Evapotranspiration increment was underestimated in China due to underrepresented land cover changes

Numerous evapotranspiration (ET) products have been produced using various approaches and diverse forcing data even as the magnitude and trends of ET show divergence. We simulated ET using updated land use and cover change (LUCC) data in China from 1900 to 2020. We found that China’s ET increased slightly from 1900 to 1980, but it increased rapidly after 1980 due to LUCC characterized by forest expansion (2.05 mm yr−1, P < 0.01). We also found that the ET trends derived from our simulation were significantly higher than other ET products (−0.70–1.47 mm yr−1, P < 0.01), implying that existing, long-term ET products might have underestimated ET trends in China during the post-1980 period because of underrepresented LUCC. These underestimated ET trends could introduce biases in the regional water budget and water resources management. We advocate for future studies to take into account the impacts of LUCC in global ET simulations.


Introduction
Evapotranspiration (ET) is one of the essential ecosystem processes linking the global water, carbon, and energy cycles (Yang et al 2023).ET in regional and global scale changes under the influence of precipitation, temperature, solar radiation, wind speed, and land cover changes.Changes in ET will feedback to the climatic system by altering the drought or wetness degrees in different regions (Guan et al 2018, Zhang et al 2018).ET also plays a critical role in plant growth, water resource utilization, ecological protection, and global climate change (Fisher et al 2017).ET changes may affect the structure and function of ecosystems (Yu et al 2022a), threatening water resource security and ecological environment stability in many countries and regions, and thus affecting human life and economic development (Cheng et al 2020, Condon et al 2020).Improving the accuracy and reliability of ET estimates will be beneficial for addressing climate change and water resources challenges (Fu et al 2022), including droughts, flooding, irrigations, and ecosystem water use efficiency (Sheffield et al 2012, Sun et al 2017a, Jalilvand et al 2019).
Many ET products are currently available, ranging from regional to global scale, and use diverse approaches, such as the remote sensing model (e.g.i.e.MOD16, Mu et al 2011), water balance model (e.g.PEW, Fu et al 2022;WBMplus, Wisser et al 2010), and process-based ecosystem model (e.g.Multi-scale synthesis and Terrestrial Model Intercomparison Project (MsTMIP), (Huntzinger et al 2013(Huntzinger et al , 2018(Huntzinger et al , 2021)); Trends in land carbon cycle (TRENDY), Friedlingstein et al 2020).However, large uncertainties remain in these ET products owing to the diverse approaches, parameterization and validation data, and climatic, atmospheric, and land cover forcing datasets used.Among all these factors, land use and cover change (LUCC) is probably the most diversified and uncertain driver in ET simulation in these products.For example, MOD16 was produced relying on Moderate Resolution Imaging Spectroradiometer (MODIS) land land cover data (e.g.MCD12Q1), a product that might have limited capability in capturing land cover changes in China (Yang and Huang 2021).CR-ET a was primarily driven by meteorological data, while it ignored land surface information such as vegetation and soil status (Ma et al 2019), which might lead to bias in ET estimation.
Human activities such as urbanization and agricultural development in China have led to drastic land-use changes and water environment problems (Qiu and Zhang 2019).In recent years, China has strengthened its efforts to protect and restore the ecological environment, and measures such as returning poor farmland to forest and grassland help to improve vegetation cover and water conservation capacity, thus increasing ET (Bai et al 2023).In addition, atmospheric nitrogen (N) deposition in China was reported to be three to four times that in the United States and Europe, stimulating forest growth, reshaping forest structure (Huang et al 2015), and further interacting with LUCC to alter ecosystem ET.To sum up, the capability of LUCC datasets to capture huge land changes determines the spatial-temporal reliability of ET products and the attributions of various factors' impacts on ET.This is also supported by a former study which reported that the contribution of the land area change caused by vegetation greening to the global ET increase reached 39% ± 18% in 1982-2011, and even higher at 62% ± 36% in 2011-2020 (Yang et al 2023).Nonetheless, previous studies revealed that the LUCC data used in international model comparison projects, such as TRENDY and MsTMIP, have largely biased the depiction of cropland and forest area dynamics in China (Yu et al 2021, 2022b, Xia et al 2023).Using the bias-corrected LUCC dataset previously developed, the present study simulated and examined the spatial-temporal changes of ET over China from 1900 to 2020.By comparing the simulated results with published ET products, we quantified the biases of the ET datasets in China and attributed the impacts of LUCC and changes in climatic and atmospheric conditions (e.g.rising CO 2 , nitrogen deposition) on terrestrial ET in China.Intensive and persistent land cover change, as characterized by forestation, is also anticipated in China for carbon sequestration enhancement to achieve the carbon neutrality goal before 2060.An accurate assessment of the historical changes in ET is essential, given the expected rise in future forestation projects in water-limited regions suffering from increasing forest failure events (Zhang et al 2022, Yu et al 2023).This study will also benefit forest and water resources management by providing spatial-temporal water consumption information.

Land-use and cover dataset
The LUCC data were developed using a top-down model (Yu et al 2018(Yu et al , 2022b) that captures the historical distribution of cropland, forest, and wetland in China spanning the period 1900-2019.The LUCC database, which assimilated land conversion signals from reports, field surveys, and satellite images, was validated with both provincial statistics and remote-sensing products in a spatial framework.Previous comparisons of the LUCC data's forest and cropland maps with other gridded products (e.g.History database of the Global Environment, Land-Use Harmonization 2 Update for the Global Carbon Budget (LUH2-GCB)) revealed that the data are advantageous in delineating land cover changes in China (Yu et al 2022b).

ET data collected from flux towers
The dynamic land ecosystem model (DLEM) was adopted for experimental simulations in this study.The DLEM has been intensively calibrated and validated at various temporal and spatial scales in previous studies (Tian et al 2012, 2018, 2020, Yu et al 2019).Besides, the current study also conducted model parameterization specific for ET using data collected from flux towers (table S1).Specifically, we found that the summarized ET retrieved from ChinaFLUX (www.nesdc.org.cn/) were different from the values reported from the studies of each flux tower.Therefore, we parameterized the model using both data separately, and the averaged simulations were used for analyses in this study.The details of the flux tower stations are listed in table S1.

Comparison of ET products
To quantify the differences of historical ET in China between different studies, we collected and examined ET values derived from both simulation results of multi-model intercomparison projects (i.e.MsTMIP and TRENDY) and published products.There are many global ET products available, while we focus on products that covers the period of 1980s to 2010s for comparison.Therefore, ET from MsTMIP, TRENDY, Wei-ET (Wei et al 2017), MTE-ET (Jung et al 2010), CR-ET a (Ma et al 2019), GLEAM (Martens et al 2017), BESS (Li et al 2023), and P-LSH (Zhang et al 2015) were used.These ET products were produced with different spatial resolutions, ranging from 8 km to 0.25 arc degree (∼25 km) at monthly intervals (table S2).These products were also validated using different flux towers (table S3, figure S1).The atmospheric chemical components, including atmospheric CO 2 concentration and nitrogen deposition data, were retrieved from the IPCC historical CO 2 data, the NACP MsTMIP (https://daac.ornl.gov/NACP),and the study of Jia et al (2018).All the datasets were prepared at or resampled to 0.5 × 0.5 degree for simulations.More details can be found in Yu et al (2022b).

Experimental design and statistical analysis
In this study, we set up simulations to distinguish and quantify the effects of LUCC, climate, CO 2 , and nitrogen deposition on ecosystem ET change in China from 1980 to 2019.First, a spin-up run was implemented to attain the initial state of the land ecosystems in 1900 (Yu et al 2018(Yu et al , 2019)).Then, a 10 year spin-up run was applied before the transient run using initial state information to avoid abrupt changes resulting from mode transition.
Two group of experiments were designed for the transient run to quantify the impacts of each major driver on ET.The first group included one experiment (Allcomb) driven by the historical varying factors over the entire study period from 1900 to 2020.The second group included four experiments designed to keep a specific environmental factor fixed at the 1980 level while varying other drivers during the entire study period.By keeping a particular environmental factor constant at the 1980 level in the second groups of experiments, the impacts of the factor on ET were excluded.Therefore, by comparing the second group of experiments and Allcomb, we were able to quantify the effects of the particular driver on ET for the period of 1980-2020.All the simulations were performed at 0.5 × 0.5 degree.

Validations using flux tower data
In general, ET values collected from published studies were higher than those derived from ChinaFlux (table S4).Results showed that all validations performed better when ET values collected from published studies were used, whereas direct validations using ET derived from ChinaFlux explained only 13%-27% of the variations (figure 1).Our simulated results explained 41% and 81% of the variations of ET from ChinaFlux and published values, respectively (figure 1(a)).
Despite the similar spatial distribution of ET between different products, the trends greatly varied between datasets (figure 3).In general, ET was dominated by increasing trends in GLEAM, P-LSH, and this study, while it was dominated by decreasing in CR-ET a and BESS (figures 3(f)-(j)).The increments of ET were mainly distributed in the northeast, western, and southwest regions of China from this study (figure 3(f)), while the increments were less pronounced in the western and southwest regions but higher in the southeast region for GLEAM and P-LSH (figures 3(h) and (i)).

Attributions of historical ET changes since 1980
Factorial experiments revealed that the largest ET increase was found in the Allcomb simulation at 2.05 mm yr −1 from 1980 to 2020 (P < 0.01), while the increment was only 0.38 mm yr −1 (P < 0.01) in the absence of LUCC since 1980 (figure 4).The inter-annual variations of ET were also observed to be at their lowest if climate was fixed at the 1980 level (figure 4).In comparison, the trends of ET were similar, with a range of 1.43-1.52mm yr −1 (P < 0.01) if the atmospheric CO 2 and nitrogen deposition were fixed at the 1980 level (figure 4).

Discussion
We found that the trends of ET were the highest in all results during the 1980-2020 period.For example, the ET trends were in the range of −0.70-1.47mm yr −1 for TRENDY, MsTMIP, Wei-ET, MTE-ET, CR-ET a , GLEAM, BESS, and P-LSH during 1982-2010, which is much lower than 1.99 mm yr −1 derived from our simulations during the corresponding period (figure 2).We suspect that the differences in ET trends were caused by the underrepresentation of LUCC in simulations (table S5).For example, P-LSH and BESS were produced using a MODIS land cover product (MCD12Q1) (Li et al 2023).However, MCD12Q1 has been criticized to be less capable of capturing land cover changes in China (Yang and Huang 2021).While in comparison, CR-ET a chose to exclude the land cover and soil information for the purpose of large-scale ET mapping (Ma et al 2019).Note that the impacts of underrepresented LUCC signal on simulated ET might be partially compensated for by other auxiliary input, such as leaf area index and normalized difference vegetation index (NDVI).We also observed significant improvements in capturing forest expansion in China with the most recent update of MCD12Q1 (v6.1) (figure S2).We anticipate that incorporating the updated LUCC forcing data may enhance the temporal change signals of existing ET products.To quantify the impacts of LUCC on ET, we set up a simulation to fixed land cover at the 1980 level and found that the ET trend was at 0.38 mm yr −1 from 1982 to 2010, which is in the range of ET trends derived from other ET products.Thus, we are confident that the former ET product underestimated ET increment in China due to underrepresented LUCC.
Cropland abandonment and forest expansion were also found to be underrepresented in China In the current study, we found that the biases in LUCC also resulted in the underestimation of ET trends (0.15-0.21 mm yr −1 since 1980 derived from MsTMIP and TRENDY).More evident, CR-ET a , which was produced without LUCC forcing, was showed decreasing ET (figure 2), implying that ET trends could be opposite if LUCC was ignored.Generally, consistent to former studies (Sun et al 2017b, Bai and Cai 2023), the simulated ET in this study also confirmed that the ET has been increasing since the 1980s.Nonetheless, the reported ET increments in former studies were also considerably lower than our results, which potentially suffered from underestimated LUCC.For example, Sun et al (2017b) reported that ET enhancement was mainly located in the Tibetan Plateau, while the southern and northeast regions were declining.However, we found that ET has been increasing in most of the regions in China since 1980.
We found that LUCC has been the largest contributor to ET increments in China since 1980, followed by climate, rising CO 2 , and nitrogen deposition.Specifically, LUCC contributed to enhanced ET, mainly in the southwest, northern, and eastern regions of China, where it occurred intensively and was dominated by forestation (figures 5(c) and S2).In comparison, climate was the major driver of ET increment in the Tibetan and northwest regions of China, but it reduced ET in southeast China (figure 5(f)).Note that our attributions are different from some former studies.For example, Bai and Cai (2023) reported that vapor pressure deficit (VPD), precipitation, and leaf area index contributed to ET increment in China by 44%, 29%, and 25% during the period of 1982-2019.These results were not directly comparable to our results due to mixed effects in factorial attributions.In addition, the changes of land cover types also matter in affecting ET, while forestation/deforestation does not necessarily lead to increased/decreased ET.For example, Liu et al (2008) found that reforestation decreased ET by 422 mm yr −1 in China, while deforestation increased it by 138 mm yr −1 .This is because most of deforested land was converted to paddy land or irrigated cropland in the early period (e.g. 1900-1960).This shift in land covers also explains the largescale forest reduction in 1900-1980 that accompanied cropland expansion, resulting in a minor increase of annual ET.However, the afforestation/reforestation after 1980 was mostly implemented on noncropland (e.g.grassland, shrubland), which further promoted water consumption and led to ET increments in China.
We also found that the interannual variations of ET were mainly attributed to climate changes, while the ET increment was stimulated by LUCC in China from 1980 to 2020 (figure 4).A previous study reported that the impacts of specific drivers (e.g.CO 2 fertilization, land use change, and nitrogen deposition) on ET are uncertain and challenging to X Wu et al quantify at the global scale, but have strong impacts at local and/or regional scales (Yang et al 2023).This study aimed to address this challenge by utilizing factorial attribution simulations.Based on the analyses, we found that, rising CO 2 and nitrogen deposition were minor factors that contributed to ET enhancement.Mao et al (2015) reported that landuse change has little impact on the trends of global ET but a huge impact on regional-scale ET.Our results confirmed that LUCC is a prominent driver of ET trends in China, where dramatic LUCC, characterized by large-scale forestation, has occurred during the last four decades since 1980.

Conclusions
To sum up, our study reveals that ET has been increasing in China due to LUCC characterized by forest expansion since 1980, with the interannual variations of ET mainly attributed to climate changes.However, existing, long-term ET products suffer from underrepresented LUCC in China, resulting in much lower ET increments.This finding might introduce biases in the water budget and water resources management in China.We advocate for future studies to consider the impact of LUCC in ET simulations.In the future, more accurate, longer-term and wider-space datasets should be selected to improve the accuracy and reliability of the existing, long-term, global ET datasets.To clarify the spatial differences and driving mechanisms of global ET trends, multiple models should be used for evaluation and comparison, along with employing multiple attribution analysis methods for comprehensive analysis.

Figure 1 .
Figure 1.ET validations using flux tower data.Panel (a): this study, panel (b): BESS, panel (c): GLEAM, panel (d): P-LSH, and panel (e): CR-ETa.The blue and red colors indicate validations using ET values collected from published studies and ET derived from ChinaFlux, respectively.

Figure 3 .
Figure 3. Spatial distribution and trends of ET in China derived from different products from 1982-2010.Panels (a)-(e): average ET for the overlapping period of 1982-2010; panels (f)-(j): trends of ET for the period of 1982-2010.Panels a and f were from this study, panels b and g were from BESS, panels c and h were from GLEAM, panels (d) and (i) were from P-LSH, and panels (e) and (j) were from CR-ETa.

Figure 4 .
Figure 4. Temporal changes of ET in China derived from factorial experiments.Black: Allcomb (driven by historical varying factors over the entire study period), red: noLC1980 (LUCC fixed at the 1980 level), purple: noClim1980 (climate fixed at the 1980 level), pink: noCO21980 (CO2 concentration fixed at the 1980 level), and cyan: noNdep1980 (N deposition fixed at the 1980 level).

X
Figure 5. Spatial distribution of ET trends from different (a) trend of annual average ET from Allcomb for the period of 1900-1980, (b)-(f) trend of annual average ET for the period of 1980-2020 from factorial experiments.(b)-(f) were Allcomb, noLC1980, noNdep1980, noCO21980, and noClim1980 experiments, respectively.Details of the abbreviations are the same as in figure 4.
Other model forcing data include climatic, agricultural management, and atmospheric chemical components data.The daily climate data, including air temperature (i.e.maximum, minimum, and average temperature) and precipitation data, were obtained from meteorological stations and published datasets covering the simulation period of 1900-2019 (Yu et al 2022b).Shortwave radiation data were provided by the National Tibetan Plateau Data Center (Tang et al 2019) and the North American Carbon Program (NACP) MsTMIP (Wei et al 2014).Agricultural management data included crop rotation maps obtained from Liu et al (2018), crop-specific nitrogen fertilizer use rates obtained from the Food and Agriculture Organization of the United Nations website (www.fao.org/faostat/) and Li et al (2010), and annual manure applications from Zhang et al (2017).