The Tibetan Plateau (TP) plays a major role in regional and global climate.
The understanding of latent heat (LE) flux can help to better describe the complex
mechanisms and interactions between land and atmosphere. Despite its
importance, accurate estimation of evapotranspiration (ET) over the TP
remains challenging. Satellite observations allow for ET estimation at high
temporal and spatial scales. The purpose of this paper is to provide a
detailed cross-comparison of existing ET products over the TP. Six available
ET products based on different approaches are included for comparison.
Results show that all products capture the seasonal variability well with
minimum ET in the winter and maximum ET in the summer. Regarding the spatial
pattern, the High resOlution Land Atmosphere surface Parameters from Space (HOLAPS) ET
demonstrator dataset is very similar to the LandFlux-EVAL dataset (a
benchmark ET product from the Global Energy and Water Cycle Experiment), with
decreasing ET from the south-east to north-west over the TP. Further comparison
against the LandFlux-EVAL over different sub-regions that are decided by
different intervals of normalised difference vegetation index (NDVI),
precipitation, and elevation reveals that HOLAPS agrees best with
LandFlux-EVAL having the highest correlation coefficient (
Evapotranspiration (ET) is an essential nexus of energy and water cycles
through the mass and energy interactions between land and atmosphere (Jung et
al., 2010; Peng et al., 2013a). The estimation of spatially distributed ET
has been advanced by the progress of satellite remote sensing technology.
However, remote sensing techniques do not allow to directly inverting ET from
space (Peng et al., 2013b; K. Zhang et al., 2016). Different methods have been
therefore developed to estimate ET with the use of physical variables that
are sensed by satellite and are related to the evaporation process (Kalma et
al., 2008; Wang and Dickinson, 2012). In recent years, a number of global ET
products have been generated with the availability of long-term global
satellite products and progress in computer science (Zhang et al., 2010;
Vinukollu et al., 2011b; Miralles et al., 2011; Fisher et al., 2008). Some of
these global products can even provide ET with a spatial resolution less than
10 km and temporal resolution less than 3 h (Mu et al., 2007; Miralles et
al., 2016; Loew et al., 2016). HOLAPS (High resOlution Land Atmosphere surface Parameters from Space) demonstrator dataset is one of them. HOLAPS is
actually a framework that can provide surface energy and water fluxes at
sub-hourly timescales and spatial resolutions at the kilometre scale. It is
also worth noting that very high spatial resolution (on the order of 10 m)
ET product at regional scale can be provided by ALEXI/DisALEXI based on
thermal observations from polar and geostationary orbiting satellites
(Anderson et al., 2011, 2007). Although these global ET products have been
applied to many applications such as multi-decadal trend analysis (Y. Zhang
et al., 2016; Zhang et al., 2015; Miralles et al., 2014; Jiménez et al.,
2011), large discrepancies exist in these products. Within the Global
Energy and Water Cycle Experiment (GEWEX) LandFlux initiative, Mueller et
al. (2011) conducted a comparison of existing global latent heat (LE) products from either
land surface models, re-analysis, or satellite estimates, and found that the
global mean LE over land was
Nevertheless, theses global ET products have great potentials for global and regional hydrological applications. In this study, the performances of the widely used global ET products will be investigated over the Tibetan Plateau (TP), as the ET over the TP is of great importance and research interest. The TP has strong impacts on weather and climate at the regional to global scale and controls climatic and environmental changes in Asia and elsewhere in the Northern Hemisphere (Ma et al., 2008). The knowledge of ET is essential for the study of land–atmosphere interactions, and assessment of the impacts of and feedbacks to the global change (Shi and Liang, 2014). In order to characterise the distribution of ET over the TP, different methods using micrometeorological measurements (Yang et al., 2003; Lee et al., 2012; Chen et al., 2013b; Zhang et al., 2007), remote sensing products (Ma et al., 2014, 2006; Chen et al., 2013a) and the combined use of both data sources (Ma et al., 2003, 2011; You et al., 2014) have been investigated over the last decades. In addition, land surface models have also been applied to simulate ET over the TP (Gerken et al., 2012; Yang et al., 2009). However, accurate estimation of ET over TP is still a challenge due to the limitations of the above approaches. Specifically, the observation-based methods are not adequate for determination of regional ET due to the limited spatial representativeness of meteorological stations, while the remote sensing products are only available under clear-sky conditions. The models results are limited by the accuracy of input parameters and the uncertainties of model parameterisation over complicated topography and highly heterogeneous areas of the TP (Shi and Liang, 2013b). The existing global ET products, especially those with high spatial and temporal resolutions such as HOLAPS, provide a potentially applicable ET dataset over the TP. Although the global ET products have been validated against FLUXNET measurements, the reliability of spatial and temporal patterns of them over the TP is still unknown. A comprehensive analysis of the characteristics of the LE over the TP based on the state-of-the-art global ET products has not yet been conducted. Therefore, the main objective of this study is to provide a detailed cross-comparison of the different existing ET products over the TP. Through this study, the following research questions will be addressed. (1) Do existing global ET products show consistent spatial and temporal patterns over the TP? (2) Are there systematic deviations between the different data products, which can be explained by different climate or surface conditions? The study will focus mainly on a cross-comparison between the different existing datasets due to a lack of appropriate reference data in the region, as will be discussed.
The TP, known as the third pole of the Earth (Qiu, 2008),
covers approximately the latitude from 26 to 40
Map of the location and topography of the Tibetan Plateau.
Different groups of algorithms have been developed to estimate ET from
satellite data. These comprise (1) surface energy balance models forced
either by satellite remote sensing or re-analysis data (Bastiaanssen et al.,
1998; Su, 2002), (2) the methods based on PM or PT equations (Fisher et al., 2008; Miralles et al., 2011; Mu et
al., 2007; Zhang et al., 2015), (3) spatial variability methods (Peng et al.,
2013b; Peng and Loew, 2014; Roerink et al., 2000). Among them, the PM
algorithm, the PT model, and the SEBS are
widely used, and have been explored by both the GEWEX LandFlux-EVAL initiative
and the Water Cycle Multi-mission Observation Strategy EvapoTranspiration
(WACMOS-ET) project. Therefore, three LE datasets based these models and
driven by the same forcing data are compared over the TP in this study. These
datasets are SEBS
SEBS is a one-source energy balance algorithm, which firstly calculates the
sensible heat flux (
Summary of the datasets used in our study.
The PM
The PT-JPL model by Fisher et al. (2008) is used to estimate
PT
The HOLAPS LE product was generated from HOLAPS framework, which makes use of
meteorological drivers coming exclusively from a globally available satellite
and re-analysis datasets and is based on a state-of-the-art land surface
scheme (Loew et al., 2016). It is based on a radiation module, a planetary
boundary layer model, a soil module, and a general module for the exchange of
energy and moisture at the surface layer. HOLAPS can ensure internal
consistency of the different energy and water fluxes and provide estimates at
high temporal (
The validation of different LE datasets against in situ measurements over the
TP is not possible for the current study period due to (a) the access to
suitable in situ measurements is not possible and (b) spatial representativeness
of the existing FLUXNET towers for areas of only several square kilometres.
Therefore, the above LE datasets are cross-compared with LandFlux-EVAL
benchmark product in the current analysis. LandFlux-EVAL is a merged
synthesis LE product based on a total of 14 datasets including land surface
model output, observations-based estimates, and atmospheric reanalyses
(Mueller et al., 2013). It provides the best guess estimate of LE for the
first time based on the existing global LE datasets, and also provides the
uncertainty range of the absolute LE values (interquartile range of the
merged synthesis LE products). Note that the merged LE dataset agreed well
with precipitation minus runoff over large river basins around the world
(Mueller et al., 2011), and it has been used to evaluate the LE simulations
of the fifth phase of the Coupled model Inter-comparison project (CMIP5)
(Mueller and Seneviratne, 2014). To further demonstrate the validity of
LandFlux-EVAL benchmark product over the TP, we also compared it to
precipitation, which is one of the most important driving factors for LE. It
should be noted here that LandFlux-EVAL also includes satellite-based LE
datasets that are estimated from PM and PT algorithms. However, the
PM
Spatial distribution of annual mean LandFlux-EVAL LE and GPCP precipitation over the TP (left panel). The scatter plots of the comparison between LE and precipitation for all the pixels (right panel).
Spatial distribution of annual mean LE for each dataset over the TP.
All of the datasets were firstly aggregated to monthly mean values over the
common time period 2001–2005, which corresponds to the temporal resolution of
LandFlux-EVAL benchmark product and the time period currently covered by the
HOLAPS demonstrator dataset (Loew et al., 2016; Mueller et al., 2013). To
make an unbiased comparison with LandFlux-EVAL dataset, HOLAPS and
SEBS
The annual mean spatial patterns of 25th percentile and 75th percentile of the LandFlux-EVAL multi-datasets ensemble.
The characteristics of all the datasets were investigated through spatial and temporal analysis. The spatial distributions of the seasonal and annual average LE over the TP were analysed, including the identification of patterns such as low and high values, and the investigation of seasonal changes. The four seasons are defined as autumn (September–October–November), winter (December–January–February), spring (March–April–May), and summer (June–July–August). The temporal analysis explored the seasonal and annual variation of all the datasets from 2001 to 2005 over the whole TP. In addition, the correlation analysis was conducted to evaluate the impacts of climate (precipitation) and surface conditions (normalised difference vegetation index and elevation) on the performance of ET estimation. The relationship between different LE products and the LandFlux-EVAL benchmark product were quantified by using correlation coefficient and root mean square deviation over the whole TP and different sub-regions, which were decided by different intervals of normalised difference vegetation index (NDVI; generated from MODIS), precipitation (Global Precipitation Climatology Project; GPCP), and elevation (Global Multi-resolution Terrain Elevation Data 2010, GMTED2010).
The spatial distributions of annual mean LandFlux-EVAL and precipitation are shown in Fig. 2. It can be seen that the LE has similar patterns as observation-based precipitation, both decreasing from south-east to north-west over the TP. The comparison of all the pixels shows a very high correlation coefficient of 0.9 between LE and precipitation. Besides precipitation, the radiation is another important driver for LE. Compared to the published studies, the LandFlux-EVAL LE also corresponds well with the merged net radiation and LE datasets, which were developed and validated over the TP by Shi and Liang (2013a, b) and Shi and Liang (2014). The spatial distribution of annual mean net radiation and LE can be found in study of Shi and Liang (2013a) and Shi and Liang (2014). Although the LandFlux-EVAL has not been validated against in situ measurements over the TP, the similar spatial patterns between LE and both observation-based precipitation and validated radiation to some extent demonstrate the validity of LandFlux-EVAL over the TP.
Differences of spatial distribution of annual mean LE between LandFlux-EVAL and other datasets over the TP.
Figure 3 displays the spatial pattern of annual mean values for different LE
datasets. Although these LE products have been reported performing well
against FLUXNET measurements at point scale, they exhibit differently in
terms of spatial pattern over the TP. In general, the LandFlux-EVAL, HOLAPS
and SEBS
Figure 4 further shows the annual mean spatial patterns of 25th percentile
and 75th percentile of the LandFlux-EVAL multi-dataset ensemble, which
quantifies the uncertainty range of the absolute LE values (interquartile
range of the merged synthesis LE products). It can be seen that HOLAPS and
most parts of PT
Besides the analysis of spatial distribution of annual mean, the seasonal
means of each LE dataset are also shown in Fig. 6. It can be seen that all
the LE datasets show clear seasonal cycles with the highest values in summer and
the lowest values in winter, which might be related to both westerlies and Asian
monsoon. Due to the influence of Asian summer monsoon, the highest LE in
LandFlux-EVAL is in south-eastern TP and the LE decreases to north-west. The
lowest LE appears in northern TP where dry westerlies dominate. Similar
patterns are also found in HOLAPS, PT
In addition to the spatial comparisons of annual and seasonal mean values,
the time evolution of all datasets is also explored. Figure 7 presents the
time series of the area mean LE for different LE datasets, and the
inter-quartile range between 25th percentile and 75th percentile of the
LandFlux-EVAL ensemble. According to Fig. 7, all products capture well the
seasonal variability with minimum LE in the winter and maximum LE in the
summer. However, the mean values of different LE products differ
substantially. There is a spread of about 35 W m
Correlation coefficient (
Spatial distribution of seasonal mean LE for each dataset over the
TP.
Statistics of the LE comparisons between the LandFlux-EVAL benchmark product and other products over the whole TP.
Statistics of the LE comparisons between the LandFlux-EVAL benchmark product and other products over different NDVI thresholds.
Temporal variability of the area-averaged LE for each dataset over the TP. The grey shadow displays the inter-quartile range between 25th percentile and 75th percentile of the LandFlux-EVAL multi-datasets ensemble.
Temporal variability of the area-averaged LE for 5-day HOLAPS over the TP.
The monthly mean scatter plots of LE between the LandFlux-EVAL benchmark product and other products over the whole TP.
Statistics of the LE comparisons between the LandFlux-EVAL benchmark product and other products over different precipitation thresholds.
The monthly mean scatter plots of LE between the LandFlux-EVAL benchmark product and other products over different NDVI thresholds.
Figure 9 presents the monthly mean scatter plots of LE between the
LandFlux-EVAL benchmark product and other products over the whole TP. The
detailed statistics are listed in Table 3. It can be seen that the model
performance varies among different LE products with statistical indices
values ranging from 0.91 to 0.99 for correlation coefficient (
The monthly mean scatter plots of LE between the LandFlux-EVAL benchmark product and other products over different precipitation thresholds.
The monthly mean scatter plots of LE between the LandFlux-EVAL benchmark product and other products over different elevation thresholds.
Statistics of the LE comparisons between the LandFlux-EVAL benchmark product and other products over different elevation thresholds.
The spatial and temporal inter-comparisons of different global LE datasets
over the TP suggest that there are large differences among different
datasets. The LandFlux-EVAL benchmark product was found to agree well with
observation-based precipitation, in situ measurements-validated
radiation (Shi and Liang, 2013a), and in situ measurements-validated
LE product (Shi and Liang, 2014). From this point of view, it can be served
as the reference dataset. The HOLAPS is found to agree temporally and
spatially well with LandFlux-EVAL benchmark product. The PT
This study provides a first comprehensive inter-comparison of existing LE
products over the TP for the period 2001–2005. The results of the study can
be summarised as follows:
The existing global LE products show substantial differences in spatial
and temporal patterns over the TP, although all these products have been
found to agree well with FLUXNET measurements in different climate
conditions. The LandFlux-EVAL benchmark product as well as the HOLAPS LE show very
similar spatial patterns, both with LE increasing from north-west to
south-east. The other LE products (SEBS Further comparison against LandFlux-EVAL benchmark dataset over the whole
TP and sub-regions that are decided by different intervals of NDVI,
precipitation, and elevation reveals that climate and surface conditions have
impacts on the performances of SEBS
Overall, there are still large uncertainties in the current global LE dataset over the TP. In order to accurately estimate LE over the TP, model calibration ad development of high accuracy forcing dataset are still needed. There is therefore a strong need for appropriate in situ flux measurements as well as other hydrological data such as, e.g. runoff measurements.
This study uses the LandFlux-EVAL merged benchmark synthesis products of ETH
Zurich produced under the aegis of the GEWEX and ILEAPS
projects
(
This research was supported by the Cluster of Excellence CliSAP (EXC177), University of Hamburg, funded through the German Science Foundation (DFG), and the MPG-CAS postdoc fellowship. The authors would like to thank Stefan Hagemann for reviewing the first version of the manuscript. The authors are also very grateful to the editor and six anonymous reviewers for their valuable comments that helped improve the manuscript.The article processing charges for this open-access publication were covered by the Max Planck Society. Edited by: L. Samaniego Reviewed by: six anonymous referees