The monthly grid data of the temperature and precipitation series from 2000 to 2014, with a spatial resolution of 0.5∘, were collected from the Climatic Research Unit (Univertisy of East Anglia Climatic Research Unit, 2015). The radiation and soil moisture data series were issued by the Global Land Data Assimilation System (GLDAS-1), with a spatial resolution of 0.25∘ (http://gdata1.sci.gsfc.nasa.gov/daac-bin/G3/gui.cgi?instance_id=GLDAS025_M). The depths of the four soil layers are 0–10, 10–40, 40–100, and 100–200 cm. The quality of the GLDAS data set was assessed against available observations from multiple sources (Zhang et al., 2008; Chen et al., 2015).

The monthly data of the Palmer Drought Severity Index (PDSI), with a spatial resolution of 2.5∘, was available at http://www.cgd.ucar.edu/cas/catalog/climind/pdsi.html. As an indicator of land-surface moisture conditions, PDSI has been widely used for the routine monitoring and assessment of global and regional drought conditions. Generally, a lower PDSI implies a drier climate. The global dry areas were defined as PDSI < -3.0, while the wet areas were defined as PDSI > +3.0 (Dai et al., 2004).

We used the Global Land Cover Characterization data from the International Geosphere–Biosphere Program (IGBP) in 2000 (http://nsidc.org/data/ease/ancillary.html#igbp_classes), along with MODIS in 2000 and 2013 (http://modis.gsfc.nasa.gov/data/dataprod/mod12.php). From these data, a routinely integrated classification of land use/cover change (LUCC) characteristics was obtained based on the feature fusion processes.

The global 1 km NPP data sets (2000–2014) are from MOD17. NPP estimations are typically model-based and biogeochemical, and are generated from a larger set of simulated C fluxes between the atmosphere and terrestrial ecosystems (Ito, 2011). A better agreement of MODIS and terrestrial NPP estimates allows the use of MODIS in large-scale estimates (Neumann et al., 2015).

The MODIS evapotranspiration data sets (2000–2014) from MOD16 are estimated using the Mu et al. (2011) improved ET algorithm over Mu et al.'s (2007) previous paper. Based on the energy-balance theory and the Penman–Monteith equation, the required MODIS data inputs for the ET algorithms include daily meteorology (temperature, actual vapor pressure, and incoming solar radiation) remotely sensed land cover, FPAR/LAI, and albedo (Friedl et al., 2010; Myneni et al., 2002).

We unified the spatiotemporal resolution of these data from different sources, based on resampling (nearest neighbor interpolation) and reclassification techniques, and combined the data from the different source data sets to form comprehensive records.

The spatial patterns of global NPP from 2000 to 2014 showed a steadily decreasing trend from the equator to the Arctic and Antarctic (Fig. 1c). Overall, the interannual series of NPP increased moderately at a rate of 0.06 PgC yr-2 over the past 15 years, and also shows different changes in the Northern and Southern hemispheres. While NPP in most parts of the NH increased (Fig. 1a), it decreased in most parts of the SH (Fig. 1b). Specifically, in the NH, 64 % of vegetated land area experienced increased NPP, including large areas of North America, western Europe, India, and eastern China. Regions with decreased NPP include eastern Europe and higher latitudes of central and west Asia. In the SH, decreased NPP accounted for about 60.3 % of vegetated land area, mainly concentrated in South America, south Africa, and western Australia. Furthermore, in the equatorial regions, Amazon rainforests had significantly decreased NPP, whereas African rainforests experienced an increasing trend (Fig. 1c). Because tropical rainforest NPP accounts for a large proportion of global NPP, decreases in SH NPP partially counteracted the increases in NH NPP.

When we combined global LUCC characteristics, the results showed that shrubland has the greatest potential increasing trend of NPP (16.5 gC m-2 yr-1) compared to other biomes, followed by grassland (12.5 gC m-2 yr-1). This may be related to the expansion of woody vegetation over the past 15 years. In the Arctic tundra (Hughes et al., 2006) and lower latitudes in arid environments (Chen et al., 2014; Li et al., 2015), experimental studies provided clear evidence that climate warming is sufficient to account for the expansion of shrubs.

Changes in vegetation albedo and emissivity exert feedback on climate, which is especially obvious in ET (Field et al., 2007). The viability of vegetation cover can substantially modulate available surface energy and partition that energy into sensible and latent heat fluxes (Matsui et al., 2005). Increased vegetation productivity and climate change may promote the rises of global ET. The average mean of estimated global annual ET is 518.6 mm yr-1, with an interannual trend of 0.46 mm yr-2. Figure 1 shows that the spatiotemporal changes of global ET are consistent with NPP variations, especially in the NH. Furthermore, where their association is less than that in the NH, NPP and ET in the SH have much higher variability. Specially, the increased rate of land ET in the NH (0.61 mm yr-2) is faster than that in the SH (0.41 mm yr-2). Anomalous warming indicates a general prospective acceleration or intensification of the global hydrological cycle and thus an alteration in the process of ET, but dry conditions have caused a reduction in vegetation productivity and a near cessation of ET growth in the SH. The spatial inconsistency in the SH mainly occurred near the equator, e.g., southern African rainforests (Fig. 1b and d). These regions have high values of average annual precipitation and stronger variability of precipitation than elsewhere, causing greater changes to ET and its components (land-surface evaporation, canopy evaporation, and transpiration). In the context of warming, places where the interannual variability of NPP is small, the ET component of land-surface evaporation will increase. In contrast, in areas with large interannual variability of NPP, such as shrubland and grass-dominant regions, the ET components of land-surface evaporation will decline and transpiration increase. Vegetation generally promotes land–atmosphere water exchange via transpiration through a biological process, changing soil moisture conditions and affecting the land–atmosphere feedback.

Water, temperature, and radiation interact to impose complex and varying limitations on vegetation activities in different parts of the world, thus also affect ET partly. To understand why NPP variations in the Northern and Southern hemispheres respond differently to climates, we first estimated the spatial trends of climatic control factors and then analyzed the complex multiple climatic constraints to plant growth. A comprehensive interpretation of interactive climatic controls on plant productivity showed that water, temperature, and radiation are the key factors affecting vegetation growth. Globally, growth was most strongly limited by water availability on 40 % of the Earth's vegetated surface, while temperature limitations exerted the main controlling influence on 33 % of the surface, and radiation on 27 % (Nemani et al., 2003).

From 2000 to 2014, overall trends of average annual temperature (0.007 ∘C yr-2) and precipitation (0.84 mm yr-2) experienced worldwide increases while showing different temporal change patterns (Fig. 2a and b). Eastern Europe, South America, southern Africa, and western Australia experienced warming combined with decreased precipitation, whereas southeast North America, western Europe, east Russia, and African rainforests experienced warming combined with increased precipitation. Meanwhile, net radiation increased in the equatorial tropics and arid regions in northwestern China, but decreased in the Arctic and Antarctic (Fig. 2c). The Palmer Drought Severity Index (PDSI) is a widely used index that correlates with soil moisture during warm seasons (Palmer, 1965; Dai, 2013). Global PDSI decreased at a rate of -0.04 yr-2 from 2000 to 2014, suggesting an increased risk of drought in the 21st century. Drought develops with periods of low-accumulated precipitation and is exacerbated by high temperatures. Warming-induced drying resulted from increased ET and was most prevalent. The spatial trend of PDSI shows that the eastern and northern coasts of North America, along with the African continent, Eurasia, and southern South America, exhibited obvious drought trends in 2000–2014. Northern China, parts of Mongolia, and western Russia near Lake Baikal also experienced a drying trend (Fig. 2d).

We analyzed partial correlations between NPP and temperature (T), precipitation (P), net radiation (Rn), and PDSI during growing seasons to determine their respective contributions across different regions (Fig. 3, Table 1).

In high latitudes of the NH (> 47.5 ∘N), temperature has a positive correlation with NPP (R = 0.6). Significant warming generally lengthens growing seasons for vegetation and promotes plant growth in tundra regions, so the recent warming in this region has increased NPP (Fig. 3a). In areas of high elevation such as the Tibetan Plateau (which is similar to high latitudes), temperature is the dominant control factor in vegetation growth. Climate changes have eased multiple climatic constraints to plant growth through earlier springs in some high latitudes, somewhat beneficial to plant growth, but the continuous warming may offset these benefits. For northern middle and low latitudes (< 47.5 ∘N), where large areas are classified as having an arid climate, vegetation is short-rooted. NPP has a significant correlation with P (R = 0.7, p < 0.05) (Fig. 3b) and is also correlated to Rn (Fig. 3c).

Equatorial Amazon rainforests experienced significantly decreased NPP, whereas African rainforests exhibited an increasing trend. These changes in equatorial regions are mainly related to the warming in the Amazon along with increasing precipitation in African rainforests. High temperatures caused higher rates of ET, generally reducing soil water availability for vegetation in the Amazon.

In the SH, we noted a significant correlation (R = 0.7, p < 0.05) between NPP and PDSI (Fig. 3d). The warming trend induced a much higher evaporative demand and led to a drying trend, except for the aforementioned increased precipitation in African rainforests. The PDSI in African rainforests also showed a slight increasing trend. A high T value can increase both vapor pressure deficiency and moisture deficit and lead to a dryer environment. The general drought event across the SH, which was induced by extreme heat and a precipitation deficit, has resulted in a net water availability reduction, ultimately reducing NPP. Hence, warming-associated drying directly caused the significant decreasing trend of NPP in the SH.

With respect to the impact of drought on the world's ecosystems, studies have been limited regarding the contribution of vegetation and terrestrial water cycle components to drought variations (Falloon et al., 2012; Teuling et al., 2013). However, the present lack of high-quality and long-term records of ET limit the forecasting of drought under climate change circumstances.

We used potential evapotranspiration (PET) as a surrogate measure of atmospheric moisture demand. PET is defined as the maximum quantity of water capable of being evaporated from soil and transpired from vegetation, whereas ET is the actual evaporation from water and soil, as well as transpiration from vegetation. Penman (1948) stated that ET had a proportional relationship with PET, and Bouchet (1963) hypothesized that a complementary feedback mechanism exists between ET and PET in water-limited regions. Overall, our investigation indicated that there is a proportional relationship between ET and PET in humid regions and a complementary one in arid regions (Figs. 1d and 4a). PET, as a surrogate measure of atmospheric moisture demand, has combined impacts of temperature, solar radiation, vapor pressure, and wind speed on its interaction with NPP (Fig. 4b). Over the past 15 years, global PET showed an increasing trend of 1.72 mm yr-2, while global P increased at a rate of 0.84 mm yr-2. However, precipitation increases cannot offset evaporative demand, indicating a potential moisture deficit for water supplies constrained by ET. In other words, P is mostly being lost to ET rather than being allocated to other components of the energy and water cycles (Zhang et al., 2015).

Various factors, including vegetation, affect the intensity and spatial variation of drought. Vegetation generally promotes land–atmosphere water exchange via transpiration, changing soil moisture conditions and affecting the land–atmosphere feedback. Available soil moisture is defined as the amount of water a plant can access in its root zone. Thus, spatial and temporal variations in soil moisture are closely related to vegetation growth (Davis and Pelsor, 2001; Yang et al., 2010). Figure 5 illustrates the worldwide decrease in soil moisture of four layers (0–10, 10–40, 40–100, and 100–200 cm). Soil moisture is an important sensor for measuring superficial wetness and dryness levels, and generally reflects the dryness and wetness of climate. Water vapor via transpiration can lead to increased regional atmospheric humidity over the short term, but preserve less water in the soil. Vegetation growth causes increased soil moisture evaporation, thus reducing the amount of soil water storage. This, in turn, accelerates reductions in soil moisture caused by warming. Especially in the areas where with less moisture conditions, as land-surface temperatures rise, increases in precipitation are insufficient to offset increases in evaporative demand, indicating a potential moisture deficit for water supplies constrained by ET. If low levels of soil moisture persist for long enough, reductions in vegetation cover and vigor can occur. This leads to soil water loss and reduced vegetation growth, along with higher risk of ecological drought (Meng et al., 2014; Zhang et al., 2015). Once vegetation suffers persistent drought, the vegetation biomass will rapidly decline and further intensify the ecological drought.