Estimates of net primary productivity and actual evapotranspiration over the Tibetan Plateau from the Community Land Model version 4.5 with four atmospheric forcing datasets

Abstract

The accuracy of the simulation of carbon and water processes largely relies on the selection of atmospheric forcing datasets when driving land surface models (LSM). Particularly in high-altitude regions, choosing appropriate atmospheric forcing datasets can effectively reduce uncertainties in the LSM simulations. Therefore, this study conducted four offline LSM simulations over the Tibetan Plateau (TP) using the Community Land Model version 4.5 (CLM4.5) driven by four state-of-the-art atmospheric forcing datasets. The performances of CRUNCEP (CLM4.5 model default) and three other reanalysis-based atmospheric forcing datasets (i.e. ITPCAS, GSWP3 and WFDEI) in simulating the net primary productivity (NPP) and actual evapotranspiration (ET) were evaluated based on in situ and gridded reference datasets. Compared with in situ observations, simulated results exhibited determination coefficients (R²) ranging from 0.58 to 0.84 and 0.59 to 0.87 for observed NPP and ET, respectively, among which GSWP3 and ITPCAS showed superior performance. At the plateau level, CRUNCEP-based simulations displayed the largest bias compared with the reference NPP and ET. GSWP3-based simulations demonstrated the best performance when comprehensively considering both the magnitudes and change trends of TP-averaged NPP and ET. The simulated ET increase over the TP during 1982–2010 based on ITPCAS was significantly greater than in the other three simulations and reference ET, suggesting that ITPCAS may not be appropriate for studying long-term ET changes over the TP. These results suggest that GSWP3 is recommended for driving CLM4.5 in conducting long-term carbon and water processes simulations over the TP. This study contributes to enhancing the accuracy of LSM in water–carbon simulations over alpine regions.

摘要

不同大气强迫数据驱动下陆面模式CLM4.5对青藏高原净初级生产力和实际蒸散发模拟性能评估

陆面模式中碳水过程模拟的准确性很大程度依赖于大气驱动数据。因此，选取适当的大气强迫数据可以有效减少陆面模式在高海拔地区水碳过程模拟的不确定性。本研究利用4套常用大气强迫数据(CRUNCEP、ITPCAS、GSWP3和WFDEI)分别驱动陆面模式CLM4.5，开展青藏高原净初级生产力(NPP)和实际蒸散发(ET)模拟试验；结合原位观测等多源参照数据，从多个时空尺度评估4套大气强迫数据在NPP和ET模拟中的表现。相较于原位观测的NPP和ET，4组试验模拟的NPP和ET的决定系数分别在0.58–0.84以及0.59–0.87之间，其中基于GSWP3和ITPCAS驱动的模拟表现更佳。在区域尺度上，基于CRUNCEP驱动的模拟结果相较于参照NPP和ET偏差最大。基于GSWP3驱动的模拟在青藏高原NPP、ET的量级及变化趋势方面都表现最为优异。基于ITPCAS驱动数据模拟的青藏高原ET在1982–2010年间的增加趋势显著高于参照ET和其他3组模拟结果，表明ITPCAS可能并不适用于青藏高原ET的长期变化研究。总体而言，GSWP3大气驱动数据总体表现最优，适合用于驱动CLM4.5开展青藏高原碳水过程的长期模拟研究。本研究结果可为提高陆面模式在高寒区碳水模拟的准确性提供参考。

INTRODUCTION

The Tibetan Plateau (TP) covers an immense area of approximately 2.5 million km² and boasts an average elevation exceeding 4000 m above sea level (asl) (Yang et al. 2023b; Yao et al. 2012). It plays a pivotal role in shaping regional and global ecological and hydrological processes (Jin et al. 2020; Wang et al. 2023b). Due to its unique high-altitude characteristics, the TP exhibits an extraordinary level of sensitivity to climate change and has been warming at a rate significantly higher than the global average (Hu et al. 2019; Sun et al. 2022; Zhou et al. 2023). In recent years, environmental perturbations have far-reaching consequences on the water and carbon cycles, which are fundamental components of the high-altitude ecosystem (Lin et al. 2023; Wang et al. 2020a). A thorough understanding of the spatiotemporal variability and driving factors behind the water and carbon cycles on the TP is essential for forecasting the ways in which future climate change will influence these processes.

Understanding these water and carbon processes necessitates accurate measurements. Observations, particularly those utilizing the eddy covariance method, have been pivotal in quantifying water and carbon fluxes in diverse ecosystems (Hu et al. 2024a, 2024b; Yu et al. 2019). The eddy covariance method involves the direct measurement of the exchange of gases and energy between the land surface and the atmosphere, and it has been widely recognized as the most accurate approach for assessing water and carbon fluxes (Deventer et al. 2021; Hu et al. 2018, 2023). The TP hosts several flux observation sites that provide valuable observed data. However, despite the importance of these observations, challenges still persist. The vast expanse, rugged terrain and harsh climatic conditions of the TP have given rise to data scarcity in many regions (Wang et al. 2023a; Wei et al. 2021). The northwest, in particular, suffers from a pronounced dearth of data (Wang et al. 2023a; Wei et al. 2021). Additionally, the length of data from the eddy covariance method on the TP is relatively short. This temporal limitation hinders our ability to comprehend the impacts of long-term climate change on the water and carbon processes over the TP (Wei et al. 2021).

In response to these challenges, land surface models (LSM) have emerged as a valuable tool (Forzieri et al. 2018; Nkiaka et al. 2017). These models enable high-resolution simulations of eco-hydrological processes, even in regions where observed data are scarce (Feigl et al. 2020; Xi et al. 2022). Many LSM have been successfully applied to simulate water and carbon processes on the TP, offering promising results (Kang et al. 2022; Lin et al. 2019, 2020). However, despite the promise of LSM, disparities in simulation outcomes have been observed. These disparities can be primarily attributed to the various parameterizations and differences in the atmospheric forcing data used (Gelati et al. 2018; Wehrli et al. 2018). Therefore, the choice of appropriate atmospheric forcing data is crucial for LSM to simulate water and carbon processes on the TP effectively.

A myriad of atmospheric forcing datasets are available for driving LSM, and substantial efforts have been made to evaluate their effectiveness (Gelati et al. 2018; Probst and Mauser 2022). However, the performance of these atmospheric forcing datasets varies when applied to different geographical regions (Gelati et al. 2018). This underscores the necessity of evaluating their performance in specific research areas, a task that remains underexplored, especially in data-scarce regions like the TP (Kang et al. 2022). However, current evaluations of atmospheric forcing data largely rely on comparisons with either site-based observations or reanalysis data, and few studies have taken a comprehensive approach by simultaneously comparing and validating atmospheric forcing data using both site-specific and regional datasets. Such an approach can provide a more comprehensive understanding of the suitability of different atmospheric forcing datasets for simulating water and carbon processes in data-scarce regions on the TP.

As one of the most popular LSM, the Community Land Model version 4.5 (CLM4.5) has been effectively utilized in studies focusing on the ecological and hydrological processes of the TP, showcasing its exceptional performance in alpine regions (Lin et al. 2019, 2020). Atmospheric forcing data are a critical component in driving CLM4.5, exerting a profound influence on the precision and reliability of eco-hydrological simulations. Numerous global atmospheric forcing datasets are available, notably CRUNCEP (Climatic Research Unit-National Center for Environmental Prediction), GSWP3 (Global Soil Wetness Project Phase 3) and WFDEI (WATCH-Forcing-Data-ERA-Interim), which are extensively utilized in global and regional land surface modeling owing to their superior spatial resolution and century-scale temporal coverage (Ma et al. 2023). At the regional level, the Institute of Tibetan Plateau Research, Chinese Academy of Sciences also developed a high-precision atmospheric forcing dataset (ITPCAS) for driving land surface process simulations over China (Ma et al. 2019; Yang et al. 2023a). As the model default forcing dataset for CLM4.5, CRUNCEP was more widely used than other forcing datasets in driving CLM4.5 to simulate land surface processes at different spatial scales (Guo et al. 2018; Lin et al. 2016). However, this default forcing dataset has not been systematically evaluated for simulating carbon–water fluxes over the TP. WFDEI has hitherto solely been utilized for investigation pertaining to land surface temperature, with its appropriateness for modeling the carbon–water cycle in TP remaining obscure. When compared with the other two datasets, GSWP3 and ITPCAS have been effectively employed to drive the CLM model in simulating alterations in carbon flux or water flux over the TP (Deng et al. 2020; Lin et al. 2019). Nevertheless, additional examination is necessary to ascertain which of these two datasets is more appropriate for the TP.

Therefore, the present study conducted four types of model simulation experiments over the TP, in which CLM4.5 was driven by four prevalent global and regionally tailored atmospheric forcing datasets. By comparing the CLM4.5 simulated results with site-specific observations, flux data and multiple regional-scale datasets, we will conduct a comprehensive assessment of the performance of different atmospheric forcing datasets in simulating net primary productivity (NPP) and actual evapotranspiration (ET) at various spatial and temporal scales. Through the simulations and comparisons, two questions will be addressed: (i) Is the CRUNCEP (default forcing dataset for CLM4.5) appropriate for forcing CLM4.5 in simulating water and carbon fluxes over the TP? (ii) Which atmospheric forcing datasets give the best NPP and ET simulation according to the comparisons at different spatiotemporal scales. Our study will facilitate reducing uncertainties from the choice of the atmospheric forcing datasets and improve the accuracy of LSM in simulating water and carbon processes in high-altitude regions.

MATERIALS AND METHODS

Study area

The TP is situated in the southwest region of China (Fig. 1), with around 75% of its area characterized by elevations exceeding 4000 m (Wang et al. 2020a). It can be categorized into five climate sub-regions, including arid region, semi-arid region, semi-humid region, humid/sub-humid region and humid region as delineated by Zhang et al. (2021). Alpine grasslands dominate the landscape of the TP, while forests are primarily concentrated in the southern and southeastern regions, with patches of barren or sparsely vegetated lands prevalent in the northwest (Ma et al. 2020).

Atmospheric forcing data

Four atmospheric forcing datasets were selected for our study to drive CLM4.5 simulations over the TP. These datasets draw from diverse data sources, including ground-based in situ measurements, satellite-based remote sensing retrievals and reanalysis datasets. They encompass seven key variables: near-surface air temperature, near-surface atmospheric pressure, near-surface specific humidity, downward shortwave radiation, downward longwave radiation, precipitation rate and near-surface wind speed. The detailed descriptions of these four forcing datasets are presented below.

ITPCAS

Addressing the lack of high-resolution near-surface reanalysis data for China’s region and for driving land surface process simulations, the Hydro-Meteorological Research Group of the Institute of Tibetan Plateau Research, Chinese Academy of Sciences, developed a high-precision (spatial resolution: 0.1° × 0.1°; temporal resolution: 3 h) atmospheric forcing dataset, referred to as ITPCAS (Chen et al. 2011; Yang et al. 2010). Covering the land region of China from 1979 to 2018, the ITPCAS dataset integrates near-surface air temperature, wind speed, atmospheric pressure and specific humidity from Princeton meteorological forcing data and observational data from the China Meteorological Administration. Precipitation is obtained by integrating various precipitation products, including data from the China Meteorological Administration and the Tropical Rainfall Measuring Mission. Downward shortwave radiation comes from the calibrated Global Energy and Water Exchanges–Surface Radiation Budget shortwave radiation product. The reliability of this dataset has been confirmed, and it has been widely used for land surface process simulations, land surface parameter retrievals and data assimilation (Zheng et al. 2016). Moreover, its reliability over the TP has been verified in studies by Xue et al. (2013) and Guo et al. (2013), who utilized ITPCAS atmospheric forcing data to drive distributed hydrological models and investigate permafrost distribution on the TP.

GSWP3

To facilitate research on energy–water–carbon cycling and their interactions, as well as for model evaluation purposes, the Global Soil Wetness Project Phase 3 (GSWP3) developed a century-scale climate forcing dataset (Kim 2017). GSWP3 atmospheric forcing data features a spatial resolution of 0.5° × 0.5° and a temporal resolution of 3 h, encompassing both land and ocean and spanning from 1901 to 2014. Precipitation data in GSWP3 is subjected to a ‘two-step’ bias correction, initially using the Climate Prediction Center’s CPC-Unified precipitation product for daily-scale correction and subsequently applying the Global Precipitation Climatology Centre’s precipitation product for (GPCC) long-term monthly scale correction. Additionally, bias correction was performed on other variables using the Climate Research Unit dataset (CRU) and Radiation Budget Measurement radiation products. This series of corrections enhances the reliability of GSWP3 atmospheric forcing data for high-resolution meteorological forcing over a century-scale timeframe. The dataset has found wide applications in meteorological and hydrological research at global and regional scales (Bonan et al. 2019; Tangdamrongsub et al. 2018). GSWP3 has been employed for driving LSM to simulate changes in land surface processes with respect to the TP (Lin et al. 2020; Wang et al. 2020b).

CRUNCEP

CRUNCEP atmospheric forcing data are the default dataset for the CLM4.5. It combines monthly observational data (CRU TS) with a spatial resolution of 0.5° × 0.5° (Mitchell and Jones 2005; Sun 2017) and 6-h reanalysis data from NCEP/NCAR with a spatial resolution of 2.5° × 2.5° (Fan et al. 2016; Gao 2017; Kalnay et al. 1996). In CRUNCEP, air temperature, precipitation, specific humidity and shortwave radiation exhibit the same diurnal variations as NCEP/NCAR reanalysis data but are adjusted to monthly means using CRU TS climatology data. Other variables, such as near-surface atmospheric pressure, downward longwave radiation and wind speed, are obtained directly through interpolation from NCEP/NCAR reanalysis data. In this study, we used the 7th version of CRUNCEP data, with a spatial resolution of 0.5° × 0.5° and a temporal resolution of 6 h, spanning from 1901 to 2016. CRUNCEP atmospheric forcing data have been employed in various studies involving CLM model simulations, including those related to vegetation productivity and ET (Mao et al. 2012; Shi et al. 2013). For the TP region, past investigations have employed the CRUNCEP dataset to drive CLM4.5 in the simulation and analysis of soil moisture and soil temperature (Guo et al. 2017; Yuan et al. 2021).

WFDEI

Weedon et al. (2014) developed a global 0.5° × 0.5° climate forcing dataset referred to as WFDEI. The WEDEI dataset has been derived through the application of a consistent methodology employed in the WATCH Forcing Data, utilizing the ERA-Interim reanalysis dataset. In the WFDEI, variables such as air temperature, near-surface atmospheric pressure, specific humidity and downward longwave radiation flux are elevation corrected. Precipitation was corrected based on gridded observations, including CRU TS and GPCC datasets. WFDEI data have been proven suitable for driving hydrological model and LSM on a global scale, especially for simulating hydrological impacts in large river basins. Regarding the TP region, the WFDEI forcing dataset has been utilized for driving LSM to conduct land surface temperature modeling (Ma et al. 2023).

Validation data

Gridded NPP and ET products for comparison

The dataset used to evaluate the spatial and temporal pattern of simulated annual NPP was obtained from the National Earth System Science Data Center of China (http://www.geodata.cn). This dataset is derived from the Integrated Biosphere Simulator (IBIS) model, CRU meteorological data and satellite observations. It covers the time period from 1950 to 2012 and has been verified on global and regional scales (Zhang et al. 2013). The Coupled Model Intercomparison Project Phase 5 (CMIP5) provided a platform for model comparison, which can be accessed at https://esgf-node.llnl.gov/search/cmip5/. To evaluate the seasonal cycle of NPP, we gathered and created a multi-model ensemble (MME) of monthly NPP results. This was done by combining the results of five model simulations (CCSM4, MRI-ESM1, NorESM1, CanESM2 and MPI-ESM-P) from the CMIP5 (CMIP5 MME). The CMIP5 MME NPP data covers the time period from 1850 to 2005.

Due to the lack of a benchmark global ET dataset, Mueller et al. (2013) developed a comprehensive monthly scale global ET dataset (LandFlux-EVAL) by merging multiple global land surface ET products, covering two time periods: 1989–1995 and 1989–2005. In this study, we used the spatial resolution of 1° from the 1989–2005 dataset, which integrates ET products from various methods, including remote sensing estimation, land surface modeling and reanalysis. This dataset has been shown to agree well with precipitation minus runoff measurements in large river basins worldwide (Mueller et al. 2011) and has been used to evaluate CMIP5 ET simulation results (Mueller and Seneviratne 2014) as well as for validation of ET estimates at global and regional scales (Jung et al. 2019; Wartenburger et al. 2018). Moreover, this dataset has been applied as a benchmark for ET assessment in the TP region (Peng et al. 2016; Wang et al. 2020a). Additionally, ET products from the Global Land Evaporation Amsterdam Model (GLEAM) and Global Land Surface Satellite (GLASS) were introduced to validate the long-term ET trends over the TP (Martens et al. 2017; Yao et al. 2014).

Observed data for validation

We evaluated the CLM4.5 simulated NPP results for the growing season (i.e. May–September) using observed NPP data collected from 79 sites as reported by (Lin et al. 2019). Since the simulation period of the CLM4.5 comparative experiment driven by four different atmospheric forcing datasets is from 1979 to 2010, we selected the ET flux data from nine sites, Haibei, Dangxiong, Tanggula, Arou, SETORS, NADORS, NAMORS, Maqu and Suli, which meet the temporal requirements, for validation. The ET flux data for Haibei, Dangxiong, Arou and Maqu were obtained from the China Flux Observation Network (http://www.chinaflux.org/), the ET flux data for SETORS, NADORS and NAMORS were obtained from Ma et al. (2020), while data for Tanggula and Suli were obtained from Chang et al. (2017). Details of these station locations are provided in Fig. 1 and Table 1. Observed precipitation was also introduced to validate the long-term precipitation change over the TP from different atmospheric forcing datasets. Monthly observed precipitation for the period of 1979–2010 was obtained from National Meteorological Information Center of China Meteorological Administration (Shen et al. 2010).

Experimental design

This study designed four sets of experiments. The first set of experiments used GSWP3 climate forcing data to drive the CLM4.5 (CLM_G). The second set of experiments used CRUNCEP atmospheric forcing data to drive the model (CLM_C). The third and fourth sets of experiments used ITPCAS and WFDEI climate forcing data to drive the model, referred to as CLM_I and CLM_W, respectively. All other data, except for the atmospheric forcing data, were kept consistent across the four sets of experiments. To facilitate comparative analysis, the spatial resolution of the four sets of atmospheric forcing data was interpolated to 0.1°. The study area covered the TP with a spatial resolution of 0.1°. The integration time step was 30 min, and the simulation period extended from January 1979 to December 2010.

Model evaluation metrics and data analysis methods

To evaluate the model performance of the simulated NPP and ET, three metrics were used including coefficient of determination (R²), root-mean-square error (RMSE) and the mean bias. The least-square regression method was employed to identify the long-term linear annual trends in the variables (Han et al. 2021). The statistical significance of the linear trends was assessed using an F test, and only tests with a P value below 0.05 were considered as passing the significance test (Lin et al. 2019).

RESULTS

Performance of NPP simulation with different atmospheric forcing data

Fig. 2 presents the comparison results of NPP simulated by the four experimental sets with observed NPP at observed sites. The results showed that the average value of observed NPP was 127.5 g C m⁻² growing season⁻¹, while the simulated mean NPP values for the four experimental sets ranged from 84.4 to 136 g C m⁻² growing season⁻¹. Among them, CLM_C and CLM_W simulations underestimated NPP with a bias of −43.1 and −9.1 g C m⁻² growing season⁻¹, respectively, while NPP simulated by CLM_I and CLM_G is slightly higher than observed values, with a bias of 3.7 and 8.5 g C m⁻² growing season⁻¹, respectively. CLM_I NPP had the highest R² with observed NPP (R² = 0.84), followed by CLM_G NPP (R² = 0.78), CLM_C NPP (R² = 0.65) and CLM_W NPP (R² = 0.58). The RMSE was 34.9, 41.6, 57.6 and 67.2 g C m⁻² growing season⁻¹ for the CLM_I, CLM_G, CLM_W and CLM_C simulation, respectively. Overall, taking the R², RMSE and bias values for the four simulations into consideration, the CLM4.5 model presented well performance in carbon flux simulations over the TP, while the choice of atmospheric forcing data had a significant impact on the accuracy of the model simulations. In this regard, ITPCAS and GSWP3 atmospheric forcing data performed well in NPP simulations with larger R² and lower RMSE and bias values, while CRUNCEP and WFDEI climate forcing data showed relatively poorer performance.

IBIS NPP increased from northwest to southeast over the TP. The lowest NPP values were found in the arid region in the northern part of the TP, while high NPP values were mainly distributed in the humid and humid/sub-humid regions of the TP. The spatial distribution of annual NPP for the four experimental sets was generally consistent with IBIS NPP (Fig. 3). For the entire TP, IBIS NPP had an average annual value of 195 g C m⁻² yr⁻¹. The NPP values simulated by the four experimental sets showed large divergences (120–162 g C m⁻² yr⁻¹) and were all had negative bias values compared with the IBIS NPP, with CLM_G had the lowest bias value (Supplementary Table S1). CLM_G also had the lowest RMSE value (127 g C m⁻² yr⁻¹) compared with the other three simulations (129–153 g C m⁻² yr⁻¹). Compared with the IBIS product, the four simulations performed better in the semi-arid, semi-humid and humid/sub-humid areas (Supplementary Table S1). CLM_W showed better performance than the other three simulations, with relatively higher R² and lower RMSE in these regions. The large difference between IBIS and CLM4.5 modeled NPP was found in humid and arid regions. CLM_I was found showed the worst performance compared with IBIS in these regions.

During 1979–2010, IBIS NPP showed an increasing trend in most parts of the TP, with the most significant increases occurring in the southern part of the plateau and decreases in the northern and western regions (Fig. 4). Similarly, all four experimental sets showed significant NPP increases in the southern part of the TP. For CLM_G simulation, NPP increases prevail across the entire TP, with the most significant increase occurring in the southern humid regions. NPP decreases were observed in some central and northern regions. The spatial distribution of CLM_G NPP trends was closest to that of IBIS NPP compared with the other three simulations. In the CLM_C experiment, NPP increases were significant in the southern part of the TP, while decreases were observed in the eastern, northern and western regions. In CLM_I, notable NPP increases were found in the southeast, southern and central-western regions of the TP, with significant decreases in some southern and western regions. In CLM_W, the most significant NPP increases were observed in the southeastern part of the TP, with smaller increases compared with the other experiments, and NPP decreases were mainly observed in the eastern and northeastern regions.

Interannual variations of NPP for the four experimental sets compared with IBIS NPP are shown in Fig. 5. All four simulations exhibit high R² with IBIS NPP (0.67–0.79). IBIS NPP exhibited a significant increase at a rate of 1.34 g C m⁻² yr⁻², while the NPP change rates in the CLM simulations varied between 0.67 and 1.1 g C m⁻² yr⁻². CLM_I NPP had the highest increase rate (1.1 g C m⁻² yr⁻²), followed by CLM_G NPP (0.89 g C m⁻² yr⁻²), while CLM_C NPP and CLM_W NPP had similar increase rates of 0.67 and 0.74 g C m⁻² yr⁻², respectively.

The monthly changes of NPP simulated by CLM4.5 compared with CMIP5 MME NPP are relatively consistent (Fig. 6). CMIP5 MME NPP peaked in July–August (38 g C m⁻² month⁻¹). All four simulated NPP peaked in July, with values ranging from 25 to 30 g C m⁻² month⁻¹. CLM_G NPP had the highest peak values, while CLM_C NPP had the lowest peak values. CMIP5 MME NPP and the four simulated NPP values all reached their minimum values in January. Compared with the NPP variations from CLM_I and CLM_C, the seasonal variations from CLM_G and CLM_W were closer to the CMIP5 MME NPP. CMIP5 MME NPP had a monthly value of 15.8 g C m⁻² month⁻¹. Among the four simulations, CLM_G NPP had the highest monthly value, slightly below CMIP5 MME NPP, while CLM_C NPP has the lowest monthly value (9.8 g C m⁻² month⁻¹). The negative bias from all four simulations showed underestimations.

Performance of ET simulation with different atmospheric forcing data

The comparison with observed ET at flux sites showed that the CLM4.5 model can capture the characteristics of ET in the TP, but the choice of atmospheric forcing data strongly influenced the accuracy of the ET simulations (Fig. 7). It can be found that the model performance varied among the four simulations with statistical indices values ranging from 0.59 to 0.87 for R², and from 14.5 to 30 mm for RMSE. Four simulations underestimated ET compared with observed ET, with bias ranging from −7.2 to −19.7 mm. Both CLM_G and CLM_I exhibited noteworthy performance, displaying similar values for R², RMSE and bias. Moreover, better performance was observed in the CLM_G and CLM_I simulations, exhibiting the highest R² and lowest RMSE values when compared with observed ET. CLM_W exhibited the lowest skill, with an R² value of 0.59 and an RMSE value of 30 mm, indicative of its inferior performance compared with other simulations. This indicated that GSWP3 and ITPCAS atmospheric forcing provided the closest match to observed ET at the site scales, while WFDEI atmospheric forcing data performed the poorest.

The spatial distribution pattern of annual ET for the four experimental sets was generally consistent with EVAL ET, which increased from northwest to southeast over the TP (Fig. 8). The simulated annual ET over the TP was 309, 317, 332 and 317 mm for the simulations based on GSWP3, ITPCAS, CRUNCEP and WFDEI, respectively. Among the four simulations, the annual ET over the TP from the GSWP3-based simulation was the closest to the EVAL product. At the plateau level, the R² between the four simulated annual ET and EVAL ET ranged from 0.43 to 0.81. The largest R² with EVAL ET was obtained by CLM_G simulation, while the lowest R² was obtained by CLM_W simulation. CLM_G demonstrated the lowest RMSE value of 95 mm with EVAL ET in comparison to the other three simulations. These results indicated that in characterizing the spatial distribution of ET over the TP, GSWP3 forcing performed the best, while WFDEI atmospheric forcing provided a rougher characterization of the spatial distribution of ET. Compared with EVAL ET, CLM_G exhibited a more pronounced negative bias in arid regions, while demonstrated a positive bias in other climate zones (Supplementary Table S2). WFDEI-based ET simulation closely aligned with EVAL ET in arid and semi-humid regions. The largest difference in annual ET values among the four simulations was observed in humid regions, with a range from 499 mm in WFDEI-based simulation to 965 mm in CRUNCEP-based simulation.

During 1989–2005, the spatial distribution of trends in EVAL ET showed increasing trends in most parts of the TP, with the largest increases in the eastern, southern and some western regions (Fig. 9). However, the patterns of annual ET trends were highly heterogeneous among the four simulations. In the CLM_G simulation, significant increases in ET were observed in the western and southeastern parts of the TP, while decreases were observed in the northeastern and southwestern regions. The spatial distribution of CLM_C and CLM_W simulated ET trends was relatively similar, with significant increases in the northeastern, central and western parts of the TP and decreased in the southeastern and northwestern regions. In the CLM_I simulation, ET increased significantly in the western and northern regions of the TP, with increases exceeding 6 mm yr⁻¹ in most areas, while decreases were observed in the southeastern, southwestern and some central regions.

When compared with EVAL ET, the ET interannual variability was well captured by all four experimental sets (Fig. 10), with R² ranging from 0.69 to 0.81. The R² between the CLM_G simulated ET and the EVAL ET was still the highest among the four simulations, indicating that GSWP3 forcing was the most suitable for characterizing the interannual variation of ET over the TP. The annual EVAL ET showed a significant increasing trend during the period from 1989 to 2005, with a rate of 1.29 mm yr⁻¹. All four simulations also showed increasing trends during 1989–2005, but with varying magnitudes. CLM_G, CLM_C and CLM_W simulated ET trends were similar to EVAL, with rates of 1.54, 1.24 and 1.71 mm yr⁻¹, respectively. The noteworthy point was that the increase rate of ET for the CLM_I simulation was significantly higher than that of the EVAL, with a growth rate of 6.5 mm yr⁻¹. This was mainly caused by a significant increase in precipitation in the ITPCAS atmospheric forcing data from 1989 to 2005.

The monthly changes of ET simulated by CLM4.5 compared with EVAL were generally consistent (Fig. 11). The peak values of CLM4.5 simulated ET and EVAL ET all occur in July, similar to the NPP patterns. CLM_G, CLM_C and CLM_I simulations all had peak values exceeding 57 mm, while CLM_W had a peak value of 50 mm. The monthly average ET of EVAL was 24.8 mm, and the monthly average values for the four simulations ranged from 25.7 to 27.7 mm, with CLM_G being the closest to EVAL (25.7 mm).

DISCUSSION

CLM4.5 default forcing dataset is not a good choice for simulating NPP and ET over the TP

Atmospheric forcing data are a critical component in driving LSM, exerting a profound influence on the precision and reliability of simulations within the domains of ecohydrology and climatology (Gavahi et al. 2022; Humphrey et al. 2021). This comprehensive utilization of atmospheric forcing data allowed us to gain insights into the intricate interactions between the land surface and the atmosphere. The CRUNCEP atmospheric forcing dataset was designed to drive the CLM over a long period and served as the default global forcing dataset for CLM4.5. It has been utilized to drive CLM4.5 in investigations of vegetation growth, surface energy budget, as well as for the TRENDY (trends in net land-atmosphere carbon exchange over the period 1980–2010) project, along with numerous other applications. Some research has evaluated the dependability of the CRUNCEP forcing dataset at global or regional scales. For example, Guo et al. (2018) simulated changes in the near-surface soil freeze/thaw cycle in Norther Hemisphere using CLM4.5 driven with four atmospheric forcing datasets and suggested the CRUNCEP dataset can be preferentially considered as the basic atmospheric forcing dataset for future prediction. However, our findings suggested that the utilization of the CLM4.5 default atmospheric forcing dataset CRUNCEP for simulating the carbon and water fluxes over the TP yielded unsatisfactory outcomes.

Specially, compared with the observed NPP at sites scale, it was found that CRUNCEP-based NPP simulation performed the worst among the four forcing datasets, with a relatively smaller R² (0.65) and the largest RMSE (67.2 g C m⁻² growing season⁻¹), and the most significant underestimation of NPP values (Fig. 2). At the plateau level, the reference NPP dataset IBIS had an average annual value of 0.48 Pg C yr⁻¹, consistent with the results from Gao et al. (2013) who showed the annual NPP over the TP was 0.47 Pg C yr⁻¹ based on a light use efficiency model. However, the simulated annual NPP over the TP based on the CRUNCEP forcing dataset was only 0.30 Pg C yr⁻¹, which showed the largest bias among the four forcing products. CRUNCEP-based NPP also showed the poorest performance in producing the long-term change trend over the TP, showing twice lower than IBIS NPP. According to the validation results at the ET sites scale, the CRUNCEP forcing dataset continued to exhibit inferior performance when compared with GSWP3 and ITPCAS, with CRUNCEP consistently underestimating ET. The annual ET over the TP from the CRUNCEP-based simulation was 332 mm, which showed the largest bias compared with the reference ET product of EVAL (298 mm). It was found that CRUNCEP severely overestimated ET in humid areas, resulting in higher TP-averaged ET values when compared with the other three simulations and EVAL. Consistently, our findings indicated that CRUNCEP may not be the suitable option for accurate simulating TP NPP and ET regarding magnitudes, spatial patterns and long-term change trends.

Can we obtain the optimal forcing dataset for conducting the carbon and water simulation over the TP?

Observational data from various sites across the TP are crucial for validating the accuracy of LSM simulations. By comparing with observed data at these sites, it was discovered that GSWP3 and ITPCAS exhibited better performance than the other two atmospheric forcing datasets (Figs 2 and 7). However, site-level validation did not conclusively show that either ITPCAS or GSWP3 significantly outperformed in both NPP and ET simulations. For instance, while the simulated NPP based on the ITPCAS forcing dataset was closer to the observed NPP than that driven by GSWP3, the simulated ET driven by GSWP3 and ITPCAS exhibited very similar performance when compared with observed ET.

At the plateau level, the GSWP3-based simulation had the highest TP-averaged NPP value among the four atmospheric forcing datasets, which closely matched the TP-averaged NPP according to the reference IBIS dataset and previous studies regarding NPP over the TP (Gao et al. 2013; Guo et al. 2020; Luo et al. 2018). In terms of NPP magnitudes, the GSWP3-based simulation seemed to exhibit greater accuracy compared with the ITPCAS-based simulation. Nonetheless, both datasets demonstrated similar performance in replicating the long-term trends in NPP over the TP.

Different choices in estimated models and input-driven data can lead to discrepancies in ET estimation, even within the same study area (Lin et al. 2021). For example, the annual ET over the TP was 496 mm reported by Han et al. (2021), whereas Wang et al. (2020a) estimated the ET over the TP based on a revised generalized nonlinear-CR model, and showed the TP-averaged annual ET was approximately 295 mm, which was close to the EVAL reference ET dataset we used (298 mm). Chang et al. (2023) also suggested that the annual ET over the TP was 300 mm based on analyzing the GLEAM product. The CLM4.5 simulated ET over the TP ranged from 309 to 332 mm among the four simulations, indicating that input-driven data could lead to large discrepancies in estimating ET.

Comparing observed ET at the sites scale and ET magnitude at the plateau level, it appeared that running CLM4.5 with either GSWP3 or ITPCAS could effectively simulate ET over the TP, with no significant difference between them. However, although GSWP3-based and ITPCAS-based ET simulations showed similar ET values, there were notable differences in their long-term ET trends. During 1989–2005, the GSWP3-based simulated ET over the TP showed a significant increasing trend with a rate of 1.54 mm yr⁻¹, which was a little higher than the reference EVAL ET (1.29 mm yr⁻¹). However, the ITPCAS-based simulated ET over the TP has an increasing trend of 6.5 mm yr⁻¹, which was much higher than EVAL ET. For the convenience of comparing with more ET products over a longer time scale, we further calculated the TP ET trends during 1982–2010. The CLM4.5 simulated ET over the TP driven by ITPCAS experienced a notable upward trend of 3.91 mm yr⁻¹ from 1982 to 2010, which was roughly 3.5 times greater than the rate observed in the GSWP3-based simulated ET (1.13 mm yr⁻¹). The ET trend for TP during 1982–2010 derived from ITPCAS-based simulation was also 2.5–5 times larger compared with the ET trends from GLASS ET (0.77 mm yr⁻¹) and GLEAM ET product (1.62 mm yr⁻¹). Therefore, simulations of ET using GSWP3 could yield results that are more consistent with other widely used products in terms of temporal trends.

It is essential to carefully consider the precipitation variable and its trend in the atmospheric driving data when employing the CLM4.5 model to simulate TP ET (Lin et al. 2020). In the ITPCAS dataset, the precipitation exhibited a rising trend at a rate of 4.1 mm yr⁻¹ during 1979–2010, which was notably higher than the precipitation in GSWP3 (1.02 mm yr⁻¹) and the observed precipitation (1.01 mm yr⁻¹). Considering that the ET simulation in CLM4.5 was particularly sensitive to changes in precipitation, we believe that GSWP3 was a more suitable option than ITPCAS for long-term ET simulation in the TP region. Ma and Zhang (2022) also confirmed that ET increasing over the TP was mainly attributed to enhanced precipitation during 1982–2016, in which the ET simulation was based on a water–carbon coupled biophysical model mostly driving by ITPCAS forcing dataset. They showed the ET trend over the TP was 1.87 mm yr⁻¹, which was 2.2 times greater than the ET trend (0.84 mm yr⁻¹) calculated from ensemble ET data in Chen et al. (2024). Their higher ET trends may be also associated with the usage of ITPCAS precipitation data, even after integrating two additional precipitation products to mitigate the uncertainty of precipitation.

Ecosystems do not operate in isolation but instead respond to a multitude of environmental drivers (Robaina-Estevez and Nikoloski 2020). The multifaceted interactions among these variables give rise to the intricacies of ecological responses. While a specific atmospheric forcing dataset may excel in simulating water and carbon processes in a particular region, it did not guarantee that individual variables within that dataset can be faithfully represented or simulated. Previous studies found that the combined influence of temperature, precipitation and solar radiation significantly affects vegetation growth and carbon cycling (Sun et al. 2021; Zhang et al. 2022). Magnan et al. (2022) underscored the importance of considering the collective performance of meteorological variables in eco-hydrological modeling. Therefore, it is imperative to recognize that, when applying atmospheric forcing data to drive LSM, the effectiveness of the entire dataset takes precedence over the performance of individual variables.

Implications

In our study, the disparities in model results for NPP and ET from four simulations were solely caused by the combined effects of all forcing variables between different atmospheric forcing datasets, highlighting the influence of uncertainties in forcing datasets on CLM4.5 simulations. The model simulations based on CRUNCEP (CLM4.5 default forcing dataset) performed worse than simulations from GSWP3 and ITPCAS. Based on validation outcomes across various spatiotemporal scales, GSWP3 emerges as the most appropriate dataset among the four forcing datasets for driving CLM4.5 in examining the water–carbon coupling process over the TP. It is imperative to emphasize that the aforementioned conclusions were drawn based on CLM4.5, and further investigations utilizing other LSMs are indispensable to validate its robustness. Previous studies indicated that LSM accuracy depended not only on the atmospheric forcing used but also on the model itself, including its structure and parameterization (Bonan et al. 2019). While the best-performing simulated results from CLM4.5 do exhibit some biases compared with observed NPP and ET at the sites scale, discrepancies in spatial scale between modeling products and in situ observations, along with uncertainties in atmospheric forcing datasets, contribute to these biases. Additionally, model deficiencies play a role. In the forthcoming period, refining CLM model parameters with substantial observational data sourced from the TP, along with evaluating multiple sets of atmospheric forcing datasets to identify optimal sets for merging, is expected to significantly improve the simulation performance of the CLM model tailored specifically for the TP region.

CONCLUSIONS

In order to improve the accuracy of CLM4.5 simulations in the TP region, this study evaluated four commonly used atmospheric forcing datasets. Through validation with observational station data and comparative analysis with regional-scale NPP and ET products, it was found that CLM4.5 model default atmospheric forcing dataset CRUNCEP was not suitable for accurately driving CLM4.5 in simulating TP NPP and ET due to large uncertainties in magnitudes, spatial patterns and long-term change trends. Comparing with observed data at various sites across the TP, it was found that GSWP3 and ITPCAS exhibited better performance than the other two atmospheric forcing datasets. However, site-level validation did not conclusively show that ITPCAS or GSWP3 significantly outperformed in both NPP and ET simulations. In terms of TP-averaged NPP and ET magnitudes, GSWP3-based simulation appeared to be more accurate compared with ITPCAS-based simulations. Concerning temporal trends, ITPCAS-based simulated ET showed very large increasing trends compared with the other three CLM4.5 simulations and widely used ET products, primarily due to notably higher precipitation trends in the ITPCAS dataset. Given the validation outcomes across various spatiotemporal scales, the GSWP3 atmospheric forcing dataset stands out as the most appropriate dataset among the four forcing datasets for driving CLM4.5 in examining the water–carbon coupling process over the TP.

Supplementary Material

Supplementary material is available at Journal of Plant Ecology online.

Table S1: Statistical metrics of annual NPP between four simulations and IBIS NPP over the TP and its five sub-regions.

Table S2: Statistical metrics of annual ET between four simulations and EVAL ET over the TP and its five sub-regions.

Funding

This work was supported by the National Key Research and Development Program of China (2022YFC3201702), the National Natural Science Foundation of China (42201146, U2240226), the Science and Technology Project of Sichuan Province (2022NSFSC1001) and Fundamental Research Funds for The Central Universities (YJ2021133).

Data Availability

The GSWP3 and CRUNCEP atmospheric forcing dataset is available online at https://svn-ccsm-inputdata.cgd.ucar.edu/trunk/inputdata/atm/datm7/. The WFDEI dataset is available online at https://rda.ucar.edu/datasets/ds314.2/. The ITPCAS atmospheric forcing dataset is available online at https://data.tpdc.ac.cn/zh-hans/data/. The CMIP5 NPP dataset is available online at https://esgf-node.llnl.gov/search/cmip5/. The EVAL ET is available online at https://data.iac.ethz.ch/landflux/. The IBIS NPP dataset is available online at http://www.geodata.cn. The ET flux data are available online at http://www.chinaflux.org/ and http://data.tpdc.ac.cn.

Acknowledgements

We would like to thank the data support from ‘National Earth System Science Data Center, National Science & Technology Infrastructure of China (http://www.geodata.cn)’ and ‘National Tibetan Plateau/Third Pole Environment Data Center (http://data.tpdc.ac.cn)’. In addition, the authors appreciated the constructive comments from anonymous reviewers to improve the quality of this article.

Conflict of interest statement. The authors declare that they have no conflict of interest.

REFERENCES