Evaluating Cumulative Drought Effect on Global Vegetation Photosynthesis Using Numerous GPP Products

The increasing trend in drought events under the background of global warming makes it more important to understand the drought effect on vegetation photosynthesis. While diverse global gross primary production (GPP) datasets were adopted to investigate the drought impact on photosynthesis, few studies focused on the discrepancies of drought response among different GPP datasets, especially for the cumulative drought impact. Therefore, a total of twenty-six global GPP datasets based on process, machine learning (ML), and light-use efficiency (LUE) model schemes were obtained to appraise the cumulative impact of drought stress on photosynthesis from 2001 to 2010. Moreover, a relatively reliable global pattern of drought’s cumulative effect on vegetation photosynthesis was acquired from these global GPP products through probability analysis. The results illustrated that the cumulative impact of drought existed in 52.11% of vegetation cover land with the cumulative time scales dominantly at a short term (1–4 months, 31.81%). Obvious heterogeneity of the drought cumulative effect in space and different vegetation functional types was observed, as the reliability of the drought effect decreased with latitude decreasing and a higher sensitivity to drought in herbaceous vegetation than woody plants. Our findings highlighted the importance of ways in characterizing moisture conditions across vegetation types among various GPP models and the necessity of GPP dataset selection in investigating drought effect on photosynthesis.


INTRODUCTION
Global climate change will cause significant effects on plant photosynthesis (Cramer et al., 2001;Richardson et al., 2013;Chen et al., 2014), especially with an increasing trend of infrequently severe drought events under the background of global warming (AghaKouchak et al., 2014;Naumann et al., 2018). In this way, it is essential to understand the various effects of drought on vegetation status and further investigate the terrestrial ecosystem carbon cycle (van der Molen et al., 2011;Barman et al., 2014;Anderegg et al., 2015).
Drought, as a complex and intermittent disturbance among climatic phenomena, can influence the traits of the terrestrial ecosystem and vegetation status (van der Molen et al., 2011;Reichstein et al., 2013;Anderegg et al., 2015), such as reducing the expansion of foliage and stomatal conductance (Passioura, 1991), inducing plant mortality (Huang et al., 2010;Rao et al., 2019), or even causing biotic disturbance and wildfire (Wendler et al., 2011;Huang et al., 2017). As the observable changes in plant status induced by drought are hard to be timely detected, it is difficult to understand the reactions of the plant to drought . Moreover, except for the current moisture conditions, early drought events may also control vegetation growth (Peng et al., 2019;Yuan et al., 2020). In this way, knowing the various time-scales of plant reactions to the drought effect is critical for understanding the interactions between climate and plants (Zhao et al., 2018;Peng et al., 2019;Wen et al., 2019). The drought effect on plants can be summarized in the cumulative and time-lag responses primarily . Time-lag impacts illustrate the effect of early drought events at a specific time on the current plant status, and cumulative impacts are associated with lagged effects and climatic dynamics during a given period . While there were an increasing number of studies that surveyed the time-lag impact of drought on the plant, it was found that it varies across species and ranges (Braswell et al., 1997;Huang et al., 2018;Peng et al., 2019). Some cumulative events (e.g., cumulative water deficits) could intensify drought effects on terrestrial ecosystems (Ivits et al., 2016). Therefore, it is necessary to assess the cumulative drought impact on global plant photosynthesis more robustly.
In general, the response sensitivity of vegetation photosynthesis (i.e., GPP) to dry conditions at the ecosystem scale was used to characterize drought effect on vegetation (Ciais et al., 2005;Sun et al., 2021). With the development of satellite remote sensing technology (Xiao et al., 2019;Chu et al., 2021;Pan et al., 2021;Guo et al., 2022), massive remotely sensed images provided unique opportunities for researching the ecosystem carbon cycle at multiple scales (e.g., regional, continental, or global). On this basis, many researchers have established a variety of models driven by remote sensing data to simulate GPP over the past few decades (Tan et al., 2012;Anav et al., 2015;Lees et al., 2018;Xie and Li, 2020b;a). Generally, all these remotely sensed data-driven GPP models could be divided into three parts based on different schemes, namely, process, machine learning (ML), and light-use efficiency (LUE) models. The process-oriented models always comprehensively considered the effects of several main biophysical and chemical processes on vegetation photosynthesis over the terrestrial ecosystem (Ito, 2010;Hayes et al., 2011;Tian et al., 2011;Zhu et al., 2014;Jiang and Ryu, 2016). In this way, GPP estimations obtained from them were more mechanistic and rigorous, while the ML models established complex nonlinear statistical relations between physiologically relevant input data and GPP, thus needing huge amounts of data for model training (Tramontana et al., 2016;Zheng et al., 2020). For LUE models, which simulated GPP through the theory of radiation conversion efficiency (Monteith, 1972), it is assumed that GPP had a direct relationship with the combination of absorbed photosynthetically active radiation and factual LUE (Xiao et al., 2004;Yuan et al., 2010;Guan et al., 2022) as the factual LUE was linked with the potential ceiling amount of LUE and various regulations of environmental factors on it (Hilker et al., 2008;Guan et al., 2021). Based on these various models and large amounts of remotely sensed and auxiliary data (e.g., meteorological reanalysis product), diverse GPP datasets have been generated and used to investigate the global climate change and carbon cycle (Campbell et al., 2017;Curasi et al., 2019). However, previous studies indicated that the model structures, model parameters, and input data among different models could bring uncertainties to final GPP simulations (Zhao et al., 2006;Xiao et al., 2011;Sanchez et al., 2015;Zhou et al., 2016;Zheng et al., 2018). In this way, the global pattern of drought effect on photosynthesis acquired from different GPP products might exist in unavoidable discrepancies.
As there were more and more global GPP datasets that could be accessed directly online (Zhang et al., 2017;Huntzinger et al., 2018;Zheng et al., 2020), they provided a feasible opportunity to derive a more reliable and robust global pattern of drought impact on vegetation photosynthesis. Considering these findings mentioned above, this study aimed to evaluate the cumulative impact of drought on plant photosynthesis quantitatively based on multiple global GPP products and a multiscale time-series global Standardized Precipitation Evapotranspiration Index (SPEI) dataset. Here, the major objectives of this work were as follows: 1) to explore the spatial pattern of cumulative drought impact on global photosynthesis, 2) to analyze the discrepancies of drought cumulative effect on photosynthesis across different vegetation types, and 3) to derive a global reliability pattern of drought cumulative effect on photosynthesis using the probability statistic based on numerous global GPP products.

Global GPP Products
In this study, a total of twenty-six global GPP products were acquired to evaluate the cumulative drought impact on photosynthesis. These GPP products contain seventeen process-oriented models (i.e., BESS, BIOME-BGC, CLASS-CTEM-N+, CLM4, CLM4VIC, DLEM, GTEC, ISAM, JULES, LPJ-wsl, ORCHIDEE-LSCE, SIB3, SIBCASA, TEM6, TRIPLEX-GHG, VEGAS2.1, and VISIT), five LUE models (i.e., EC-LUE, MOD17, OPT-LUE, RC-LUE, and VPM), and four ML models (i.e., ANN, MARS, MTE, and RF). While these GPP datasets had different spatial and temporal resolutions, all GPP datasets were processed into a uniform format (i.e., 0.5°and monthly) based on the spatial average resampling and accumulated GPP within a month. In addition, the overlapped years of these GPP datasets (i.e., 2001-2010) were used as the study period in this work. More detailed information about these GPP datasets can be found in Supplementary Material.

Multiscale Global SPEI Data
In this work, monthly SPEI data at multiple time scales (1-12 months) from 2001 to 2010 were acquired from the SPEIbase v.2.5 datasets (Beguería et al., 2017) to identify the duration and intensity of drought. This product contained monthly SPEI data at 1-48 months' time-scales from 1901-2015 with a spatial resolution of 0.5°, while SPEI at i time-scale (i.e., i month SPEI) represented the cumulative climatic moisture cycling process for the earlier i months. In this way, SPEI data can be used to characterize different types of drought (e.g., short, middle, and long period drought) and various effects on vegetation status (Begueria et al., 2010;Vicente-Serrano et al., 2010;Peng et al., 2019). As for this product, the SPEI value at each grid was computed through the discrepancy among precipitation and reference evapotranspiration first, and then the final SPEI value was a standardized parameter related to the climatic moisture cycling process that obeys a log-logistic distribution . In addition, a more large positive SPEI value indicated surplus water supply after meeting the coincident moisture requirement, and a smaller negative value represented a more exigent water deficit (Begueria et al., 2014). In this way, SPEI was a reasonable parameter to characterize the degree of drought. Here, the 1-12 month SPEI data for 2001-2010 were adopted to explore the drought's cumulative impact on global photosynthesis according to previous studies (Kang et al., 2018;Peng et al., 2019).

Global Land Cover Data
In this work, moderate-resolution imaging spectroradiometer yearly global land cover types dataset Version 6 (MCD12C1 v006) was adopted to compare the cumulative effect on photosynthesis across vegetation types. The MCD12C1 dataset (Friedl and Sulla-Menashe., 2015) provided three land cover classification scheme layers at 0.05°spatial resolution yearly, and the International Geosphere-Biosphere Program (IGBP) layer was used in this work. Moreover, the 0.05°resolution land cover maps were resampled to the 0.5°resolution ones based on the assumption that the land cover of every 0.5-grid was the major stamp (i.e., with the maximal area percentage) among all the 0.05°subpixels. As for demonstrating the cumulative drought impact on global photosynthesis and reducing the uncertainties attracted by land cover alterations, merely the unchanged vegetation cover pixels (i.e., these pixels that remained the same vegetation type during 2001-2010) were selected as the study area. In this way, the unchanged vegetation cover map is based on the IGBP scheme as shown in Figure 1. A total of eleven vegetation types were selected from the unchanged vegetation cover map (the total amount of pixels was 77,930, accounting for 93.04% of all the land surface). All the used vegetation types, their total numbers of pixels, and area proportions to the land surface are shown below: deciduous broadleaf forests (DBFs, 941, 1.12%), deciduous needleleaf forests (DNFs, 120, 0.14%), evergreen broadleaf forests (EBFs, 4008, 4.79%), evergreen needleleaf forests (ENFs, 1144, 1.37%), mixed forests (MFs, 2204, 2.63%), closed shrublands (CSHs, 116, 0.14%), open shrublands (OSHs, 6638, 7.93%), savannas (SAVs, 6217, 7.42%), woody savannas (WSAs, 4412, 5.27%), croplands (CROs, 4717, 5.63%), and grasslands (GRAs, 12592, 15.03%).

Determining the Cumulative Drought Impact on Plant Photosynthesis
The cumulative drought effect on photosynthesis was concluded through correlation analysis in which time scale SPEI (i.e., any month in 1-12 months) had the maximum significant correlation with monthly GPP. For example, assuming that the i month SPEI data (i could be anyone in the range of 1-12) showed the largest correlation with monthly GPP at a pixel, the cumulative impact of drought on photosynthesis would be set as i months for this pixel, which indicated that the earlier i month climatic water balance was important to affect vegetation photosynthesis. The specific processes to determine the cumulative effect of drought for each GPP dataset could be summarized into three steps. At first, time series monthly GPP product and 1-12 months SPEI in the timeperiod of 2001-2010 were extracted. Second, the reaction of monthly GPP to 1-12 months SPEI was characterized through Pearson's correlation coefficient, while the significance level (i.e., p) was set to 0.05 for each pixel. Third, the accumulated months and intensity of drought cumulative impact on FIGURE 1 | Spatial distribution of unchanged IGBP land cover types obtained from MCD12C1 Version six during the study period.
Frontiers in Environmental Science | www.frontiersin.org May 2022 | Volume 10 | Article 908875 photosynthesis was determined as i month and R max-cum (R maxcum with a range of −1 to 1), while the absolute largest significant Pearson's correlation coefficient (i.e., |R max-cum | when p < 0.05) happened between monthly GPP and i month SPEI.

Statistical Analyses
Based on Pearson's correlation coefficients between monthly GPP data and multiscale monthly SPEI, the R max-cum and accumulated mouth maps were obtained for each GPP dataset according to the method mentioned in Section 2.2. Furthermore, in order to compare the cumulative drought impacts on global photosynthesis which were acquired through different GPP datasets, the R max-cum were divided into three types, including a significant positive correlation (i.e., R max-cum > 0 when p < 0.05), significant negative correlation (i.e., R max-cum < 0 when p < 0.05), and no significant correlation (i.e., p ≥ 0.05). Moreover, the accumulated months could be ranged likewise into three terms, namely, short term (1 ≤ months ≤4), medium term (5 ≤ months ≤8), and long term (9 ≤ months ≤12). In this way, there were both three possible results of R max-cum (i.e., positive, negative, and no significant correlation) and accumulated months (i.e., short, medium, and long term) for different GPP datasets at each pixel. Based on these results of R max-cum acquired from multiple GPP datasets, the final R max-cum outcome for each pixel was determined through probability statistics wherein one result had the highest proportion within all these GPP datasets. Moreover, the reliability (Re) of the final R max-cum for each pixel was determined as follows: where N Rmax-cum represented the total number of final R max-cum from all GPP datasets and N GPP was the total amount of GPP products. At the same time, the final accumulated month outcome and its reliability for each pixel were also obtained using the same processes as the final R max-cum . As for comparing the cumulative drought impact on photosynthesis across vegetation types, the percentage (Pe) of different types of R max-cum for each vegetation cover was obtained as follows: where j was the types of R max-cum (i.e., positive, negative, and no significant correlation), M was the number of pixels that showed R max-cum , and M all was the total pixels for each vegetation type. Similar to the R max-cum , the percentage of different types of accumulated months for each vegetation cover land was also acquired as follows: where i was the types of accumulated months (i.e., short, medium, and long term), S was the number of pixels showing accumulated months, and S all was all the significant pixels for one vegetation type. In addition, the variation of mean reliability with latitudes can be obtained as follows: where Re Mean was the averaged reliability for each latitude, P was the reliability of pixels at the same latitude, and i was the number of pixels with reliability for the same latitude.

Spatial Pattern of Drought Cumulative Effect on Photosynthesis
The spatial distribution of R max-cum which is based on probability statistics is shown in Figure 2. The results indicated that 52.11% of vegetation cover lands have shown significant R max-cum , and the positive correlations had a higher proportion (38.30%) than negative correlations (13.81%). Meanwhile, the positive R maxcum was primarily observed in North America, South America, South-central Africa, Central Asia, South Asia, Oceania, and some parts of East Asia. The negative R max-cum was mostly discovered at the medium and high latitudes of the Northern Hemisphere, such as eastern North America, Northern Asia, and FIGURE 2 | Spatial distribution of R max-cum based on probability statistic. FIGURE 3 | Spatial distribution of accumulated months based on probability statistic.
Frontiers in Environmental Science | www.frontiersin.org May 2022 | Volume 10 | Article 908875 some parts of Eastern Europe. As shown in Figure 3, the spatial distribution of accumulated months was significantly different at the global scale, while the accumulated months were mainly concentrated in the short term (31.81%), followed by medium term (16.32%) and long term (3.98%). The accumulated months in the short term were mostly found in North America, South America, Central Africa, Northern Asia, and Oceania. The medium-term accumulated months were mainly observed among Central Asia, Southern Africa, and some parts of East Asia. Moreover, the long term accumulation mainly occurred in East and South Asia.

Cumulative Drought Impact on Photosynthesis Across Plant Types
As shown in

Global Reliability Pattern of Drought Cumulative Effect on Photosynthesis
The global reliability pattern of R max-cum thought probability analysis is shown in Figure 4. The R max-cum had relatively high reliability at the global scale (characterized by the averaged value of 80.37 ± 17.33%). Moreover, the R max-cum showed high reliability in high latitudes and low reliability around the equatorial regions. The reliability of accumulated   months displayed a similar global pattern with R max-cum ( Figure 5), characterized by the mean value of 74.68 ± 20.79%. Notably, there were obvious discrepancies in the reliability of R max-cum and accumulated months around Oceania, South Asia, and South Africa, while the R max-cum presented high reliability and the accumulated months showed low reliability in these areas. Figure 6 showed the variability trend of reliability in R max-cum and accumulated months across latitudes. The reliability of R maxcum was decreased with the decreasing of latitude, while the reliability of accumulated months was undulated in the Southern Hemisphere. As for the reliabilities of both R max-cum and accumulated months, they were various across vegetation types ( Table 2). The averaged reliabilities of R max-cum in diverse plant types followed the order: DNFs (

DISCUSSION
This study evaluated the cumulative drought impact on global photosynthesis across vegetation types grounded on numerous GPP products and derived a global reliability pattern of drought cumulative effect through probability analysis. The results illustrated that the cumulative effects of drought occurred in 52.11% of the vegetation cover areas (Figure 2) and the dominant accumulated months were short term (1-4 months, 31.81%). Our study indicated the high spatial heterogeneity of drought effect at the global scale and the importance of the GPP dataset in evaluating the cumulative drought impact on photosynthesis.
In this study, the global reliability pattern of drought cumulative effect on photosynthesis was acquired through numerous GPP products, and both R max-cum and accumulated months exhibited spatially heterogenous distributions (Figure 4 and Figure 5). Moreover, the reliability of cumulative effect in time scales was more undulated, especially in the southern hemisphere ( Figure 6). These findings revealed that the water supply and availability could play a key role in GPP simulation. Generally, a total of twenty-six global GPP datasets generated from different model schemes (i.e., based on LUE, ML, and process) were used as indicators to represent the photosynthesis of vegetation. Among various models, these process-oriented models always considered many physiological or biogeochemical processes of photosynthesis mechanistically (Jiang and Ryu, 2016;Wagle et al., 2016;Xie et al., 2021), thus making the GPP estimation more rigorous. While these LUE models were developed through the theory of radiation conversion efficiency (Monteith, 1972;Yuan et al., 2014), the ML models estimated GPP based on the statistical relationship established between input data and outcome (Jiang and Ryu, 2016;Wolanin et al., 2019;Xie et al., 2022). In this way, different theoretical foundations among diverse models would bring unavoidable uncertainties in characterizing global photosynthesis (Zheng et al., 2020;Wang S. et al., 2021;Wang Z. et al., 2021). In addition, multiple GPP models have used various strategies to describe the water stress of vegetation. For example, the EC-LUE model used the rate between evapotranspiration and net radiation to consider the moisture regulations on LUE (Yuan et al., 2010), while the VPM model adopted a scaled Land Surface Water Index to characterize the moisture limitation for photosynthesis (Zhang et al., 2017). Moreover, the water limiting processes could be more complex in process-oriented models (Jiang and Ryu, 2016;Slevin et al., 2017). These different ways of regulating the moisture status of plants might also lead to a diverse response of photosynthesis to drought among different GPP datasets.
The cumulative drought impact on photosynthesis and its reliability showed obvious discrepancies among vegetation types  Table 1 and Table 2). The results showed that there were higher proportions of significant R max-cum pixels in CSHs, OSHs, and GRAs than those in DNFs, MFs, ENFs, and WSAs, confirming the importance of water supply and availability for grasslands and shrubs (Wu et al., 2015;Peng et al., 2019;Fan et al., 2020), while this might be attributed to the different structural and functional characteristics of rhizomes between herbs and woody plants (Chimento and Amaducci, 2015;Hudek et al., 2017;Lobmann et al., 2020), such as root diameter (Hudek et al., 2017), root depth (Seghieri, 1995), and stem specific gravity (McCoy-Sulentic et al., 2017). Herbaceous vegetations usually have abundant fine and shallow roots in the topsoil layer, making them respond more sensitively to drought due to the limitation in water availability through more deep layers of soil (Dodd et al., 1998;Wu et al., 2018;Lobmann et al., 2020). However, woody plants always have taproot structures with thick roots, and deeper roots facilitate the absorption of groundwater which causes them to be more resistant to drought (Dodd et al., 1998;Midwood et al., 1998;Bleby et al., 2010;Wang et al., 2020). Moreover, the roots of woody plants are further in evolution than herbaceous plants and are characterized by secondary growth (Ma et al., 2018;Wang et al., 2020), helping them withstand stress caused by drought. Different water storage modes may also result in diverse responses to drought across vegetations (Tian et al., 2018) because woody vegetations hold most plant water in their woody tissues and have more complex hydraulic strategies, while the moisture storage of herbs primarily depends on the amount of foliage (Sternberg and Shoshany, 2001;Morris et al., 2016).

CONCLUSION
In this work, a total of twenty-six global GPP products based on process, LUE, and ML models were obtained to evaluate and derive a global reliability pattern of drought's cumulative effect on photosynthesis. The results illustrated that the cumulative drought impact was observed in 52.11% of vegetation cover land primarily in the short term (31.81%). Obvious discrepancies of drought cumulative effect were observed in different plant functional types, while herbaceous vegetations were more sensitive to drought than woody plants. This could be attributed to the difference in root functional traits among vegetation functional types. Because the reliability pattern of drought depended on multiple GPP datasets, the results also showed marked heterogeneity in terms of space and vegetation type, while lower reliability was found in the tropics. Our findings highlighted the importance of water characterization strategies under different vegetation functional types in GPP models and the necessity of GPP dataset selection in assessing drought effect on vegetation.

DATA AVAILABILITY STATEMENT
The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.