E.A., McVicar, T.R., Van
Supplementary data to this article can be found online at https://doi. Dijk, A.I., Van Niel, T.G., Yebra, M., Zhang, Y., 2011. Actual evapotranspiration
org/10.1016/j.jaridenv.2021.104513. estimation by ground and remote sensing methods: the australian experience.
Hydrol. Process. 25 (26), 4103–4116.
Glenn, E.P., Huete, A.R., Nagler, P.L., Hirschboeck, K.K., Brown, P., 2007. Integrating
References remote sensing and ground methods to estimate evapotranspiration. Crit. Rev. Plant
Sci. 26 (3), 139–168.
Abdullah, S.S., Malek, M., Abdullah, N.S., Kisi, O., Yap, K.S., 2015. Extreme learning Gocić, M., Motamedi, S., Shamshirband, S., Petković, D., Ch, S., Hashim, R., Arif, M.,
machines: a new approach for prediction of reference evapotranspiration. J. Hydrol. 2015. Soft computing approaches for forecasting reference evapotranspiration.
527, 184–195. Comput. Electron. Agric. 113, 164–173. https://doi.org/10.1016/j.
Baldocchi, D., Falge, E., Gu, L., Olson, R., Hollinger, D., Running, S., Anthoni, P., compag.2015.02.010. ISSN 0168-1699. http://www.sciencedirect.com/science/
Bernhofer, C., Davis, K., Evans, R., et al., 2001. Fluxnet: a new tool to study the article/pii/S0168169915000526.
temporal and spatial variability of ecosystem-scale carbon dioxide, water vapor, and Good, S.P., Soderberg, K., Guan, K., King, E.G., Scanlon, T.M., Caylor, K.K., 2014. δ2h
energy flux densities. Bull. Am. Meteorol. Soc. 82 (11), 2415–2434. isotopic flux partitioning of evapotranspiration over a grass field following a water
Barraza, V., Grings, F., Franco, M., Douna, V., Entekhabi, D., Restrepo-Coupe, N., pulse and subsequent dry down. Water Resour. Res. 50 (2), 1410–1432.
Huete, A., Gassmann, M., Roitberg, E., 2019. Estimation of latent heat flux using Guerschman, J.P., Van Dijk, A.I., Mattersdorf, G., Beringer, J., Hutley, L.B., Leuning, R.,
satellite land surface temperature and a variational data assimilation scheme over a Pipunic, R.C., Sherman, B.S., 2009. Scaling of potential evapotranspiration with
eucalypt forest savanna in northern Australia. Agric. For. Meteorol. 268, 341–353. modis data reproduces flux observations and catchment water balance observations
https://doi.org/10.1016/j.agrformet.2019.01.032. ISSN 0168-1923. across Australia. J. Hydrol. 369 (1–2), 107–119.
Barraza, V., Grings, F., Restrepo-Coupe, N., Huete, A., 2018. Comparison of the Halladay, K., Good, P., 2017. Non-linear interactions between co2 radiative and
performance of latent heat flux products over southern hemisphere forest physiological effects on amazonian evapotranspiration in an earth system model.
ecosystems: estimating latent heat flux error structure using in situ measurements Clim. Dynam. 49 (7–8), 2471–2490.
and the triple collocation method. Int. J. Rem. Sens. 39 (19), 6300–6315. https:// Hoek van Dijke, A.J., Mallick, K., Teuling, A.J., Schlerf, M., Machwitz, M., Hassler, S.K.,
doi.org/10.1080/01431161.2018.1458348. Blume, T., Herold, M., 2019. Does the normalized difference vegetation index
Barraza, V., Restrepo-Coupe, N., Huete, A., Grings, F., Beringer, J., Cleverly, J., explain spatial and temporal variability in sap velocity in temperate forest
Eamus, D., 2017. Estimation of latent heat flux over savannah vegetation across the ecosystems? Hydrol. Earth Syst. Sci. 23 (4), 2077–2091. https://doi.org/10.5194/
north australian tropical transect from multiple sensors and global meteorological hess-23-2077-2019. https://www.hydrol-earth-syst-sci.net/23/2077/2019/.
data. Agric. For. Meteorol. 232, 689–703. Hu, G., Jia, L., Menenti, M., 2015. Comparison of mod16 and lsa-saf msg
Beringer, J., Hutley, L.B., McHugh, I., Arndt, S.K., Campbell, D., Cleugh, H.A., evapotranspiration products over europe for 2011. Rem. Sens. Environ. 156,
Cleverly, J., De Dios, V.R., Eamus, D., Evans, B., et al., 2016. An Introduction to the 510–526.
Australian and new zealand Flux Tower Network-Ozflux. Biogeosciences. Huntington, T.G., 2006. Evidence for intensification of the global water cycle: review and
Berkelhammer, M., Noone, D., Wong, T., Burns, S., Knowles, J., Kaushik, A., Blanken, P., synthesis. J. Hydrol. 319 (1–4), 83–95.
Williams, M., 2016. Convergent approaches to determine an ecosystem’s Huo, Z., Feng, S., Kang, S., Dai, X., 2012. Artificial neural network models for reference
transpiration fraction. Global Biogeochem. Cycles 30 (6), 933–951. evapotranspiration in an arid area of northwest China. J. Arid Environ. 82, 81–90.
Bhattarai, N., Mallick, K., Stuart, J., Vishwakarma, B.D., Niraula, R., Sen, S., Jain, M., https://doi.org/10.1016/j.jaridenv.2012.01.016. ISSN 0140-1963. http://www.sci
2019. An automated multi-model evapotranspiration mapping framework using encedirect.com/science/article/pii/S0140196312000481.
remotely sensed and reanalysis data. Rem. Sens. Environ. 229, 69–92. https://doi. Hutley, L.B., Beringer, J., Isaac, P.R., Hacker, J.M., Cernusak, L.A., 2011. A sub-
org/10.1016/j.rse.2019.04.026. ISSN 0034-4257. continental scale living laboratory: spatial patterns of savanna vegetation over a
Breiman, L., Oct 2001. Random forests. Mach. Learn. 45 (1), 5–32. https://doi.org/ rainfall gradient in northern Australia. Agric. For. Meteorol. 151 (11), 1417–1428.
10.1023/A:1010933404324. ISSN 1573-0565. Isaac, P., Cleverly, J., McHugh, I., Gorsel, E.v., Ewenz, C., Beringer, J., 2017. Ozflux data:
Bureau of Meteorology, A.G., 2015. Annual climate statement. http://www.bom.gov.au/ network integration from collection to curation. Biogeosciences 14 (12), 2903–2928.
climate/current/annual/aus/2014/. Jain, S., Nayak, P., Sudheer, K., 2008. Models for estimating evapotranspiration using
Cai, Y., Guan, K., Lobell, D., Potgieter, A.B., Wang, S., Peng, J., Xu, T., Asseng, S., artificial neural networks, and their physical interpretation. Hydrol. Process.: Int. J.
Zhang, Y., You, L., Peng, B., 2019. Integrating satellite and climate data to predict 22 (13), 2225–2234.
wheat yield in Australia using machine learning approaches. Agric. For. Meteorol. Jouybari-Moghaddam, Y., Saradjian, M.R., 2019. A semi-empirical approach for the
274, 144–159. https://doi.org/10.1016/j.agrformet.2019.03.010. ISSN 0168-1923. estimation of land-surface emissivity from satellite data based on spectral index
http://www.sciencedirect.com/science/article/pii/S0168192319301224. fusion using ensemble regression. Int. J. Rem. Sens. 40 (11), 4213–4243. https://doi.
Chen, Y., Xia, J., Liang, S., Feng, J., Fisher, J.B., Li, X., Li, X., Liu, S., Ma, Z., Miyata, A., org/10.1080/01431161.2018.1562261.
et al., 2014. Comparison of satellite-based evapotranspiration models over terrestrial Jung, M., Reichstein, M., Ciais, P., Seneviratne, S.I., Sheffield, J., Goulden, M.L.,
ecosystems in China. Rem. Sens. Environ. 140, 279–293. Bonan, G., Cescatti, A., Chen, J., De Jeu, R., et al., 2010. Recent decline in the global
Cleugh, H.A., Leuning, R., Mu, Q., Running, S.W., 2007. Regional evaporation estimates land evapotranspiration trend due to limited moisture supply. Nature 467 (7318),
from flux tower and modis satellite data. Rem. Sens. Environ. 106 (3), 285–304. 951.
Crisci, C., Ghattas, B., Perera, G., 2012. A review of supervised machine learning Kalma, J.D., McVicar, T.R., McCabe, M.F., 2008. Estimating land surface evaporation: a
algorithms and their applications to ecological data. Ecol. Model. 240, 113–122. review of methods using remotely sensed surface temperature data. Surv. Geophys.
https://doi.org/10.1016/j.ecolmodel.2012.03.001. ISSN 0304-3800. 29 (4–5), 421–469.
Dirmeyer, P.A., Gao, X., Zhao, M., Guo, Z., Oki, T., Hanasaki, N., 2006. Gswp-2: Kanniah, K.D., Beringer, J., Hutley, L.B., 2011. Environmental controls on the spatial
multimodel analysis and implications for our perception of the land surface. Bull. variability of savanna productivity in the northern territory, Australia. Agric. For.
Am. Meteorol. Soc. 87 (10), 1381–1398. https://doi.org/10.1175/BAMS-87-10- Meteorol. 151 (11), 1429–1439.
1381. Ke, Y., Im, J., Park, S., Gong, H., 2016. Downscaling of modis one kilometer
Dos Santos, V., Laurent, F., Abe, C., Messner, F., 2018. Hydrologic response to land use evapotranspiration using landsat-8 data and machine learning approaches. Rem.
change in a large basin in eastern amazon. Water 10 (4). https://doi.org/10.3390/ Sens. 8 (3), 215.
w10040429. ISSN 2073-4441. https://www.mdpi.com/2073-4441/10/4/429. Koch, G.W., Vitousek, P.M., Steffen, W.L., Walker, B.H., 1995. Terrestrial transects for
Dou, X., Yang, Y., 2018. Evapotranspiration estimation using four different machine global change research. Vegetatio 121 (1–2), 53–65.
learning approaches in different terrestrial ecosystems. Comput. Electron. Agric. Kousari, M.R., Hosseini, M.E., Ahani, H., Hakimelahi, H., 2017. Introducing an
148, 95–106. https://doi.org/10.1016/j.compag.2018.03.010. ISSN 0168-1699. operational method to forecast long-term regional drought based on the application
http://www.sciencedirect.com/science/article/pii/S016816991731476X. of artificial intelligence capabilities. Theor. Appl. Climatol. 127 (1–2), 361–380.
Eamus, D., Cleverly, J., Boulain, N., Grant, N., Faux, R., Villalobos-Vega, R., 2013. Kumar, M., Raghuwanshi, N.S., Singh, R., Jan 2011. Artificial neural networks approach
Carbon and water fluxes in an arid-zone acacia savanna woodland: an analyses of in evapotranspiration modeling: a review. Irrigat. Sci. 29 (1), 11–25. https://doi.
seasonal patterns and responses to rainfall events. Agric. For. Meteorol. 182, org/10.1007/s00271-010-0230-8, 1432-1319.
225–238. Lary, D.J., Zewdie, G.K., Liu, X., Wu, D., Levetin, E., Allee, R.J., Malakar, N., Walker, A.,
Eamus, D., O’Grady, A., Hutley, L., 2000. Dry season conditions determine wet season Mussa, H., Mannino, A., Aurin, D., 2018. Machine Learning Applications for Earth
water use in the wet–tropical savannas of northern Australia. Tree Physiol. 20 (18), Observation. Springer International Publishing, Cham, ISBN 978-3-319-65633-5,
1219–1226. pp. 165–218. https://doi.org/10.1007/978-3-319-65633-5˙8.
Ershadi, A., McCabe, M., Evans, J., Wood, E.F., 2015. Impact of model structure and Leuning, R., Zhang, Y.Q., Rajaud, A., Cleugh, H., Tu, K., 2008. A simple surface
parameterization on penman–monteith type evaporation models. J. Hydrol. 525, conductance model to estimate regional evaporation using modis leaf area index and
521–535. the penman-monteith equation. Water Resour. Res. 44 (10) https://doi.org/
Fisher, J.B., Melton, F., Middleton, E., Hain, C., Anderson, M., Allen, R., McCabe, M.F., 10.1029/2007WR006562. https://agupubs.onlinelibrary.wiley.com/doi/abs/10.10
Hook, S., Baldocchi, D., Townsend, P.A., et al., 2017. The future of 29/2007WR006562.
evapotranspiration: global requirements for ecosystem functioning, carbon and Li, X., Liu, S., Li, H., Ma, Y., Wang, J., Zhang, Y., Xu, Z., Xu, T., Song, L., Yang, X., et al.,
climate feedbacks, agricultural management, and water resources. Water Resour. 2018. Intercomparison of six upscaling evapotranspiration methods: from site to the
Res. 53 (4), 2618–2626. satellite pixel. J. Geophys. Res.: Atmos. 123 (13), 6777–6803.
Gleason, C.J., Im, J., 2012. Forest biomass estimation from airborne lidar data using Liu, M., Liu, X., Liu, D., Ding, C., Jiang, J., 2015. Multivariable integration method for
machine learning approaches. Rem. Sens. Environ. 125, 80–91. https://doi.org/ estimating sea surface salinity in coastal waters from in situ data and remotely
sensed data using random forest algorithm. Comput. Geosci. 75, 44–56. https://doi.

14
V. Douna et al. Journal of Arid Environments 191 (2021) 104513

org/10.1016/j.cageo.2014.10.016. ISSN 0098-3004. http://www.sciencedirect.co Shrestha, N., Shukla, S., 2015. Support vector machine based modeling of
m/science/article/pii/S0098300414002490. evapotranspiration using hydro-climatic variables in a sub-tropical environment.
Liu, S., Xu, Z., Song, L., Zhao, Q., Ge, Y., Xu, T., Ma, Y., Zhu, Z., Jia, Z., Zhang, F., 2016. Agric. For. Meteorol. 200, 172–184. https://doi.org/10.1016/j.
Upscaling evapotranspiration measurements from multi-site to the satellite pixel agrformet.2014.09.025. ISSN 0168-1923. http://www.sciencedirect.com/science/
scale over heterogeneous land surfaces. Agric. For. Meteorol. 230, 97–113. article/pii/S0168192314002494.
Majozi, N.P., Mannaerts, C.M., Ramoelo, A., Mathieu, R., Mudau, A.E., Verhoef, W., Stoffelen, A., 1998. Toward the true near-surface wind speed: error modeling and
2017. An intercomparison of satellite-based daily evapotranspiration estimates calibration using triple collocation. J. Geophys. Res.: Oceans 103 (C4), 7755–7766.
under different eco-climatic regions in South Africa. Rem. Sens. 9 (4) https://doi. Suppiah, R., 1992. The australian summer monsoon: a review. Prog. Phys. Geogr. 16 (3),
org/10.3390/rs9040307. ISSN 2072-4292. https://www.mdpi.com/2072-4292/9 283–318.
/4/307. Tabari, H., Martinez, C., Ezani, A., Talaee, P.H., 2013. Applicability of support vector
Mallick, K., Wandera, L., Bhattarai, N., Hostache, R., Kleniewska, M., Chormanski, J., machines and adaptive neurofuzzy inference system for modeling potato crop
2018. A critical evaluation on the role of aerodynamic and canopy–surface evapotranspiration. Irrigat. Sci. 31 (4), 575–588.
conductance parameterization in seb and svat models for simulating Teuling, A.J., Hirschi, M., Ohmura, A., Wild, M., Reichstein, M., Ciais, P., Buchmann, N.,
evapotranspiration: a case study in the upper biebrza national park wetland in Ammann, C., Montagnani, L., Richardson, A.D., Wohlfahrt, G., Seneviratne, S.I.,
Poland. Water 10 (12), 1753. 2009. A regional perspective on trends in continental evaporation. Geophys. Res.
Mehdizadeh, S., Behmanesh, J., Khalili, K., 2017. Using mars, svm, gep and empirical Lett. 36 (2) https://doi.org/10.1029/2008GL036584. https://agupubs.onlinelibrary
equations for estimation of monthly mean reference evapotranspiration. Comput. .wiley.com/doi/abs/10.1029/2008GL036584.
Electron. Agric. 139, 103–114. https://doi.org/10.1016/j.compag.2017.05.002. Ukkola, A., Prentice, I., et al., 2013. A worldwide analysis of trends in water-balance
ISSN 0168-1699. http://www.sciencedirect.com/science/article/pii/S01681699 evapotranspiration. Hydrol. Earth Syst. Sci. 17 (10), 4177–4187.
16311954. Verstraeten, W., Veroustraete, F., Feyen, J., 2008. Assessment of evapotranspiration and
Moore, C.E., Beringer, J., Evans, B., Hutley, L.B., McHugh, I., Tapper, N.J., 2016. The soil moisture content across different scales of observation. Sensors 8 (1), 70–117.
contribution of trees and grasses to productivity of an australian tropical savanna. Wang, L., Good, S.P., Caylor, K.K., 2014. Global synthesis of vegetation control on
Biogeosciences 13 (8), 2387–2403. https://doi.org/10.5194/bg-13-2387-2016. htt evapotranspiration partitioning. Geophys. Res. Lett. 41 (19), 6753–6757. https://
ps://bg.copernicus.org/articles/13/2387/2016/. doi.org/10.1002/2014GL061439. https://agupubs.onlinelibrary.wiley.com/doi/a
Mu, Q., Heinsch, F.A., Zhao, M., Running, S.W., 2007. Development of a global bs/10.1002/2014GL061439.
evapotranspiration algorithm based on modis and global meteorology data. Rem. Wang, L., Zhou, X., Zhu, X., Dong, Z., Guo, W., 2016. Estimation of biomass in wheat
Sens. Environ. 111 (4), 519–536. https://doi.org/10.1016/j.rse.2007.04.015. ISSN using random forest regression algorithm and remote sensing data. Crop J. 4 (3),
0034-4257. http://www.sciencedirect.com/science/article/pii/S0034425 212–219. https://doi.org/10.1016/j.cj.2016.01.008. ISSN 2214-5141. http://www.
707001903. sciencedirect.com/science/article/pii/S2214514116300162.
Mu, Q., Zhao, M., Running, S.W., 2011. Improvements to a modis global terrestrial Copernicus Website. Global land service. http://land.copernicus.vgt.vito.be/PDF/porta
evapotranspiration algorithm. Rem. Sens. Environ. 115 (8), 1781–1800. https://doi. l/Application.html.
org/10.1016/j.rse.2011.02.019. ISSN 0034-4257. http://www.sciencedirect.com/ Wei, Z., Yoshimura, K., Wang, L., Miralles, D.G., Jasechko, S., Lee, X., 2017. Revisiting
science/article/pii/S0034425711000691. the contribution of transpiration to global terrestrial evapotranspiration. Geophys.
Mutanga, O., Adam, E., Cho, M.A., 2012. High density biomass estimation for wetland Res. Lett. 44 (6), 2792–2801.
vegetation using worldview-2 imagery and random forest regression algorithm. Int. Xu, T., Guo, Z., Liu, S., He, X., Meng, Y., Xu, Z., Xia, Y., Xiao, J., Zhang, Y., Ma, Y., et al.,
J. Appl. Earth Obs. Geoinf. 18, 399–406. 2018. Evaluating different machine learning methods for upscaling
Mystakidis, S., Davin, E.L., Gruber, N., Seneviratne, S.I., 2016. Constraining future evapotranspiration from flux towers to the regional scale. J. Geophys. Res.: Atmos.
terrestrial carbon cycle projections using observation-based water and carbon flux 123 (16), 8674–8690.
estimates. Global Change Biol. 22 (6), 2198–2215. Yang, F., White, M.A., Michaelis, A.R., Ichii, K., Hashimoto, H., Votava, P., Zhu, A.-X.,
Otgonbayar, M., Atzberger, C., Chambers, J., Damdinsuren, A., 2019. Mapping pasture Nemani, R.R., 2006. Prediction of continental-scale evapotranspiration by
biomass in Mongolia using partial least squares, random forest regression and combining modis and ameriflux data through support vector machine. IEEE Trans.
landsat 8 imagery. Int. J. Rem. Sens. 40 (8), 3204–3226. https://doi.org/10.1080/ Geosci. Rem. Sens. 44 (11), 3452–3461.
01431161.2018.1541110. Yang, Y., Cao, C., Pan, X., Li, X., Zhu, X., Jul 2017. Downscaling land surface temperature
Park, S., Im, J., Jang, E., Rhee, J., 2016. Drought assessment and monitoring through in an arid area by using multiple remote sensing indices with random forest
blending of multi-sensor indices using machine learning approaches for different regression. Rem. Sens. 9 (8), 789. https://doi.org/10.3390/rs9080789. ISSN 2072-
climate regions. Agric. For. Meteorol. 216, 157–169. 4292.
Petković, D., Gocic, M., Trajkovic, S., Shamshirband, S., Motamedi, S., Hashim, R., Yao, Y., Liang, S., Li, X., Hong, Y., Fisher, J.B., Zhang, N., Chen, J., Cheng, J., Zhao, S.,
Bonakdari, H., 2015. Determination of the most influential weather parameters on Zhang, X., Jiang, B., Sun, L., Jia, K., Wang, K., Chen, Y., Mu, Q., Feng, F., 2014.
reference evapotranspiration by adaptive neuro-fuzzy methodology. Comput. Bayesian multimodel estimation of global terrestrial latent heat flux from eddy
Electron. Agric. 114, 277–284. https://doi.org/10.1016/j.compag.2015.04.012. covariance, meteorological, and satellite observations. J. Geophys. Res.: Atmos. 119
ISSN 0168-1699. http://www.sciencedirect.com/science/article/pii/S01681699 (8), 4521–4545. https://doi.org/10.1002/2013JD020864. https://agupubs.onlinel
15001192. ibrary.wiley.com/doi/abs/10.1002/2013JD020864.
Raghavendra, S., Deka, P.C., 2014. Support vector machine applications in the field of Yao, Y., Liang, S., Qin, Q., Wang, K., Zhao, S., 2011. Monitoring global land surface
hydrology: a review. Appl. Soft Comput. 19, 372–386. https://doi.org/10.1016/j. drought based on a hybrid evapotranspiration model. Int. J. Appl. Earth Obs. Geoinf.
asoc.2014.02.002. ISSN 1568-4946. http://www.sciencedirect.com/science/article/ 13 (3), 447–457. https://doi.org/10.1016/j.jag.2010.09.009. ISSN 0303-2434.
pii/S1568494614000611. http://www.sciencedirect.com/science/article/pii/S0303243410001133.
Ramoelo, A., Majozi, N., Mathieu, R., Jovanovic, N., Nickless, A., Dzikiti, S., 2014. Yebra, M., Dijk, A.V., Leuning, R., Huete, A., Guerschman, J.P., 2013. Evaluation of
Validation of global evapotranspiration product (mod16) using flux tower data in optical remote sensing to estimate actual evapotranspiration and canopy
the african savanna, South Africa. Rem. Sens. 6, 7406–7423. https://doi.org/ conductance. Rem. Sens. Environ. 129, 250–261. https://doi.org/10.1016/j.
10.3390/rs6087406, 08. rse.2012.11.004. ISSN 0034-4257. http://www.sciencedirect.com/science/article/
Restrepo-Coupe, N., Huete, A., Davies, K., Cleverly, J., Beringer, J., Eamus, D., Van pii/S0034425712004270.
Gorsel, E., Hutley, L., Meyer, W., 2016. Modis vegetation products as proxies of Zhang, J., Bai, Y., Yan, H., Guo, H., Yang, S., Wang, J., 2020. Linking observation,
photosynthetic potential along a gradient of meteorologically and biologically modelling and satellite-based estimation of global land evapotranspiration. Big Earth
driven ecosystem productivity. Biogeosciences 13 (19), 5587–5608. https://doi.org/ Data 4 (2), 94–127. https://doi.org/10.1080/20964471.2020.1743612.
10.5194/bg-13-5587-2016. Zhang, K., Kimball, J.S., Running, S.W., 2016. A review of remote sensing based actual
Schlesinger, W.H., Jasechko, S., 2014. Transpiration in the global water cycle. Agric. For. evapotranspiration estimation. Wiley Interdiscipl. Rev. Water 3 (6), 834–853.
Meteorol. 189, 115–117. Zhao, W., Sánchez, N., Lu, H., Li, A., 2018. A spatial downscaling approach for the smap
Sea, W.B., Choler, P., Beringer, J., Weinmann, R.A., Hutley, L.B., Leuning, R., 2011. passive surface soil moisture product using random forest regression. J. Hydrol. 563,
Documenting improvement in leaf area index estimates from modis using 1009–1024. https://doi.org/10.1016/j.jhydrol.2018.06.081. ISSN 0022-1694.
hemispherical photos for australian savannas. Agric. For. Meteorol. 151 (11), http://www.sciencedirect.com/science/article/pii/S0022169418305031.
1453–1461.

15

You might also like