Performance of the Remotely-Derived Products in Monitoring Gross Primary Production across Arid and Semi-Arid Ecosystems in Northwest China

: As an important component to quantify the carbon budget, accurate evaluation of terrestrial gross primary production (GPP) is crucial for large-scale applications, especially in dryland ecosystems. Based on the in situ data from six ﬂux sites in northwestern China from 2014 to 2016, this study compares seasonal and interannual dynamics of carbon ﬂuxes between these arid and semi-arid ecosystems and the atmosphere. Meanwhile, the reliability of multiple remotely-derived GPP products in representative drylands was examined, including the Breathing Earth System Simulator (BESS), the Moderate Resolution Imaging Spectroradiometer (MODIS) and data derived from the OCO-2 solar-induced chlorophyll ﬂuorescence (GOSIF). The results indicated that the carbon ﬂuxes had clear seasonal patterns, with all ecosystems functioning as carbon sinks. The maize cropland had the highest GPP with 1183 g C m − 2 y − 1 . Although the net ecosystem carbon exchange (NEE) in the Tamarix spp. ecosystem was the smallest among these ﬂux sites, it reached 208 g C m − 2 y − 1 . Furthermore, distinct advantages of GOSIF GPP (with R 2 = 0.85–0.98, and RMSE = 0.87–2.66 g C m − 2 d − 1 ) were found with good performance. However, large underestimations in three GPP products existed during the growing seasons, except in grassland ecosystems. The main reasons can be ascribed to the uncertainties in the key model parameters, including the underestimated light use e ﬃ ciency of the MODIS GPP, the same coarse land cover product for the BESS and MODIS GPP, the coarse gridded meteorological data, and distribution of C 3 and C 4 plants. Therefore, it still requires more work to accurately quantify the GPP across these dryland ecosystems. based on discrete OCO-2 SIF soundings, MODIS data, and MERRA-2 meteorological data. In from and based on the geographical location of each ﬂux site for completing the performance.


Introduction
Climate change and climate extremes always exert a substantial impact on the biosphere, but the response, as well as the underlying mechanism, vary among different vegetation types [1]. Especially under the context of global warming, the arid and semi-arid (ASA) regions are projected to expand further in the coming decades [2,3]. Koutroulis [4] found that the dryland had enlarged by 4-8% The S_Dm site and the S_Zy site are located in the Daman and Yingke irrigated areas near Zhangye city Gansu province, respectively. They are also in the midstream of the Heihe river (artificial oasis experimental area) having a continental arid climate [32]. The S_Zy site is situated on a wetland with a reed surface in the Zhangye National Wetland Park. The S_Ar and the S_Dsl sites are part of the upstream cold region experimental area. The former is in a valley with an east-west direction and a maximum width of 3 km [15]. The latter is covered with snow most of the time [34]. The S_Dm site and the S_Zy site are located in the Daman and Yingke irrigated areas near Zhangye city Gansu province, respectively. They are also in the midstream of the Heihe river (artificial oasis experimental area) having a continental arid climate [32]. The S_Zy site is situated on a wetland with a reed surface in the Zhangye National Wetland Park. The S_Ar and the S_Dsl sites are part of the upstream cold region experimental area. The former is in a valley with an east-west direction and a maximum width of 3 km [15]. The latter is covered with snow most of the time [34]. The S_Sdq and the S_Hhl sites are located on the natural oasis experimental area in the downstream. More information can be found in Table 1 and Figure 1. The mean temperature represents the average daily temperature from 2014 to 2016 and precipitation represents cumulative average annual precipitation from 2014 to 2016 for each flux site. The standard deviations of mean temperature and precipitation are also attached.

Flux Data Acquisition and Processing
Each flux tower site has two systems including an EC and an automatic weather station (AWS) system, to obtain the energy fluxes and meteorological data. A three-dimensional sonic anemometer (CSAT3, Campbell Scientific, Logan, UT, USA) and a Li7500A open path CO 2 /H 2 O gas analyzer (LI-COR Inc., NE, USA) form an EC system to measure ecosystem scale CO 2 and H 2 O fluxes. The sampling rate was set to 10 Hz. Before releasing data to the public, the 10 Hz data was processed into a half-hour flux dataset by the Edire post-processing software (University of Edinburgh, Edinburgh, UK). The spike detection, lag correction of H 2 O/CO 2 , sonic virtual temperature correction, coordinate rotation, Webb-Pearman-Leuning correction (WPL) [35], and frequency response correction were included in the procedures. In addition, sensor-faults, rain, and other bad weather can easily cause data loss or low quality. Therefore, data with poor quality was checked and excluded from the half-hour data [36]. The AWS system measures the relevant weather conditions including air temperature (Ta) ( • C), precipitation (mm), etc. [37]. The data of the AWS system were recorded every ten minutes.
In this study, three-year flux data and meteorological data of these typical sites were obtained from the Environmental and Ecological Science Data Center of West China (http://westdc.westgis.ac.cn). Then, preprocessing such as outlier elimination was carried out. Gap-filling by a standardized algorithm and flux partitioning by the nighttime-based method were carried out with the R package by the Max Planck Institute for Biogeochemistry (https://www.bgc-jena.mpg.de/bgi/index.php/Services/ REddyProcWeb). Flux partition was used to partition NEE into the GPP and ecosystem respiration (R e ). In particular, the latent heat flux and carbon dioxide flux data were missing from 1st May 2016 to 23rd August 2016 because of a malfunction of the Li7500 at the S_Sdq site and we gap-filled this part with multi-year mean imputation. Finally, we integrated these half-hour data into the daily time scale and then got the average of each eight-day period to compare with the BESS, MODIS, and GOSIF products.

Remote Sensing Products and Processing
The BESS is designed to continuously measure global carbon and water fluxes [29] and provides the global 1 km resolution GPP product with an interval of eight days during the period from 2000 to 2015. In the BESS-derived GPP, radiation, photosynthetic active radiation, and near-infrared radiation at the top of the canopy were calculated by an atmospheric radiative transfer model, and then these parameters, absorbed by the sunlit and shade leaves, were gained by a two-leaf canopy radiative transfer model. Furthermore, it used the biochemical photosynthesis models for C 3 and C 4 plants, and parameterized maximum carboxylation rate based on plant functional type, climatic zones, and the season. The BESS products used seven MODIS atmospheric and land products, four other satellites, four reanalyses, and three auxiliary datasets.
We used the MODIS GPP version 6 (MOD17A2H) obtained from the Terra platform, with 500 m spatial resolution [36], which is produced using the LUE model as follows: where LUE max is the maximum value of light use efficiency, while SWRad denotes incident shortwave radiation. T min is the daily minimum temperature, VPD is the daylight average vapor pressure deficit, and FPAR is the fractional photo-synthetically active radiation absorbed by the canopy. The algorithm logic and model parameters can also be found in the MOD17 user guide [38]. The model assumed that the GPP is proportional to plant absorption and effective radiation under ideal conditions [39]. A set of biome-specific radiation-use efficiency parameters was acquired from the updated Biome Property Look Up Tables. The product also used a new version of the daily Global Modeling and Assimilation Office (GMAO) meteorological data [39]. The product used in this study was downloaded from NASA's Earth Observing System Data and Information System (https://search.earthdata.nasa.gov).
We re-sampled it to 1 km pixel size before being extracted based on site locations. The GOSIF GPP product was developed by Li and Xiao [24], with a 0.05 • spatial resolution and an eight-day time step for the period from 2000 to 2018. The GOSIF GPP product benefited from the breakthrough of two key matters, the relationship between SIF and GPP, and SIF data with high spatial and temporal resolutions. For the relationship between SIF and GPP, Li and Xiao [40] found a nearly universal SIF-GPP relationship. The SIF data used the global OCO-2 SIF data set (GOSIF) with a 0.05 • spatial resolution and an eight-day time step. The GOSIF product was developed by Li and Xiao [41], based on discrete OCO-2 SIF soundings, MODIS data, and MERRA-2 meteorological data. In this study, the BESS, MODIS, and GOSIF GPP data from 2014-2015 were selected uniformly, and we extracted the value based on the geographical location of each flux site for completing the performance.

Statistical Analyses
To quantify the relationships between two GPP products and EC-based data in major ecosystems, we used two statistical criteria: the root-mean-square error (RMSE) and R 2 . The equations of RMSE and R 2 were as follows: where y andŷ represent the observed and remotely-derived GPP, respectively; n is the number of observations (eight-day periods). The best products generally had the highest R 2 and the lowest RMSE. All statistical analyses were carried out using the OriginPro software (OriginLab, OriginPro 8.5, Northampton, MA, USA).

Dynamics of GPP, R e , and NEE in Dryland Ecosystems
Firstly, the seasonal dynamics of carbon fluxes (GPP, R e , and NEE) are shown in Figure 2. As illustrated in Figure 2, the GPP, R e , and NEE of all flux sites showed clear seasonal patterns from 2014 to 2016. Particularly, the GPP and R e exhibited consistent trends. During the dormant seasons, the vegetation GPP was nearly zero because photosynthesis stopped. With the increase of temperature during springtime, the GPP and R e increased gradually and reached a peak in July or August. Inversely, the NEE had deep valleys in the growing seasons. It decreased gradually after approximately 18th week of year (WOY 18), to the valley around in July or August, and then gradually increased. In addition, the S_Dm site had the highest GPP at the peak period among all sites. Although consistent seasonal dynamics exist, different ecosystems still had their own characteristics, especially in forest ecosystems (the S_Sdq and S_Hhl sites). The time when GPP began to grow rapidly was earlier at S_Hhl than the rest of the other sites, and the period of carbon sequestration was the longest in every year among all ecosystems. In the S_Sdq site, we found that the peak periods came later year after year from 2014 to 2016.

Dynamics of GPP, Re, and NEE in Dryland Ecosystems
Firstly, the seasonal dynamics of carbon fluxes (GPP, Re, and NEE) are shown in Figure 2. As illustrated in Figure 2, the GPP, Re, and NEE of all flux sites showed clear seasonal patterns from 2014 to 2016. Particularly, the GPP and Re exhibited consistent trends. During the dormant seasons, the vegetation GPP was nearly zero because photosynthesis stopped. With the increase of temperature during springtime, the GPP and Re increased gradually and reached a peak in July or August. Inversely, the NEE had deep valleys in the growing seasons. It decreased gradually after approximately 18th week of year (WOY 18), to the valley around in July or August, and then gradually increased. In addition, the S_Dm site had the highest GPP at the peak period among all sites. Although consistent seasonal dynamics exist, different ecosystems still had their own characteristics, especially in forest ecosystems (the S_Sdq and S_Hhl sites). The time when GPP began to grow rapidly was earlier at S_Hhl than the rest of the other sites, and the period of carbon sequestration was the longest in every year among all ecosystems. In the S_Sdq site, we found that the peak periods came later year after year from 2014 to 2016. To analyze the differences in carbon fluxes between different ecosystems, we conducted a statistical analysis of carbon fluxes (GPP, Re, and NEE) from 2014 to 2016. The annual mean carbon fluxes of different ecosystems were statistically calculated, presented in Table 2. Overall, the average annual GPP in all flux sites was greater than the values of Re, and according to the NEE annual mean value, it implied that all ecosystems also behaved as carbon sinks. In particular, the S_Dm site had the largest annual mean GPP and NEE, which was near double the GPP and NEE values of other ecosystems. Furthermore, the lowest annual mean GPP value appeared in the S_Dsl site, in which the annual mean Re value was also the lowest, reaching up to 208 ± 49 g C m −2 y −1 . However, although both the annual GPP and Re in the S_Dsl site were far lower than those in the S_Ar site, and there was To analyze the differences in carbon fluxes between different ecosystems, we conducted a statistical analysis of carbon fluxes (GPP, R e , and NEE) from 2014 to 2016. The annual mean carbon fluxes of different ecosystems were statistically calculated, presented in Table 2. Overall, the average annual GPP in all flux sites was greater than the values of R e , and according to the NEE annual mean value, it implied that all ecosystems also behaved as carbon sinks. In particular, the S_Dm site had the largest annual mean GPP and NEE, which was near double the GPP and NEE values of other ecosystems. Furthermore, the lowest annual mean GPP value appeared in the S_Dsl site, in which the annual mean R e value was also the lowest, reaching up to 208 ± 49 g C m −2 y −1 . However, although both the annual GPP and R e in the S_Dsl site were far lower than those in the S_Ar site, and there was no obvious difference in the annual mean NEE between the two sites. The largest annual mean NEE was found at the S_Sdq site, and it implied that the lowest carbon sequestration was there, about 208 ± 56 g C m −2 y −1 .

Performance of Satellite-Based Seasonal GPP
In this study, three fundamental methods contributing to the analysis of the differences in the seasonal performances of the BESS, MODIS, and GOSIF GPP products were carried out. It is worth noting that the BESS and MODIS GPP extracted based on the S_Sd site locations were both invalid data. Firstly, based on the EC-based observations GPP (GPP EC ), we assessed the performance across seasonal courses (  Figure 3, we can find that the performances of the three GPP products varied dramatically with different ecosystems. On the whole, the seasonal variation of the three GPP products was the most consistent with the flux site GPP EC of grassland ecosystems (S_Ar and S_Dsl) but remained largely underestimated for the growing seasons at the other EC sites. Furthermore, the three GPP products had well peaks matching at the S_Dm site and a degree of peak delay in both grassland ecosystems (the S_Ar and S_Dsl sites). Moreover, the GPP SIF performed better than the other GPP products generally, but it was obviously overestimated during the growing season of the S_Ar site. Linear regression analysis was performed to examine the performance of the BESS, MODIS, and GOSIF GPP products on an eight-day period at the five flux sites between 2014 and 2015. As the values of both GPPBESS and GPPMOD were null at the S_Zy site, the performance of these two products could not be further analyzed. As shown in Figure 4, different GPP products varied with different ecosystems. Overall, the three GPP products' performance was underestimated or overestimated in  Linear regression analysis was performed to examine the performance of the BESS, MODIS, and GOSIF GPP products on an eight-day period at the five flux sites between 2014 and 2015. As the values of both GPP BESS and GPP MOD were null at the S_Zy site, the performance of these two products could not be further analyzed. As shown in Figure 4, different GPP products varied with different ecosystems. Overall, the three GPP products' performance was underestimated or overestimated in all EC sites and the GPP SIF performed better than the other GPP products. Especially, the GPP SIF had the highest consistency with GPP EC at the S_Dm, the S_Sdq, and the S_Hhl sites, with R 2 values ranging from 0.92 to 0.98 and RMSE from 1.5 to 2.66 g C m −2 d −1 . However, the GPP SIF (R 2 = 0.95, RMSE = 1.47 g C m −2 d −1 ) was obviously overestimated at the S_Ar site, having no superiority over the GPP BESS and GPP MOD . At the S_Ar and S_Dsl, The GPP BESS and GPP MOD were distributed closer to the 1:1 line like the GPP SIF , which is consistent with the results of Figure 3.
In order to examine the different performances of the GPP BESS , GPP MOD , and GPP SIF in spring, summer, and autumn, we carried out a boxplot analysis, as shown in Figure 5. In terms of grassland ecosystems (S_Ar and S_Dsl), the GPP MOD had better distribution characteristics than the GPP BESS and GPP SIF in summer, while the GPP BESS was closer to distribution characteristics of the GPP EC in spring and autumn. In addition, the GPP SIF performed better only in spring for the S_Ar site. The GPP SIF also had better distribution characteristics than the GPP BESS and GPP MOD at the S_Dm site in spring, summer, and autumn. Furthermore, the GPP BESS , GPP MOD , and GPP SIF did not perform well at the S_Sdq and S_Hhl sites.

Annual Mean Patterns of Satellite-Based GPP
The annual mean GPP from flux measurements (GPPEC), BESS (GPPBESS), the MODIS (GPPMOD), and GOSIF (GPPSIF) at the six flux sites from 2014 to 2015 are shown in Figure 6. Firstly, at the S_Ar site, although both the GPPBESS and GPPSIF were overestimated, the GPPBESS had the best performance. Furthermore, the three GPP products were underestimated in varying degrees at other EC sites. For instance, at the S_Sdq and S_Hhl sites, the three GPPs were severely underestimated by more than half. Similarly, the GPPBESS and GPPMOD were only approximately half of the GPPEC at the S_Dm site. Finally, the GPPSIF had an obvious advantage at the S_Dm site among the three GPP products. Figure 5. Comparisons of the seasonal patterns of GPP observations (GPP EC ) and the remotely-derived products (GPP BESS , GPP MOD , and GPP SIF ) at the six flux sites during the periods of spring, summer, autumn, and winter, respectively. The box range and whisker range were 25%-75% and 10%-90%, respectively.

Annual Mean Patterns of Satellite-Based GPP
The annual mean GPP from flux measurements (GPP EC ), BESS (GPP BESS ), the MODIS (GPP MOD ), and GOSIF (GPP SIF ) at the six flux sites from 2014 to 2015 are shown in Figure 6. Firstly, at the S_Ar site, although both the GPP BESS and GPP SIF were overestimated, the GPP BESS had the best performance. Furthermore, the three GPP products were underestimated in varying degrees at other EC sites. For instance, at the S_Sdq and S_Hhl sites, the three GPPs were severely underestimated by more than half. Similarly, the GPP BESS and GPP MOD were only approximately half of the GPP EC at the S_Dm site. Finally, the GPP SIF had an obvious advantage at the S_Dm site among the three GPP products.

Uncertainty in EC-Based GPP
Compromised due to instrument malfunctions, inadequate power supply, and bad weather, certain carbon fluxes and meteorological data can have poor quality and are even lost, which is inevitable. In the above cases, the data with an occasional segment missing or quality problems have been dealt with by the methods of gap-filling using the R package. However, for data gaps or quality problems of more than one month, we gap-filled with multi-year mean imputation, which was mainly used at the S_Sdq site in 2016. To some extent, it covered up the difference left by the specific environment, which showed some uncertainties. Besides, previous work has reported the overestimation of plant photosynthetic uptake when using the integrated open-path eddy covariance system [42]. Despite uncertainties, the six flux sites selected in this study represent the most typical ASA ecosystems in western China. The results obtained in this study also clearly reveal the quadratic functional characteristics of carbon fluxes (GPP, Re, and NEE) concerning the seasonal dynamics in these ecosystems. Further, the results suggest that the cropland ecosystems achieved the highest annual GPP values of 1182.68 ± 38.62 g C m −2 y −1 among all ASA ecosystems, which was consistent with the result of Gu [15]. On the annual scale, the values of GPPEC at the S_Ar and S_Dsl sites were about 741 g C m −2 y −1 and 528 g C m −2 y −1 . The heavy precipitation events (over 10 mm day −1 ) as well as precipitation intervals have been reported to impact GPP in the temperate steppe or the alpine meadow ecosystems [43]. Compared with the S_Dsl site, we found that the S_Ar had more frequent precipitation events (an average of 13 days during the growing season of 2014-2015), and annual average precipitation was about 460 ± 63 mm, which was likely to favor the higher GPP value of the S-Ar site. For semi-arid grasslands in the western United States, Álvarez-Taboada [44] examined three grassland ecosystems under water-limited conditions and found that the EC-based GPPs were about 500 g C m −2 y −1 . Especially at the Vaira Ranch-Ione site with annual cumulative precipitation of approximately 630 mm, the mean GPP from 2004 to 2006 was approximately 822 g C m −2 y −1 . Therefore, the carbon fixation in the arid and semi-arid ecosystems was strongly constrained by water availability and fluctuated annually.

Uncertainty in EC-Based GPP
Compromised due to instrument malfunctions, inadequate power supply, and bad weather, certain carbon fluxes and meteorological data can have poor quality and are even lost, which is inevitable. In the above cases, the data with an occasional segment missing or quality problems have been dealt with by the methods of gap-filling using the R package. However, for data gaps or quality problems of more than one month, we gap-filled with multi-year mean imputation, which was mainly used at the S_Sdq site in 2016. To some extent, it covered up the difference left by the specific environment, which showed some uncertainties. Besides, previous work has reported the overestimation of plant photosynthetic uptake when using the integrated open-path eddy covariance system [42]. Despite uncertainties, the six flux sites selected in this study represent the most typical ASA ecosystems in western China. The results obtained in this study also clearly reveal the quadratic functional characteristics of carbon fluxes (GPP, R e , and NEE) concerning the seasonal dynamics in these ecosystems. Further, the results suggest that the cropland ecosystems achieved the highest annual GPP values of 1182.68 ± 38.62 g C m −2 y −1 among all ASA ecosystems, which was consistent with the result of Gu [15]. On the annual scale, the values of GPP EC at the S_Ar and S_Dsl sites were about 741 g C m −2 y −1 and 528 g C m −2 y −1 . The heavy precipitation events (over 10 mm day −1 ) as well as precipitation intervals have been reported to impact GPP in the temperate steppe or the alpine meadow ecosystems [43]. Compared with the S_Dsl site, we found that the S_Ar had more frequent precipitation events (an average of 13 days during the growing season of 2014-2015), and annual average precipitation was about 460 ± 63 mm, which was likely to favor the higher GPP value of the S-Ar site. For semi-arid grasslands in the western United States, Álvarez-Taboada [44] examined three grassland ecosystems under water-limited conditions and found that the EC-based GPPs were about 500 g C m −2 y −1 . Especially at the Vaira Ranch-Ione site with annual cumulative precipitation of approximately 630 mm, the mean GPP from 2004 to 2006 was approximately 822 g C m −2 y −1 . Therefore, the carbon fixation in the arid and semi-arid ecosystems was strongly constrained by water availability and fluctuated annually.

Error Sources of Remotely-Derived Products
This study found that both the MODIS and BESS GPP performed with good accuracy for grasslands, which was consistent with previous studies [29,45], but larger uncertainties existed in other ecosystems. Regardless of the algorithm error, the precision of the model parameters used for the estimation plays an indispensable role in the accuracy of GPP estimation. For the input parameters of the MODIS GPP algorithm, we can take knowledge from the MOD17 user guide [38], while for the model parameters of the BESS GPP, the description is very clear in the research of Jiang and Ryu [29]. The BESS model contains the six key modules algorithm [17], and most of its input parameters come from the MODIS products whose resolution difference is large enough to make uncertainty inevitable in the process of the resolution consistency. For the GOSIF GPP, the GOSIF data set, the key part of the input data was affected by MERRA-2 with coarse spatial resolution [41]. Therefore, these products had their own errors.
In this study, we found that these BESS, MODIS, and GOSIF GPP products largely underestimated the GPP during the growing period, except the S_Ar site. This can be partly explained by the open-path EC-based flux measurements. Such systems would inevitably lead to the overestimation of the GPP of terrestrial ecosystems because of the limitation of technical means. However, the GOSIF GPP had obvious advantages in seasonal and interannual dynamics, which can be ascribed to the universal GOSIF-GPP relationship and negligible impact of land covers [41]. Through a thorough analysis of the performances of the MODIS and BESS GPP, we found that both the BESS and MODIS GPP underestimated the GPP data by more than half at S_Hhl (Figures 3 and 6). The reason can be largely explained by the fact that both the BESS and MODIS GPP products used the same land cover product (MCD12Q1) [29,46], which had significant classification uncertainty. Different judgments of land cover resulted in inaccurate LUEmax and affected the estimation results. Figure 7 showed that the S_Hhl and S_Sdq sites were located on grasslands surface, but not on Populus euphratica and Tamarix spp. For the MODIS GPP product, LUEmax, Tmin, and VPD of grasslands are much lower than those of forest land, which is another important reason for underestimation. Zhao [47] found that the Tmin and VPD from the GMAO were generally underestimated. Furthermore, the LUEmax value in MODIS GPP depends on the simple look-up table approach. Among the same vegetation types, the LUEmax value was also generally underestimated [48]. After Wang [28] optimized the LUEmax using the site-observed GPP, the performance of the MODIS GPP had obvious improvement. The footprints of these flux sites can reach several hundred meters around the tower. However, owing to the spatial heterogeneity, its effect on the remote sensing-based GPP products cannot be neglected, which would lead to inconsistency between the site-level measurement and satellite-based product. In addition, soil moisture has been indicated as a major factor limiting the GPP in the water-constrained ecosystems [15,49]. However, the BESS GPP assumed that the water stress is embedded in the variability of leaf area index, which lacked an accurate proxy for soil moisture dynamics [50].
The BESS and MODIS GPP applied different types of models, which made it hard to compare from the model structure. However, the most important difference is that the BESS model distinguished the difference between C 3 and C 4 plants while MODIS GPP did not. The cropland ecosystems mainly represent rice at MCD12Q1 land cover products. Wang [51] also found that the GPP of maize was obviously higher than the other C 3 crops. Rice at the S_Dm is a C 3 plant different from C 4 plants (maize), which is the main reason for the underestimation of MODIS GPP. Although considering the differences in ecophysiological characteristics of C 3 and C 4 species, the BESS GPP underestimated by about 50% of actual measurement (Figures 3 and 6). This is because the BESS GPP depended on the relative proportion of two species on a pixel, and C 4 plants account for only less than 0.205 at the S_Dm [52]. Land 2020, 9,   The BESS and MODIS GPP applied different types of models, which made it hard to compare from the model structure. However, the most important difference is that the BESS model distinguished the difference between C3 and C4 plants while MODIS GPP did not. The cropland ecosystems mainly represent rice at MCD12Q1 land cover products. Wang [51] also found that the GPP of maize was obviously higher than the other C3 crops. Rice at the S_Dm is a C3 plant different from C4 plants (maize), which is the main reason for the underestimation of MODIS GPP. Although considering the differences in ecophysiological characteristics of C3 and C4 species, the BESS GPP underestimated by about 50% of actual measurement (Figures 3 and 6). This is because the BESS GPP depended on the relative proportion of two species on a pixel, and C4 plants account for only less than 0.205 at the S_Dm [52].
Although the GOSIF GPP had also underestimated at the S_Dm, S_Hhl, and S_Sdq sites, it had the best performed among the three GPP products (Figures 3 and 4). Reasons abound, for instance, Li and Xiao [41] found that the impact of land cover on product accuracy is much smaller than that of training samples. Therefore, the land cover maps were not taken into account when the GOSIF products are produced. Similarly, Li and Xiao [24] used a nearly universal rather than biome-specific SIF-GPP relationship to get the GOSIF GPP product. In a nutshell, the GOSIF GPP reduced the impact of land cover classification uncertainty to some extent.

Effect on Global Climate Change
With changing climate, the influence of carbon sequestration in semi-arid biomes may be of more importance for global carbon sink than in tropical rainforests inter-annually (<10 years) in the future Although the GOSIF GPP had also underestimated at the S_Dm, S_Hhl, and S_Sdq sites, it had the best performed among the three GPP products (Figures 3 and 4). Reasons abound, for instance, Li and Xiao [41] found that the impact of land cover on product accuracy is much smaller than that of training samples. Therefore, the land cover maps were not taken into account when the GOSIF products are produced. Similarly, Li and Xiao [24] used a nearly universal rather than biome-specific SIF-GPP relationship to get the GOSIF GPP product. In a nutshell, the GOSIF GPP reduced the impact of land cover classification uncertainty to some extent.

Effect on Global Climate Change
With changing climate, the influence of carbon sequestration in semi-arid biomes may be of more importance for global carbon sink than in tropical rainforests inter-annually (<10 years) in the future [53]. In this study, we found that carbon fluxes varied in different years and different dryland ecosystems. On one hand, it suggested that with restrictions on water resources, ASA ecosystems remained a large carbon sink. On the other hand, there was much uncertainty for carbon sequestration across ASA regions. Meanwhile, large areas of grassland and woodland in Australia has been invaded by shrubs, which lead to a serious decline in pastoral productivity [54]. In this study, we found that the grassland ecosystem is the highest carbon-sequestered vegetation other than farmland. Gu [15] found that soil moisture showed a large positive correlation with the GPP in ASA regions in China. Soil moisture content can reflect soil texture to some extent, which means that grassland has excellent water conservation function and can alleviate the continuous expansion of drought. Therefore, accurately estimating the GPP provided an important scientific method and data support for quantifying carbon sink and source in ASA ecosystems to contribute to the development of regional carbon management planning and decision-making.

Conclusions
Based on six flux sites in northwest China, this study examined the differences in seasonal and interannual carbon fluxes across the typical ASA ecosystems in northwest China. The results indicate that the carbon fluxes of these dryland ecosystems have clear seasonal dynamics and vary inter-annually, and all ecosystems function as carbon sinks on the annual time scale. Moreover, we also assessed the performance of these remotely-derived BESS, MODIS, and GOSIF GPP products, and got some evidence. Except for at the S_Ar and S_Dsl sites, some distinct advantages of the GOSIF GPP were picked up, not only in the seasonal dynamic but also in annual mean difference. The universal GOSIF-GPP relationship and less impact of land cover helped the GOSIF GPP algorithm. Furthermore, both the MODIS and BESS products largely underestimated the GPP during the growing period. Compared with the MODIS GPP, the BESS GPP was more reliable for swamp meadow and alpine grassland ecosystems. However, large uncertainties in the BEES, MODIS, and GOSIF GPP products remain. Therefore, precise input parameters should be used to improve the products, such as accurate land cover types and LUEmax at high spatial resolution. Meanwhile, the MODIS GPP will be greatly improved by considering the differences in sun and shade leaves, and C 3 /C 4 plants. The accuracy of the BESS GPP can be also improved by adding the soil moisture effect in water-limited areas, and the accuracy of the GOSIF GPP could be further improved with high-quality meteorological data. All these analyses would be helpful for providing reliable GPP products in the ecologically-fragile dryland ecosystems and supporting policies for sustainable development.