Increased carbon uptake and water use efficiency in global semi-arid ecosystems

We used flux and meteorological data from the tier-1 sites in the FLUXNET2015 database (https://fluxnet.fluxdata.org/data/fluxnet2015-dataset/). The dataset contains 212 sites, and 166 of these sites are tier-1 sites that are free and open for scientific analysis. These tier-1 sites are distributed across North America, Central and South America, Europe, Asia, Africa, and Australia and encompass a wide range of vegetation types and climate zones. Of the 79 sites in our study, 31 sites are forests, 7 are shrublands, 13 are savannas (SAV), and 19 are grasslands (GRA) (figure 1, table S1 is available online at stacks.iop.org/ERL/15/034022/mmedia). Forests include evergreen needleleaf forests, evergreen broadleaf forests, deciduous broadleaf forests, and mixed forests. Shrublands include both open and closed shrublands (OSH/CSH), and savannas consist of woody savannas and savannas (WSA/SAV). For each of these site-years, we aggregated the daily GPP, NEP, and ET data to 15-day intervals. We excluded the bi-weekly data if the percentage of gap-filled data was more than 20%. The resulting 15-day GPP, NEP, and ET data were then used for model development and accuracy assessment.

Zoom In Zoom Out Reset image size

We also used the GIMMS NDVI3g dataset (http://ecocast.arc.nasa.gov/data/pub/gimms/, Pinzon and Tucker [23]) from 1982 to 2015. The GIMMS NDVI3g dataset provides global NDVI maps with 15-day intervals and 1/12 degree spatial resolution, which is currently the longest NDVI dataset and is suitable for many long-term analyses [3, 24]. To minimize the effects of sensor changes among the NOAA satellites and orbital decay on the quality of the AVHRR data, calibration was performed using Bayesian methods and high quality, well-calibrated NDVI data from the sea-viewing wide field-of-view sensor (SeaWiFS). The GIMMS NDVI3g dataset has better quality and a much longer record than its predecessor, namely the GIMMS NDVI dataset (1981–2006).

We used gridded climate data from the MERRA-2 (modern era retrospective analysis for research and applications) [25] from NASA's global modeling and assimilation office (GMAO) (http://gtnao.gsfc.nasa.gov/research/merra/). MERRA was developed utilizing the Goddard earth observing system (GEOS) atmospheric model and data assimilation system (DAS) and makes use of conventional observations and satellite radiance data from both operational and research instruments. The conventional observations include measurements of standard atmospheric variables made by instrumentation on weather stations, balloons, aircraft, ships, buoys, and satellites. MERRA reduces the uncertainty in precipitation and inter-annual variability by improving the representation of the water cycle in reanalyses. The long-term MERRA-2 reanalysis dataset (1979–present) provides meteorological data with a spatial resolution of 0.5° × 0.625°. In this study, daily precipitation, temperature, and photosynthetically active radiation (PAR) data from MERRA for the period 1982–2015 were used. To match the spatial and temporal resolutions of the GIMMS NDVI3g dataset, the daily climate data were resampled to 8 km and then aggregated to a 15-day time step.

According to the MODIS land cover map (MCD12Q1, 500 m spatial resolution), we identified the five dominant vegetation types in the global semi-arid regions: forests (5.3%), shrublands (22.6%), savannas (15.3%), grasslands (33.6%), and croplands (21.1%) (figure 1). The percentage that each broad vegetation type accounted for the semi-arid region of each continent is summarized in table S2. Grasslands are the major land cover type in semi-arid regions of North America (51.9%), Asia (43.0%), Africa (29.3%), and Oceania (Australia) (23.1%).

In this study, we used a data-driven upscaling approach to estimate gross primary productivity (GPP), net ecosystem productivity (NEP), evapotranspiration (ET), and water use efficiency (WUE) for each grid cell at 8 km spatial resolution and bi-weekly time step for global semi-arid regions over the period 1982–2015. The upscaling models were established using a boosting-like regression-tree ensemble technique. This approach has been used to quantify terrestrial CO₂ exchange at multiple spatio-temporal scales [16, 17, 20].

For each of the fluxes (GPP, NEP, and ET), we first developed a set of models with different input variables. The flux tower data were used for model training and validation, while remote sensing and climatic data sets were used as explanatory (or predictive) variables. We then identified the most important explanatory variables and determined their empirical relationships with each flux. For each flux, the final predictive model was composed of five committee models, and each committee model consisted of several rule-based models each of which assigned higher weights to the outliers of the previous model, which help generate smooth flux maps by encouraging variations in the regression stratification thresholds among models.

We used ten-fold cross-validation to evaluate the accuracy of the predictive model for each flux. Specifically, the flux data were partitioned into ten equally-sized subsamples; one subsample was retained for model validation, while the remaining nine subsamples were used to train the model; this process was then repeated nine times. We compared the model-estimated fluxes with the tower fluxes using the Pearson correlation coefficient (r) and root mean square error (RMSE). The models estimated the carbon and water fluxes fairly well, with r values of 0.89, 0.75, and 0.85 for GPP, NEP, and ET, respectively, and RMSE values of 1.13 g C m⁻² d⁻¹, 0.78 g C m⁻² d⁻¹, and 0.72 mm d⁻¹ for GPP, NEP, and ET, respectively (figure 2).

Zoom In Zoom Out Reset image size

We used the final predictive models to map GPP, NEP, and ET globally at 8 km spatial resolution and 15-day time step for semi-arid regions over the period 1982–2015. For each grid cell and each 15-day time step, the final prediction of a given flux was the average of the estimates from all committee models. We also calculated WUE (the ratio of GPP/ET) globally based on our GPP and ET estimates.

We aggregated the 15-day GPP, NEP, ET, and WUE estimates to the annual scales. We first calculated the spatial means of annul GPP, NEP, ET, and WUE by averaging the annual values of each variable across the study region (continents and biomes) for each year over the period 1982–2015. We then analyzed the magnitude and spatial patterns of annual GPP, NEP, ET, and WUE at the global scale and across different continents and vegetation types. For each grid cell, we also examined the long-term trends in GPP, NEP, ET, and WUE during the 34-year period with the linear regression method. For each variable, a positive slope (b > 0) indicates an increasing trend, while a negative slope (b < 0) indicates a decreasing trend. Finally, we assessed the effects of climate factors (precipitation, temperature, PAR) and atmospheric CO₂ on GPP, NEP, ET, and WUE using correlation analyses.

The global semi-arid annual GPP averaged over the period 1982–2015 was 628.6 ± 24.6 g C m⁻² yr⁻¹ (8.03 ± 0.31 Pg C yr⁻¹) (table S3). The total GPP of the northern hemisphere accounted for 71.2% of the total GPP of global semi-arid ecosystems; the total productivity in the southern hemisphere was relatively low, although the forests in the semi-regions in the southern hemisphere showed high productivity on a per-unit-area basis (>1600 g C m⁻² yr⁻¹) (figure 3(a)). The long-term mean GPP was 569.02 g C m⁻² yr⁻¹ (equivalent to 6.21 Pg C yr⁻¹) and 865.90 g C m⁻² yr⁻¹ (equivalent to 2.51 Pg C yr⁻¹) in the northern and southern hemispheres, respectively (figure 4(a)). During the 34-year period, the annual NEP of semi-arid regions varied from –163.1 to 343.6 g C m⁻² yr⁻¹ globally (figure 3(b)). Higher NEP was mainly observed in the central and southern portions of North America and croplands in Russia. Central Africa and northern Australia were generally carbon neutral, and eastern Australia was a weak carbon source. The global semi-arid regions had a net carbon uptake (NEP) of 9.6 g C m⁻² yr⁻¹ (0.11 Pg C yr⁻¹) (table S3). The annul NEP was –1.22 ± 11.8 g C m⁻² yr⁻¹ and 52.60 ± 10.24 g C m⁻² yr⁻¹ for the northern and southern hemispheres, respectively (figure 4(b)).

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

During the 34-year period, annual ET varied from 110 to 1100 mm yr⁻¹ (figure 3(c)). Generally, ET was higher in the areas with larger GPP. For example, the ET in the forests of central South America and croplands in India exceeded 1000 mm yr⁻¹. The long-term mean annual ET of global semi-arid regions was 463.1 ± 10.5 mm yr⁻¹ (5934.77 ± 134.4 km³ yr⁻¹) (table S3). In the northern hemisphere, the ET was relatively high (about 800 mm yr⁻¹) in the east of the Great Plains dominated by grasslands and relatively low (∼100–400 mm yr⁻¹) in other regions, especially in central Asia (figure 3(c)). The annul ET was 440.1 mm yr⁻¹ (equivalent to 4806.56 km³ yr⁻¹) and 554.37 mm yr⁻¹ (equivalent to 1609.16 km³ yr⁻¹) for the northern and southern hemispheres, respectively (figure 4(c)). The long term mean annual WUE (annual GPP/annual ET) varied from 0.5 to 4 g C kg⁻¹ H₂O (figure 3(d)). The long term mean WUE of global semi-arid regions was 1.60 ± 0.02 g C kg⁻¹ H₂O (table S3). As with annual GPP and ET, the annual WUE was the highest in croplands of the northern hemisphere and subtropical forests in the southern hemisphere.

Annual GPP increased at the rate of 1.96 g C m⁻² yr⁻¹ (p < 0.001) (figure 4(a)), especially in Europe (slope = 2.80, p < 0.001), Oceania (Australia) (slope = 2.81, p < 0.001), and Asia (slope = 3.09, p < 0.001) (table S3). The GPP increased (p-value < 0.001) by 2.19 g C m⁻² yr⁻¹ and 1.06 g C m⁻² yr⁻¹ in the northern and southern hemispheres, respectively (figure 4(a)). The areas with significant increasing and decreasing GPP accounted for about 31.2% and 6.5% of the global semi-arid regions, respectively (figure 5(a)). In the northern hemisphere, the areas with significant increasing and decreasing GPP accounted for 30.9% and 5.9% of semi-arid regions, respectively (figure 5(a)). In the southern hemisphere, 16.7% and 12.0% of the semi-arid regions exhibited significant increasing and decreasing GPP, respectively (figure 5(a)). The annual NEP showed large variations with a significant increase of 0.78 g C m⁻² yr⁻¹ (p < 0.001) (figure 4(b)). The grid cells with significant increasing and decreasing trends in NEP accounted for 29.4% and 6.5% of the global semi-arid regions, respectively (figure 5(b)). The annul NEP increased by 0.82 and 0.65 g C m⁻² yr⁻¹ in the northern and southern hemispheres, respectively (figure 4(b)). In the northern hemisphere, 30.0% and 4.5% of the semi-arid regions had significant increasing and decreasing NEP, respectively; in the southern hemisphere, 17.9% and 10.9% of the semi-arid regions showed significant increasing and decreasing NEP, respectively (figure 5(b)).

Zoom In Zoom Out Reset image size

Over the 34 years (1982–2015), the globally-averaged annual ET showed a significant increase at the rate of 0.75 mm yr⁻¹ (figure 4(c)). The annul ET increased by 0.78 mm yr⁻¹ (p-value < 0.001) and 0.62 mm yr⁻¹ (p-value = 0.09) for the northern and southern hemispheres, respectively (figure 4(c)). The areas with significantly increasing ET accounted for 21.4% of global semi-arid regions that are mainly distributed in the Mongolia prairie, grasslands in Africa and North America, croplands in Russia, Kazakhstan, and northern China, while areas with significantly decreasing ET only accounted for 8.3% of global semi-arid regions, such as the south of North America and central South America (figure 5(c)). The globally-averaged annual WUE showed a significant increase in the same direction as GPP and ET (figure 4(d)). 21.8% and 6.8% of the global semi-arid regions showed significant increasing (e.g., croplands of North America and India) and decreasing (e.g., Asia) trends in WUE, respectively (figure 5(d)).

Overall, South America had the highest carbon sink (NEP) (0.24 Pg C yr⁻¹ or 274.4 g C m⁻² yr⁻¹), the second largest total GPP (1.43 Pg C yr⁻¹, or 1625.9 g C m⁻² yr⁻¹), and the fourth largest total ET (710.0 km³ yr⁻¹, or 809.6 mm yr⁻¹) (figure 6). The total GPP and ET in South America accounted for 17.8% and 12.0% of the total GPP and ET of the global semi-arid ecosystems, respectively. North America had the second highest annual NEP (0.05 Pg C yr⁻¹, or 27.2 g C m⁻² yr⁻¹) and the third highest annual GPP (1.08 Pg C yr⁻¹, or 537.5 g C m⁻² yr⁻¹). The other four continents were weak carbon sources or were nearly carbon neutral. The mean annual WUE was 1.25, 1.15, 1.09, and 1.08 g C kg⁻¹ H₂O for Asia, Australia, Africa, and South America, respectively.

Zoom In Zoom Out Reset image size

Over the 34-year period, the total annual GPP, NEP, and ET showed significant increases for Asia, Africa, and Australia; both GPP and ET significantly increased for Europe; both GPP and NEP had significant increases for North America; both GPP and ET exhibited decreases for South America (figure S1). The annual WUE significantly increased in North America, Asia, Africa, and Australia by 0.0019, 0.0023, 0.0012, and 0.002 g C kg⁻¹ H₂O (p < 0.001), respectively. In the croplands of northeastern North America and India, annual WUE increased at a slightly higher rate (>0.015 g C kg⁻¹ H₂O). In contrast, annual WUE showed decreasing trends in some areas of central Asia grasslands and northeastern China grasslands. Although the trend in WUE was not statistically significant in most areas, the globally-averaged annual WUE significantly increased (0.0018 g C kg⁻¹ H₂O) (figure 4).

Among the five land vegetation types, the forests had the highest annual GPP (1670.34 ± 22.62 g C m⁻² yr⁻¹, or 1.72 Pg C yr⁻¹) which accounted for 21.4% of global semi-arid GPP (figure 7). Forests also had the highest carbon sink (271.61 g C m⁻² yr⁻¹ or 0.28 Pg C yr⁻¹). The GPP and NEP of forests showed insignificant trends over the 34-year period (figure S2). The savannas had the second highest annual GPP (964.4 g C m⁻² yr⁻¹) which accounted for 10.1% of global semi-arid GPP (figure 7). The NEP, GPP, and ET of savannas all significantly increased (figure S2). Croplands had intermediate GPP and ET (figure 7). The NEP, GPP, and ET of croplands all significantly increased (figure S2). The annual mean GPP of shrublands was 0.30 Pg C yr⁻¹ (or 512.18 g C m⁻² yr⁻¹) and only accounted for 3.7% of the total annual GPP of the global semi-arid areas (figure 7). The shrublands were weak carbon sinks (40.46 g C m⁻² yr⁻¹, equivalent to 0.02 Pg C yr⁻¹). Both NEP and GPP of shrublands significantly increased (figure S2). Among the five land vegetation types, grasslands had the lowest annual GPP (360.39 ± 11.67 g C m⁻² yr⁻¹) on a per-unit-area basis, while the total annual GPP of grasslands (∼2.31 Pg C yr⁻¹) accounted for 28.8% of the global semi-arid GPP (figure 7). The GPP, NEP, and ET of grasslands increased by 0.78 g C m⁻² yr⁻¹, 0.30 g C m⁻² yr⁻¹, and 0.78 mm yr⁻¹, respectively (figure S2). The grasslands were nearly carbon neutral with a mean annual NEP of –5.51 g C m⁻² yr⁻¹ (equivalent to –0.04 Pg C yr⁻¹). The grasslands area accounted for 59.9% of the semi-arid area in the Great Plains (US and Canada), 23.1% in Australia, 66.5% in Central Asia, and 43.0% in Central Africa and were weak carbon sinks (Central Asia, Great Plains, Central Africa) or sources (Australia) (figure S3).

Zoom In Zoom Out Reset image size

The WUE differed among vegetation types in the semi-arid regions (figure 7). On average, forests had the highest annual WUE (2.39 g C kg⁻¹ H₂O), followed by savannas (1.50 g C kg⁻¹ H₂O), croplands (1.33 g C kg⁻¹ H₂O), and shrublands (1.13 g C kg⁻¹ H₂O); grasslands had the lowest WUE (0.95 g C kg⁻¹ H₂O). The WUE of shrublands, savannas, grasslands, and croplands significantly increased (figure S2). In the four typical grassland regions (i.e., Great Plains, Australia, Central Asia, and Central Africa), the WUE was relatively stable (figure S3) because both GPP and ET significantly increased. The WUE significantly increased for Australian grasslands.

The influences of climate factors on carbon and water fluxes varied among regions (figure 8). Overall, precipitation had positive influences on GPP, NEP and WUE, but had a negative influence on ET. Temperature and PAR mainly had positive influences on NEP, but mainly negative influences on GPP, ET and WUE. The influences of PAR on GPP and ET showed similar spatial patterns.

Zoom In Zoom Out Reset image size

We further examined the integrated effects of three climate variables on GPP, NEP, ET, and WUE across the global semi-arid regions (figure 9). Precipitation was the most important variable influencing GPP (figure 9(a), table 1). About 50.2% of the semi-arid region was significantly influenced by precipitation, and precipitation alone affected 16.4% of the region. The semi-arid regions in Australia, and most grasslands in Africa, China, Mongolia and Kazakhstan, were significantly affected by precipitation. PAR was the second most important factor affecting GPP. About 38.0% of the semi-arid region was significantly influenced by PAR, and PAR alone affected 9.6% of the region. The PAR-influenced regions were mainly distributed in parts of South America (forested areas between 0° and 30 °S) and western and central Asia (especially in India dominated by cropland), and the influence of PAR on GPP was mainly negative (figure 8(a)). Temperature affected GPP for 33.4 % of the global semi-arid regions, while temperature alone affected only 5.7% of the regions. Temperature mainly positively influenced GPP in croplands and forests of Russia and the western portions of Asia and Europe. The combined effects of precipitation and temperature, precipitation and PAR, temperature and PAR, and the three factors together accounted for 8.9%, 9.8%, 3.6%, and 15.2% of the semi-arid regions, respectively. The combined influence of temperature and PAR was prominent in the north of Asia, while the combined influence of precipitation and temperature was prominent in most regions of Australia. In North America, eastern Europe and central Africa, the combined influence of the three factors was relatively complex likely due to the diversity of vegetation types. In general, besides precipitation, GPP in the northern hemisphere was more affected by PAR, while GPP in the southern hemisphere was more affected by temperature.

Zoom In Zoom Out Reset image size

Overall, the three climate factors had almost equally important effects on NEP (figures 8(b), 9(b), table 1). Precipitation, temperature, and PAR affected 28.5%, 27.7%, and 25.8% of the semi-arid regions, respectively. Both each factor alone and the three factors combined together affected about 10% of the global semi-arid regions. In India, Russia, most areas of North America, and southern Australia, PAR alone had a significant impact on NEP by affecting 10.6% of these regions. PAR and temperature had positive effects on NEP in most regions and negative effects in the eastern part of North America and the grasslands of northern Kazakhstan. In the grasslands of northern China, precipitation had strong independent positive influence, while in the grasslands of Kazakhstan, the independent influence of temperature and precipitation was dominant. In the Great Plains, the forested areas in central South America, and the northern parts of Australia, the combination of temperature and precipitation affected NEP for 7.7% of these regions.

PAR was the most important variable influencing ET (figures 8(c), 9(c), table 1). 44.6% of the global semi-arid regions was significantly influenced by PAR; PAR alone affected 23.2% of the regions. ET was significantly affected by PAR alone in most semi-arid regions in Asia and Europe (table 1), most grasslands in North America, eastern Australia, and most forests of South America. The combined influence of temperature and precipitation was also important and affected 12.6% of the global semi-arid regions such as forests of central South America and parts of grasslands in Mongolia, Kazakhstan, Africa, and North America. The combined effect of temperature, precipitation and PAR on ET accounted for 21.3% of the global semi-arid regions.

Overall, precipitation had the largest impact on WUE (figures 8(d), 9(d), table 1). 40.9% of the global semi-arid regions was positively influenced by precipitation, and precipitation alone affected 12.7% of the regions mainly including grasslands of Kazakhstan and North America. The influence of PAR and temperature alone accounted for about 30.2% and 27.5% of the global semi-arid regions, respectively, and the effects were mainly negative. In India and western Russia, PAR alone had a significant effect on WUE. In the central portions of South America and Africa, the combined influence of temperature and precipitation was prominent. The combined influence of precipitation and temperature was mainly distributed in the north of Australia, while the combined influence of PAR and precipitation was mainly observed in the south of Australia. The combined influence of precipitation and temperature, precipitation and PAR, temperature and PAR and the three factors together accounted for 7.7%, 8.1%, 2.3%, and 10.6% of the global semi-arid regions, respectively (figure 9(d)).

Atmospheric CO₂ had mainly positive effects on carbon fluxes and WUE, especially on NEP (figure 8, table 1). The negative effects of CO₂ on GPP and ET were mainly observed in South America and some areas of central and southern North America, western edge of Asia, and central Africa. The influence of PAR and CO₂ on ET was generally consistent with that of temperature. Atmospheric CO₂ had negative effects on WUE in some parts of central Asia (figure 8(d)).