Abstract

Evapotranspiration (ET) is an important component of the water and energy cycle. To obtain more accurate actual ET estimation, three commonly used actual ET products (i.e., GLEAM, GLDAS, and SSEBop) are merged by the simple Taylor skill score fusion method. Validation results show that the merged actual ET data exhibits better correlation with eddy-covariance ET observation than single-source ET data. Based on the merged ET, the spatiotemporal variations of ET over China and their links to changes of three meteorological factors, i.e., precipitation, air temperature (Ta), and shortwave radiation (Rs), are analyzed. During 2003–2016, there is a general upward trend in ET with an increase rate of 1.6 mm/year over China. ET increases very significantly in northeastern China, the southern coastal region, and the northwestern corner of China, whereas it decreases significantly in Huang-Huai-Hai Plain. There is a clear spatial pattern of the relationship between ET and its driving factors, i.e., the variation of ET in the humid southeast is dominated by Ta and Rs; in the semi-humid and semi-arid zone, the variation of ET is controlled mainly by precipitation, followed by Ta; and in arid northwestern China, the variation of ET is controlled mainly by precipitation.

HIGHLIGHTS

  • The paper offers a feasible and effective method to merge different evapotranspiration data.

  • The quality assessment of merged ET and single ET product indicates the merged data is more reliable and accurate.

  • The Spatiotemporal variation of actual ET over China during 2003–2015 is analyzed.

  • The dominant meteorological factors driving changes of actual ET in different climate regions in China are identified.

INTRODUCTION

Actual evapotranspiration (ETa) is an indispensable element of the global water and energy cycle. Although the spatial and temporal characteristics of ETa in China have been investigated by many researchers, most of them were conducted on regional scales, e.g., the North China Plain, Yangtze River Basin, Haihe River Basin, and the Pearl River Basin (Wang et al. 2011; Li et al. 2013; Matin & Bourque 2013; Zhu et al. 2013; Cao et al. 2014; Gao et al. 2019).The major concern in investigating the spatiotemporal variation of ETa is the lack of reliable ET products. Susceptible to complex geographical conditions and variable climatic conditions (e.g., vegetation type, soil moisture, precipitation, etc.), ETa is difficult to accurately measure (Jiménez et al. 2011). For instance, Gao et al. (2007) showed that ETa decreased in most areas east of 100°E over China but increased in the west and the north parts of northeast China during 1960–2002, and for most parts of China the change in precipitation played a key role in the change of ETa. They recognized that the water balance model which was used in their study to estimate ETa was highly simplified and therefore affected the reliability of their analysis.

Many global ET data products, such as MODIS MOD16, GLEAM, and GLDAS, with broad spatial coverage and long time series, are now available for regional climate research. For instance, He & Shao (2014) used MODIS MOD16 data to analyze the land surface ET over China from 2001 to 2010. Tian et al. (2012) applied the NOAH Land Surface Model to analyze the characteristics of surface ET over China, as well as impact factors, during 1986–2008; Mo et al. (2015) used satellite observations of Normalized Difference Vegetation Index (NDVI) from NOAA-AVHRR and Terra-MODIS, together with climatic data in a physical evapotranspiration (ET) model, to investigate the spatiotemporal variability of ET in terrestrial China from 1981 to 2010. However, while many ET products can provide relatively acceptable ET estimation, they have some degrees of uncertainty in different regions due to model structure, environmental variables involved, quality of data input, and complex local conditions (Miralles et al. 2011a; Chen et al. 2016; Yao et al. 2017). For example, GLEAM ET has good accuracy in semi-arid areas of grassland yet poor performance in semi-humid areas (Yang et al. 2015), whereas MODIS MOD16 ET performs better in exorheic basins than in arid endorheic basins over China (Jiang et al. 2017).

To reduce uncertainties of the research results caused by a single model, the merged data are applied in many studies, and have been proven to exhibit better accuracy and reliability. For instance, Yao et al. (2014) merged five ET products to optimize daily ET using the BMA method, acquiring merged ET with smaller root mean square error (RMSE). Zhu et al. (2016) merged four ET products over north China, discovering that the BMA method for ET data fusion has the advantage of generating more skilful and reliable results than the simple model averaging (SMA) scheme did. Li et al. (2017) applied the Spatial and Temporal Adaptive Reflectance Merged Model to merge ET from three sources. Khan et al. (2018) generated spatially merged ET by choosing ET products with the least uncertainties in different land cover systems. To reduce the complexity of the fusion method and to generate global ET products with high spatial resolution, Yao et al. (2017) developed a simple Taylor skill fusion (STS) method to improve terrestrial ET estimation. Despite simple algorithms and high calculating efficiency, it shows excellent performance in enhancing the accuracy of ET estimation.

To investigate the spatiotemporal variation of evapotranspiration and its response to changes of major meteorological factors over China in the 14 years from 2003 to 2016 on a reliable ET data basis, the STS score fusion method is applied in this study to merge three commonly used ET products. The main objectives of this study are three-fold: (1) to evaluate the applicability of merging multiple ET products to improve the accuracy of ET estimation; (2) to reveal general spatiotemporal variation of ET over China in a recent period using the merged ET; and (3) to reveal the major driving meteorological factors for the variation of ET in different hydrothermal conditions over China.

DATA

ET data products

GLEAM is a global-scale actual ET product based on the Priestley–Taylor equations and sophisticated land surface model using satellite forcing data (Miralles et al. 2011b).

GLDAS-2.1 (Global Land Data Assimilation System) (Rodell et al. 2004) applies land surface model (including NOAH, etc.) and data assimilation techniques to generate land surface variables including actual ET.

The SSEBop (operational Simplified Surface Energy Balance) model produces global actual ET data based on a simplified energy balance approach using remotely sensed observations and climatological data with unique parameterization for operational applications (Senay et al. 2013).

Details about these data products are listed in Table 1.

Table 1

Data used in this study

DataProductUnitWebsiteSpatial resolutionTemporal resolutionTime spanModel or algorithm
ET data GLEAM mm www.gleam.eu 0.25° Daily, aggregated to monthly 2003–2016 Global Land Evaporation Amsterdam Model 
GLDAS 2.1 mm https://ldas.gsfc.nasa.gov/gldas/ 0.25° 3-Hourly, aggregated to monthly 2003–2016 GLDAS 2.1 NOAH 
SSEBop mm https://earlywarning.usgs.gov/fews/product/458 0.25° Monthly 2003–2016 Simplified Surface Energy Balance 
Meteorological data Precipitation mm https://data.cma.cn/data/index.html 0.5° Monthly 2003–2016 Ground-based observation and interpolation 
Reanalysis air temperature; shortwave radiation KW/ https://ldas.gsfc.nasa.gov/gldas/forcing-data 0.25° 3-Hourly, aggregated to monthly 2003–2016 GLDAS 2.1 NOAH 
DataProductUnitWebsiteSpatial resolutionTemporal resolutionTime spanModel or algorithm
ET data GLEAM mm www.gleam.eu 0.25° Daily, aggregated to monthly 2003–2016 Global Land Evaporation Amsterdam Model 
GLDAS 2.1 mm https://ldas.gsfc.nasa.gov/gldas/ 0.25° 3-Hourly, aggregated to monthly 2003–2016 GLDAS 2.1 NOAH 
SSEBop mm https://earlywarning.usgs.gov/fews/product/458 0.25° Monthly 2003–2016 Simplified Surface Energy Balance 
Meteorological data Precipitation mm https://data.cma.cn/data/index.html 0.5° Monthly 2003–2016 Ground-based observation and interpolation 
Reanalysis air temperature; shortwave radiation KW/ https://ldas.gsfc.nasa.gov/gldas/forcing-data 0.25° 3-Hourly, aggregated to monthly 2003–2016 GLDAS 2.1 NOAH 

ET observation from flux tower sites

Monthly ET observation is measured by the eddy-covariance (EC) method in eight flux tower sites (Figure 1) during 2004 and 2005. It is provided by the Chinese Flux Observation and Research Network (ChinaFLUX) (http://www.chinaflux.org/) and used to assess the accuracy of ET by different products and the merged ET data.

Figure 1

Location of eight flux observation sites.

Figure 1

Location of eight flux observation sites.

Meteorological data

Monthly precipitation data are provided by the China Meteorological Data Service Center (http://data.cma.cn/). This gridded dataset is produced by interpolating the precipitation data of 2,472 ground stations over China with the TPS (Thin Plate Spline) method.

Monthly air temperature (Ta) and shortwave radiation (Rs) data are retrieved from the GLDAS 2.1 NOAH model.

Detailed information about all the data used in the present study is presented in Table 1.

RESEARCH METHODS

The simple Taylor skill fusion (STS) method

The STS fusion method uses a weighted average of the individual ET products with weights determined by their Taylor skill scores (Yao et al. 2017). The weight of each ET product can be expressed as:
formula
formula
where is the weight of the ET product i; is the Taylor skill score of ET product i; n is the number of ET products ( is 3 in this paper); is the correlation coefficient between the estimated ET of the product i and the ground measurement of ET with the method of eddy covariance; is the maximum correlation coefficient set as 1.0 in this paper; is the ratio of the standard deviation of the estimated ET of product i to the standard deviation of the corresponding EC ground measurement ET; S varies from zero (least proficient) to one (most proficient).

Data quality measures

Correlation coefficient (r), mean bias error (MBE), mean relative absolute error (MRAE), and root mean square error (RMSE) are selected to assess the quality of each ET dataset. r measures how strong the relationship is between the estimated value and the observed value. Both MBE and MRAE indicate the deviation from the observed value. RMSE is the standard deviation of the estimated value from the observed value.

Mann–Kendall trend test

The Mann–Kendall (MK) trend test (Hamed & Rao 1998) is a commonly used method for detecting the trend of climatic and hydrological time series changes. The τ value of the MK trend test reflects the correlation strength of the sequence with time, and the p-value is the probability of obtaining test results at least as extreme as the results actually observed assuming that the null hypothesis of no trend is correct. We classify the trends into six types in terms of τ value and p-value, as shown in Table 2.

Table 2

Classification of trend based on MK trend test

τp-valueLabelTrend type
τ > 0 p ≤ 0.05 Significant increase 
0.05 < p ≤ 0.1 Slight significant increase 
p > 0.1 Insignificant increase 
τ < 0 p > 0.1 − 1 Insignificant decrease 
0.05 < p ≤ 0.1 − 2 Slight significant decrease 
p ≤ 0.05 − 3 Significant decrease 
τp-valueLabelTrend type
τ > 0 p ≤ 0.05 Significant increase 
0.05 < p ≤ 0.1 Slight significant increase 
p > 0.1 Insignificant increase 
τ < 0 p > 0.1 − 1 Insignificant decrease 
0.05 < p ≤ 0.1 − 2 Slight significant decrease 
p ≤ 0.05 − 3 Significant decrease 

Partial correlation analysis

Partial correlation coefficient (PCC) is an indicator that measures the degree of linear correlation between two of multiple variables by eliminating influences of remaining variables. In our study, PCC among four variables is considered. Among four variables , , , and , PCC between and , denoted as , can be calculated as:
formula
formula
where is the Pearson correlation coefficient between and ; is first-order PCC between and when eliminating the influence of ; is second-order PCC between and when eliminating the influence of and . A null hypothesis is set that and has no correlation and the partial correlation coefficient . Student's t test is performed to test the hypothesis, and the test statistic is calculated as:
formula
where N is the sample size; n is number of variables; r is the PCC. When or p, the null hypothesis is rejected, and , are believed to have correlation. In this study, the critical value of PCC between ET and meteorological factors is 0.217 at the 0.05 significance level. An absolute value of PCC greater than 0.217 indicates a significant correlation between ET and the meteorological factor.

RESULTS

Data quality assessment

Pearson correlation coefficient (r), MBE, MRAE, RMSE of monthly actual ET by GLEAM, GLDAS, and SSEBop are shown in Figure 2. GLEAM gives better ET estimates at three grassland sites (r from 0.93 to 0.98) than at forest and cropland sites (r from 0.58 to 0.95). MBE indicates that GLEAM overestimates ET at forest and cropland sites (MBE > 0) and slightly underestimates ET at grassland sites (MBE < 0), which is basically consistent with previous reports (Miralles et al. 2011a; Yang et al. 2017; Khan et al. 2018). GLDAS performs well in the grassland ecosystem (r from 0.83 to 0.96). Except for Neimenggu station, GLDAS underestimates ET at grassland sites (MBE < 0) and overestimates ET at forest and cropland sites (MBE > 0), which is consistent with the finding of Wang et al. (2016) that the GLDAS NOAH model overestimated ET in Chinese exorheic regions. SSEBop performs well in cropland ecosystems (r = 0.88), which is in agreement with the finding of Chen et al. (2016). It underestimates ET in grassland ecosystems (MBE <0) and overestimates ET in forest and cropland ecosystems (MBE > 0).

Figure 2

Comparison of error measures of merged ET against ET provided by GLEAM, GLDAS, and SSEBop at CBS (Changbaishan), QYZ (Qianyanzhou), DHS (Dinghushang), XSBN (Xishuangbanna), HB (Haibei), NMG (Neimenggu), DX (Dangxiong), and YC (Yucheng).

Figure 2

Comparison of error measures of merged ET against ET provided by GLEAM, GLDAS, and SSEBop at CBS (Changbaishan), QYZ (Qianyanzhou), DHS (Dinghushang), XSBN (Xishuangbanna), HB (Haibei), NMG (Neimenggu), DX (Dangxiong), and YC (Yucheng).

The weights of the GLEAM, GLDAS, and SSEBop ET data calculated by the STS method are shown in Table 3. The data quality measures of merged ET are calculated and also presented in Figure 2 for comparison with individual ET datasets. On the whole, the merged ET has a higher mean value of r (0.87) than individual products, that is, mean r of 0.86 for GLEAM, 0.81 for GLDAS, and 0.74 for SSEBop. What is more, MBE, MRAE, and RMSE of merged ET are also relatively lower than those of individual ET datasets, indicating better accuracy of the merged ET.

Table 3

Weight of ET by each model

GLEAMGLDASSSEBop
W 0.419 0.372 0.209 
GLEAMGLDASSSEBop
W 0.419 0.372 0.209 

The merged ET is compared against ET observation at eight flux tower sites, and the result is shown in Figure 3. At four forest sites, the mean value of r is 0.81 (0.5–0.98). Except at Xishuangbanna where r is quite low (r = 0.49), r at the other three sites is above 0.85, indicating good coherence between the merged ET and ET observation in forest sites. Also, compared with grassland and cropland sites, MRAE, MBE, and RMSE in forest sites are relatively high, and the ET value is overestimated (MBE > 0). In grassland, the merged ET has the best accuracy with the highest mean r of 0.95. The lower MRAE, MBE, and RMSE values prove better performance in grassland. However, the negative MBE reflects underestimation of ET in grassland. Yucheng is a cropland site, and r of 0.840 indicates good reliability of merged ET in cropland. MBE, MRAE, and RMSE values of cropland are lower than those in forest and higher than those in grassland. Overall, the merged ET matches well with ET observations at site scale, and its performance at grassland is the best.

Figure 3

Comparing merged monthly ET against observed ET at eight China flux sites.

Figure 3

Comparing merged monthly ET against observed ET at eight China flux sites.

Spatiotemporal variations in actual evapotranspiration over China

The changes of the average actual ET over China during 2003–2016 based on different datasets are shown in Figure 4. While SSEBop ET data are generally much lower (380–418 mm/a) than GLDAS or GLEAM ET data (442–500 mm/a, 454–478 mm/a), the trends of the datasets are basically consistent, all showing a general upward trend during the past 14 years. The merged ET shows a linear increasing rate of 1.6 mm/a.

Figure 4

Average actual ET over China during 2003–2016.

Figure 4

Average actual ET over China during 2003–2016.

Based on three ET datasets and the merged ET, spatial patterns of annual ET changing rate and the MK trend test results for annual ET calculated on a grid basis over China are presented in Figure 5, which shows remarkable spatial differences of variations of ET indicated by different datasets.

Figure 5

Spatial patterns of changing rate of annual ET (mm/a) of (a) Merged ET, (c) GLDAS ET, (d) GLEAM ET, and (e) SSEBop ET; and MK trend test results of (b) Merged ET, (f) GLDAS ET, (g) GLEAM ET, and (h) SSEBop ET over China during 2003–2016. (Note: for MK trend test, −3 stands for significant decrease; −2 slight decrease; −1 insignificant decrease; 1 insignificant increase; 2 slight increase; 3 significant increase.).

Figure 5

Spatial patterns of changing rate of annual ET (mm/a) of (a) Merged ET, (c) GLDAS ET, (d) GLEAM ET, and (e) SSEBop ET; and MK trend test results of (b) Merged ET, (f) GLDAS ET, (g) GLEAM ET, and (h) SSEBop ET over China during 2003–2016. (Note: for MK trend test, −3 stands for significant decrease; −2 slight decrease; −1 insignificant decrease; 1 insignificant increase; 2 slight increase; 3 significant increase.).

In the Huang-Huai-Hai Plain (HHHP), all three ET products consistently show a downward trend, as does the merged ET, where the linear declining rate is the highest (−0.7 mm/a). Besides HHHP, due to the effects of significant decrease of GLDAS ET in southern Xinjiang (i.e., Tarim Basin) and GLEAM ET in the central Tibetan Plateau (TP), merged ET also exhibits a significant decreasing trend in those two regions.

In northeast China, southern coastal areas and the northwestern corner of China (i.e., northern Xinjiang), ET estimates by the three models all show an upward trend, and consequently merged ET also shows a significant upward trend. As GLDAS ET, GLEAM ET, and SSEBop ET all exhibit fast increases in southern Guangdong Province and northern Hainan Island, the merged ET has the highest increasing rate of about 1.4 mm/a in those regions, as shown in Figure 5(a). As both GLDAS ET and GLEAM ET generally show an increasing trend in eastern TP, whereas SSEBop ET generally shows a decreasing trend in the eastern TP but with very low declining rate, the merged ET also exhibits an upward trend in the eastern part of TP.

Links between variations of merged ET and meteorological factors

Meteorological factors including precipitation, temperature, and surface radiation are crucial factors driving the change of ET, and they are usually positively correlated with ET. Zhou et al. (2009) found that annual precipitation and annual average temperature both have a significant positive correlation with annual average ET. Quan et al. (2016) confirmed that ET was linearly correlated with shortwave radiation (Rs) and highly exponential with air temperature (Ta). However, various factors (e.g., soil moisture content, vegetation type, topography, elevation, wind speed, humidity, etc.) have effects on ET, making the effects of precipitation, Ta, and surface radiation on ET vary among different regions. For example, Teuling et al. (2010) and Seneviratne et al. (2012) pointed out that soil water limited ET only when it is low; however, atmospheric conditions are the main forcing factors in most areas, especially in wet areas. Such a theory is consistent with the findings of De Boeck & Verbeeck (2011).

Links between variations of merged ET and meteorological factors are conducted at monthly time scales. As shown in Figure 6(a)–6(c), in humid southern China where annual precipitation is above 800 mm, except for the Yunnan-Guizhou Plateau (YGP), the partial correlation coefficients (PCC) of monthly actual ET and monthly precipitation are close to zero. However, PCC between monthly ET and monthly average Rs, as well as that between monthly ET and monthly average Ta, is generally higher than 0.4 (>0.217 at the 0.05 significance level). In the southeastern coastal region and the cold and wet southeastern edge of the Tibetan Plateau (TP), PCC between ET and Ta even reaches up to 0.9. Since the southern humid region is rich in water supply for ET, the change of precipitation correlates little with the change of ET. By contrast, the thermal factors, including Ta and Rs that provide energy for the ET process, are the main driving factors of ET variations in these regions.

Figure 6

PCC between monthly ET and (a) precipitation, (b) Rs, and (c) Ta; and MK trend test results of (d) precipitation, (e) Rs, and (f) Ta. (Note: for MK trend test, −3 stands for significant decrease; −2 slight decrease; −1 insignificant decrease; 1 insignificant increase; 2 slight increase; 3 significant increase.).

Figure 6

PCC between monthly ET and (a) precipitation, (b) Rs, and (c) Ta; and MK trend test results of (d) precipitation, (e) Rs, and (f) Ta. (Note: for MK trend test, −3 stands for significant decrease; −2 slight decrease; −1 insignificant decrease; 1 insignificant increase; 2 slight increase; 3 significant increase.).

In the semi-humid and semi-arid areas extending from northeastern China to southwestern China with annual precipitation of 200–800 mm, radiation is generally enough for ET while temperature is moderate, and ET is dominated by water supply from precipitation. Therefore, in this region, precipitation has the highest PCC with ET reaching up to 0.9 in the central TP and northwestern Inner Mongolia, Ta generally is significantly positively related to ET, whereas Rs has the least PCC, mostly close to 0. However, in Tianshan mountain region and northern Xinjiang Province with annual precipitation between 200 and 400 mm, ET is most significantly positively related to Rs (PCC >0.217).

The arid areas of northwestern China where annual precipitation is below 200 mm are endorheic regions where all precipitation is evaporated eventually, and consequently ET and precipitation are highly correlated theoretically. But because of the modification of groundwater and soil water storage on actual ET (Balugani et al. 2017) and high uncertainty in ET estimates in arid areas (Xiong et al. 2015), PCC between ET and precipitation is less high than that in semi-arid and semi-humid regions, albeit it generally reaches 0.6. Meanwhile, PCC is slightly negative between ET and Ta, and close to zero between ET and Rs in the arid regions.

In a word, there is a clear spatial pattern of the relationship between ET and its driving factors, that is, in southeastern China where annual precipitation is mostly over 800 mm, the variation of ET is dominated by Ta and Rs; in the belt extending from northeastern China to southwestern China where annual precipitation varies between 200 and 800 mm, the variation of ET is controlled mainly by precipitation followed by Ta; and in northwestern China where annual precipitation is less than 200 mm, the variation of ET is controlled mainly by precipitation.

Comparison between Figures 6(e)–6(g) and 5(a) indicates that the increase of ET in the southern coastal areas of China is due to the significant increase in Ta; in the middle and lower area of Yangtze River Basin, the Rs shows a decreasing trend, causing the slight decline of ET; ET in the HHHP is affected by the significant decrease in Rs as well as precipitation and, therefore, shows a significant decreasing trend. In arid and semi-arid northern and western China, the spatial distribution of precipitation variation is highly consistent with that of ET. In western TP, southern Xinjiang and western Inner Mongolia, although Ta and Rs rise, the decrease in precipitation results in the decline of ET in these areas because precipitation is the main driving factor of ET variation in these arid areas.

Three regions need special notice in Figure 6. The Yunnan-Guizhou Plateau (YGP), which is located in southwestern China, is an area with annual precipitation over 800 mm, where ET should have less correlation with precipitation but be controlled by Ta and Rs. However, it is shown that ET there has good correlation with precipitation but little correlation with Rs, and lower correlation with Ta than in other areas with the same amount of precipitation. Although the YGP has annual precipitation over 800 mm, climatically it is mostly semi-humid (Wang et al. 2015), and it has been increasingly arid in recent years (Wang et al. 2018), which proves precipitation in the YGP region to be the main factor affecting the amount of actual ET. The TP is another region of concern. The TP is semi-humid with annual precipitation greater than 400 mm in the southeastern part. In other parts of China, ET has stronger correlation with precipitation than Ta and Rs in semi-arid areas. However, in the southeastern TP, ET has the strongest correlation with Ta, rather than precipitation, because it is a cold (due to the high altitude) but comparatively wet region where energy plays a more important role than water availability. The third region needing special attention is the Xinjiang province in northwestern China which is mostly an arid or semi-arid endorheic region. The local water vapor is mainly brought by the westerly winds from the Atlantic Ocean and the dry cold air from the Arctic Ocean. Blocked by high mountains, such as Tianshan Mountain and Altun Mountain, water vapor forms precipitation and makes water more available in the north and west of Xinjiang than in the south and east, in the mountains than in the basin (Su et al. 2007). Therefore, northern and western Xinjiang are more humid than other parts of Xinjiang. ET of these areas is mainly affected by Rs, while in the south and east of Xinjiang, precipitation is still the main forcing factor of ET.

DISCUSSION

  • (1)

    The uncertainties of the ET estimates may result from input errors, parameter definitions, and inadequate model structure (Chen et al. 2016). GLEAM performs better in forest and cropland ecosystems than in grassland ecosystems in estimating the ET (Khan et al. 2018). Also, the GLEAM ET product in forest and cropland ecosystems is overestimated, while this product slightly underestimates ET in grassland ecosystems (Yang et al. 2017). By considering the interception loss of low canopy and high canopy and widely used microwave measurement data, it is capable of working in cloudy weather. However, the GLEAM ET product is greatly affected by the accuracy of satellite data (Miralles et al. 2011a). Therefore, without in-situ observation adjustments, the GLEAM algorithm is generally unable to accurately calculate daily ET. The GLDAS ET product has relatively high errors when applied to estimated ET in cropland and grassland ecosystems in the present study. Over China, GLDAS monthly ET is overestimated in exorheic basins and exhibits the seasonality of ET very poorly in endorheic basins (Wang et al. 2016). Meanwhile, the accuracy of SSEBop ET differs in various land cover types; it performs best for cropland (Chen et al. 2016). SSEBop is sensitive to land surface temperature (LST) (Chen et al. 2016), whereas satellite imagery can underestimate LST on some non-vegetated surfaces with high albedo (such as desert sands) or high emissivity (such as lava rocks), consequently, SSEBop may overestimate ET for these surfaces (McShane et al. 2017). Therefore, it would be helpful to merge several ET products so as to achieve better accuracy of ET estimation. In this study, compared with a single model, the accuracy of the merged ET product is improved. The correlation between the merged ET and ET observation in eight flux stations are all strong, avoiding the poor quality of individual products at some stations (e.g., GLDAS ET at Xishuangbanna and SSEBop ET at Dangxiong).

  • (2)

    Variations of ET result from changes of multi-meteorological factors. However, when analyzing forcing factors of ET over a large scale such as China, researchers tend to analyze the links between ET and its drivers separately, ignoring the correlation between the drivers themselves. For example, Yang et al. (2015) and He & Shao (2014) simply analyzed the relationship between spatial distributions of changing rate of ET and that of forcing factors. Quan et al. (2016) applied Pearson's correlation coefficient and exponential fit to the investigation of the relationship between ET and Ta, and ET and radiation, which can only reflect correlativity between two variables. Chen et al. (2013) applied the coefficient of determination to discuss correlation between ET and several environmental variables. To illustrate the difference of correlation and partial correlation, correlation coefficient (CC) between ET and main meteorological factors is calculated and plotted in Figure 7. By comparing PCC in Figure 6 and CC in Figure 7, we can find that CCs between ET and meteorological drivers (Figure 7) are relatively higher all over China than PCCs (Figure 6), and hence, it exaggerates the correlation between ET and meteorological drivers or even indicates specious correlation in certain regions such as the middle area of Yangtze River Basin. Consequently, ignoring interrelationship between meteorological factors will cause misinterpretation of the role certain meteorological factors play in the changing process of ET. To exclude the impacts between the meteorological factors, the partial correlation is recommended for such kinds of driver impact analysis. In addition, since the variation of ET is related to other factors (e.g., wind speed, vegetation type, sunshine hours, etc.) (Dang et al. 2016; Deng et al. 2017; Hu et al. 2017), more factors should be taken into account in future research to study their impacts on ET.

Figure 7

Spatial pattern of correlation coefficient between ET and (a) precipitation, (b) Rs, and (c) Ta.

Figure 7

Spatial pattern of correlation coefficient between ET and (a) precipitation, (b) Rs, and (c) Ta.

CONCLUSIONS

Three actual ET data products, i.e., GLEAM, GLDAS, and SSEBop, are merged and applied to the investigation of the spatiotemporal variation of actual ET over China, as well as links between actual ET and meteorological drivers. The results show that:

  • (1)

    Compared with ET provided by any single product, the merged ET has better accuracy. Its mean value of r at eight eddy-covariance ET observation sites is 0.87, higher than that of GLEAM (0.86), GLDAS (0.81), and SSEBop (0.74). In grassland, it has the best performance with the highest mean value of r (0.95) and relatively lower MBE, RMSE, and MRAE.

  • (2)

    During 2003 and 2016, actual ET over China showed an overall upward trend with a linear increase rate of 1.6 mm/a. The merged ET decreases significantly in Huang-Huai-Hai Plain, while it increases significantly in northeast China, the southern coastal region, and the northwestern corner of China.

  • (3)

    The dominant factor driving the spatial-temporal variation differs depending on regional hydrothermal conditions. In humid southeastern China where annual precipitation is mostly over 800 mm, the variation of ET is dominated by Ta and Rs; in the semi-humid/semi-arid zone extending from northeastern China to southwestern China where annual precipitation varies between 200 and 800 mm, the variation of ET is controlled mainly by precipitation followed by Ta; and in northwestern China where annual precipitation is less than 200 mm the variation of ET is controlled mainly by precipitation.

ACKNOWLEDGEMENTS

This research is funded by National Key Research and Development Program project (No. 2017YFC0405801-02) and National Natural Science Foundation of China (No. 41961134003, 41830752). The authors are grateful to anonymous reviewers for their constructive comments and suggestions.

REFERENCES

Balugani
E.
Lubczynski
M. W.
Reyes-Acosta
L.
van der Tol
C.
Francés
A. P.
Metselaar
K.
2017
Groundwater and unsaturated zone evaporation and transpiration in a semi-arid open woodland
.
Journal of Hydrology
547
,
54
66
.
Chen
Y.
Xia
J. Z.
Liang
S. L.
Feng
J. M.
Fisher
J. B.
Li
X.
Li
X. L.
Liu
S. G.
Ma
Z. G.
Miyata
A.
Mu
Q. Z.
Sun
L.
Tang
J. W.
Wang
K. C.
Wen
J.
Xue
Y. J.
Yu
G. R.
Zha
T. G.
Zhang
L.
Zhang
Q.
Zhao
T. B.
Zhao
L.
Yuan
W. P.
2013
Comparison of satellite-based evapotranspiration models over terrestrial ecosystems over China
.
Remote Sensing of Environment
140
,
279
293
.
Dang
S. Z.
Jiang
X. H.
Dong
G. T.
Chen
X. C.
Bai
L.
2016
Variation of actual evapotranspiration and its causes analysis in the upper reaches of the Jinghe River Basin
.
Research of Soil and Water Conservation
23
,
143
147
(in Chinese)
.
Deng
X. Y.
Liu
Y.
Liu
Z. H.
Yao
J. Q.
2017
Temporal-spatial dynamic change characteristics of evapotranspiration in arid region of Northwest China
.
Acta Ecologica Sinica
37
,
2994
3008
(in Chinese)
.
Gao
G.
Chen
D.
Xu
C.
Simelton
E.
2007
Trend of estimated actual evapotranspiration over China during 1960–2002
.
Journal of Geophysical Research Atmosphere
112
,
D11
.
Hamed
K. H.
Rao
A. R.
1998
A modified Mann-Kendall trend test for autocorrelated data
.
Journal of Hydrology
204
,
182
196
.
He
T.
Shao
Q. Q.
2014
Spatial-temporal variation of terrestrial evapotranspiration over China from 2001 to 2010 using MOD16 products
.
Journal of Geo-Information Science
16
,
979
988
(in Chinese)
.
Hu
Q.
Dong
B.
Pan
X. B.
Wang
X. X.
Wei
P.
Zhao
H. H.
Zhang
X. T.
2017
Spatiotemporal variation and causes analysis of dry-wet climate at different time scales in North China Plain
.
Chinese Journal of Agrometeorology
38
,
267
277
(in Chinese)
.
Jiang
Y. Y.
Wang
W.
Zhou
Z. H.
2017
Evaluation of MODIS MOD16 evapotranspiration product in Chinese River Basins
.
Journal of Natural Resources
32
,
517
528
.
Jiménez
C.
Prigent
C.
Mueller
B.
Seneviratne
S. I.
McCabe
M. F.
Wood
E. F.
Rossow
W. B.
Balsamo
G.
Betts
A. K.
Dirmeyer
P. A.
Fisher
J. B.
Jung
M.
Kanamitsu
M.
Reichle
R. H.
Reichstein
M.
Rodell
M.
Sheffield
J.
Tu
K.
Wang
K.
2011
Global intercomparison of 12 land surface heat flux estimates
.
Journal of Geophysical Research
116
,
1
27
.
Li
X.
Gemmer
M.
Zhai
J.
Liu
X.
Su
B.
Wang
Y.
2013
Spatio-temporal variation of actual evapotranspiration in the Haihe River Basin of the past 50 years
.
Quaternary International
304
,
133
141
.
McShane
R. R.
Driscoll
K. P.
Sando
R.
2017
A Review of Surface Energy Balance Models for Estimating Actual Evapotranspiration with Remote Sensing at High Spatiotemporal Resolution Over Large Extents. Scientific Investigations Report 2017-5087
.
U.S. Geological Survey
,
Reston, VA
,
USA
, p.
19
.
https://doi.org/10.3133/sir20175087
Miralles
D. G.
Holmes
T. R. H.
De Jeu
R. A. M.
Gash
J. H.
Meesters
A. G. C. A.
Dolman
A. J.
2011a
Global land-surface evaporation estimated from satellite-based observations
.
Hydrology and Earth System Sciences
15
,
453
469
.
Miralles
D. G.
De Jeu
R. A. M.
Gash
J. H.
Holmes
T. R. H.
Dolman
A. J.
2011b
Magnitude and variability of land evaporation and its components at the global scale
.
Hydrology and Earth System Sciences
15
,
967
981
.
Quan
C.
Zhou
B. R.
Han
Y. X.
Zhao
T. L.
Xiao
J. S.
2016
A study of evapotranspiration on the degraded alpine wetland surface in the Yangtze River source area
.
Journal of Glaciology and Geocryology
38
(
5
),
1249
1257
(in Chinese)
.
Rodell
M.
Houser
P. R.
Jambor
U.
Gottschalck
J.
Mitchell
K.
Meng
C. J.
Arsenault
K.
Cosgrove
B.
Radakovich
J.
Bosilovich
M.
Entin
J. K.
Walker
J. P.
Lohmann
D.
Toll
D.
2004
The global land data assimilation system
.
Bulletin of the American Meteorological Society
85
,
381
394
.
Senay
G. B.
Bohms
S.
Singh
R. K.
Gowda
P. H.
Velpuri
N. M.
Alemu
H.
Verdin
J. P.
2013
Operational evapotranspiration mapping using remote sensing and weather datasets: a new parameterization for the SSEB approach
.
Journal of the American Water Resources Association
49
(
3
),
577
591
.
Seneviratne
S. I.
Lehner
I.
Gurtz
J.
Teuling
A. J.
Lang
H.
Moser
U.
Grebner
D.
Menzel
L.
Schroff
K.
Vitvar
K.
Zappa
M.
2012
Swiss prealpine Rietholzbach research catchment and lysimeter: 32-year time series and 2003 drought event
.
Water Resources Research
48
,
W06526
.
Su
H. C.
Shen
Y. P.
Han
P.
Li
J.
Lan
Y. C.
2007
Precipitation and its impact on water resources and ecological environment in Xinjiang region
.
Journal of Glaciology and Geocryology
3
,
343
350
(in Chinese)
.
Teuling
A. J.
Seneviratne
S. I.
Stöckli
R.
Reichstein
M.
Moors
E.
Ciais
P.
Luyssaert
S.
Hurk
B.
Ammann
C.
Bernhofer
C.
Dellwik
E.
Gianelle
D.
Gielen
B.
Grünwald
T.
Klumpp
K.
Montagnani
L.
Moureaux
C.
Sottocornola
M.
Wohlfahrt
G.
2010
Contrasting response of European forest and grassland energy exchange to heatwaves
.
Nature Geoscience
3
,
722
727
.
Tian
J.
Su
H. B.
Chen
S. H.
Sun
X. M.
Chen
Q. M.
2012
Spatial-temporal variations of evapotranspiration over China mainland in recent 20 years
.
Resources Science
34
,
1277
1286
(in Chinese)
.
Wang
Y.
Liu
B.
Su
B.
Zhai
J.
Gemmer
M.
2011
Trends of calculated and simulated actual evaporation in the Yangtze River Basin
.
Journal of Climate
24
,
4494
4507
.
Wang
W.
Zhu
Y.
Xu
R. G.
Liu
J. T.
2015
Drought severity change in China during 1961–2012 indicated by SPI and SPEI
.
Natural Hazards
75
(
3
),
2437
2451
.
Wang
W.
Huang
J.
Cui
W.
2018
Comparison of drought indices for remote sensing drought monitoring in Yunnan-Guizhou Plateau region
.
Transactions of the Chinese Society of Agricultural Engineering
34
,
131
139 + 309
(in Chinese)
.
Yang
X. Q.
Wang
G. J.
Pan
X.
Zhang
Y. Q.
2015
Spatial-temporal variability of terrestrial evapotranspiration over China from 1980 to 2011 based on GLEAM data
.
Transactions of the Chinese Society of Agricultural Engineering
21
,
132
141
(in Chinese)
.
Yang
X. Q.
Yong
B.
Ren
L.
Zhang
Y.
Long
D.
2017
Multi-scale validation of GLEAM evapotranspiration products over China via ChinaFLUX ET measurements
.
International Journal of Remote Sensing
38
,
5688
5709
.
Yao
Y. J.
Liang
S. L.
Li
X. L.
Zhang
Y. H.
Chen
J. Q.
Jia
K.
Zhang
X. T.
Fisher
J. B.
Wang
X. Y.
Zhang
L. L.
Xu
J.
Shao
C. L.
Posse
G.
Li
Y. N.
Magliulo
V.
Varlagin
A.
Moors
E. J.
Boike
J.
Macfarlane
C.
Kato
T.
Buchmann
N.
Billesbach
D. P.
Beringer
J.
Wolf
S.
Papuga
S. A.
Wohlfahrt
G.
Montagnani
L.
Ardo
J.
Limoges
E. P.
Emmel
C.
Hortnagl
L.
Sachs
T.
Gruening
C.
Gioliw
B.
Ballesterosx
A. L.
Steinbrechery
R.
Gielenz
B.
2014
Bayesian multi-model estimation of global terrestrial latent heat flux from eddy covariance, meteorological, and satellite observations
.
Journal of Geophysical Research Atmospheres
119
,
4521
4545
.
Yao
Y. J.
Liang
S. L.
Li
X. L.
Zhang
Y. H.
Chen
J. Q.
Jia
K.
Zhang
X. T.
Fisher
J. B.
Wang
X. Y.
Zhang
L. L.
Xu
J.
Shao
C. L.
Posse
G.
Li
Y. N.
Magliulo
V.
Varlagin
A.
Moors
E. J.
Boike
J.
Macfarlane
C.
Kato
T.
Buchmann
N.
Billesbach
D. P.
Beringer
J.
Wolf
S.
Papuga
S. A.
Wohlfahrt
G.
Montagnani
L.
Ardo
J.
Limoges
E. P.
Emmel
C.
Hortnagl
L.
Sachs
T.
Gruening
C.
Gioliw
B.
Ballesterosx
A. L.
Steinbrechery
R.
Gielenz
B.
2017
Estimation of high-resolution terrestrial evapotranspiration from Landsat data using a simple Taylor skill fusion method
.
Journal of Hydrology
553
,
508
526
.
Zhou
L.
Wang
S. Q.
Chen
J. M.
Feng
X. F.
Ju
W. M.
Wu
W. X.
2009
The spatial-temporal characteristics of evapotranspiration of China's terrestrial ecosystems during 1991–2000
.
Resources Science
31
,
962
972
(in Chinese)
.
Zhu
G. F.
Li
X.
Zhang
K.
Ding
Z. Y.
Han
T.
Ma
J. Z.
Huang
C. L.
He
J. H.
Ma
T.
2016
Multi-model ensemble prediction of terrestrial evapotranspiration across north China using Bayesian model averaging
.
Hydrological Processes
30
,
2861
2879
.
This is an Open Access article distributed under the terms of the Creative Commons Attribution Licence (CC BY-NC-ND 4.0), which permits copying and redistribution for non-commercial purposes with no derivatives, provided the original work is properly cited (http://creativecommons.org/licenses/by-nc-nd/4.0/).