As a proxy for vegetation and ecosystem productivity, we use a new LAI dataset based on the third generation Global Inventory Modeling and Mapping Studies NDVI product (GIMMS NDVI3g) [27–29]). The GIMMS NDVI3g dataset is derived from the Advanced Very High Resolution Radiometer and is available at 8 km horizontal resolution every 10–15 days from 1981. GIMMS processing includes corrections for sensor calibration, volcanic aerosols, and other sensor degradation and contamination issues. It is the only continuous and up-to-date global NDVI dataset that is continually assessed and validated [30]. Globally, trends in the GIMMS NDVI3g compare favorably with trends derived from Landsat [31] and MODIS NDVI [32]. GIMMS is generally considered the best dataset to use for the analysis of long-term trends [30,31]. The GIMMS NDVI product has been used previously to diagnose and detect land degradation and greening trends (e.g., [11,30,33,34])

LAI is defined as the green leaf area on the ground, expressed in units of m² leaf area per m² of ground. The derived LAI product [35] is generated based on a neural network algorithm that uses the GIMMS NDVI3g and an improved version of Collection 5 Terra MODIS LAI, the MODIS BNU product (Beijing Normal University version) [36]. This new LAI product (LAI3g) compares favorably with both ground based measurements of LAI and independent satellite estimates from the SPOT VEGETATION sensor [35,37], and also shows expected responses to temperature and precipitation variability [35]. For this study, we analyze the LAI3g dataset rather than the original NDVI3g data. LAI represents a variable that can be readily interpreted within an ecological context, while NDVI itself is a ratio of spectral reflectances that is a proxy for vegetation activity. As noted above, this LAI product has been verified using independent ground and satellite based LAI data. Our decision to analyze LAI instead of NDVI is also supported by precedent in the literature, including studies analyzing satellite vegetation trends using variables derived from NDVI, instead of NDVI directly (e.g., [38–40]).

We use maximum annual values of LAI (LAImax) as a proxy for standing biomass and peak annual aboveground productivity. Maximum annual NDVI/LAI has been used previously to investigate long-term greening and browning trends (e.g., [24,41,42]). Alternative approaches have used NDVI/LAI integrated annually or over the growing season (e.g., [11]). Objectively defining the growing season using remotely sensed vegetation products can be difficult, however, especially in areas where the timing and length of the growing season varies significantly from year to year [41]. Additionally, neither proxy considers belowground carbon allocation, i.e., root/shoot ratio [43].

Given that maximum and integrated values of NDVI/LAI are often strongly and positively correlated (e.g., [41]), we believe LAImax provides the most straightforward proxy indicator of aboveground biomass and productivity that is appropriate for our analyses. For the Northern Hemisphere, we calculate LAImax based on the year from January to December; for the Southern Hemisphere, we use July to June. We calculate LAImax at each 8 km resolution grid cell, and then spatially average all LAImax cells to create a new dataset at 0.5° spatial resolution, the same resolution as our climate and land cover datasets.

Three estimates of the modern and historical distribution of human land use and land cover (including pastures) are available in the literature [44]: (1) Houghton (2008) (HH, [45]); (2) HYDE 3.1 (HYDE, [46]); and (3) an updated version of Ramankutty and Foley (1999) (RF, [47]). Land cover datasets are from Meiyappan and Jain (2012) [44], and were obtained from [48]. For pasture land, substantial differences can be seen in the distribution and intensity of this land use type across the three datasets (Figure 1). The HH estimate represents the major outlier, showing significantly lower pasture areas than both HYDE and RF. The HH calculation assumes that, outside of Latin America, all pastures are derived from grasslands and forests are never converted to pastures, leading to a severe underestimation of pasture land extent. Globally, pasture estimates from HYDE and RF are similar, although significant differences are still apparent for many regions, including Africa, Australia, the Middle East, and Asia (Figure 2). Differences between HYDE and RF are primarily due to the underlying data used in the land cover reconstruction. RF uses a census report with significantly lower pasture areas than the FAO estimate [49] used in HYDE, leading to much more widespread and intensive pasture distributions in HYDE relative to RF. Since HYDE represents the most recent and up-to-date estimate of modern and historical pasture areas, we use this dataset to discriminate between pasture and non-pasture regions in our analysis. Specifically, we distinguish between cells that are natural (<10% pasture), low intensity pasture (≥10%–<50% pasture), and high intensity pasture (≥50% pasture).

Globally, the most extensive pasture regions are concentrated in the subtropics of both hemispheres and the mid-latitudes of the Northern Hemisphere (Figure 3(a)). Local minima occur in the tropics, where the landscape is dominated by tropical forests, and around 20°N, corresponding to the latitudinal band containing the Sahara desert and much of monsoon Asia, where the landscape is dominated by cultivated agriculture. Pastures represent the dominant fraction of total vegetated areas in much of the Southern Hemisphere extra-tropics (Figure 3(b)), as well as significant fractions of the total vegetation in the Northern Hemisphere tropics (0°N–20°N) and mid-latitudes.

For climate data, we use an updated version of the monthly climate grids (version 3.1) available from the Climate Research Unit (CRU) at the University of East Anglia [50]. These data, representing monthly precipitation and temperature, are available for 1901–2009, and cover all global land areas (excluding Antarctica) at 0.5° horizontal resolution. Gridded values are based upon a statistical interpolation from individual weather stations. For comparisons with LAImax, we convert the monthly data to annual means (for temperature) and annual sums (for precipitation), using seasonal intervals consistent with the calculation of LAImax (January–December for the Northern Hemisphere; July–June for the Southern Hemisphere).

Spatially, we restrict our analysis to all latitudinal bands between 60°S and 60°N, within which nearly all of the pasture areas are contained. Temporally, we restrict our analyses to the common time periods available from all datasets (1982–2008) and use land cover from the HYDE dataset for 2008. To quantify climate controls on pasture productivity, we calculate Spearman’s rank correlation coefficient (ρ) between the linearly detrended climate variables (annual temperature and precipitation) and LAImax. Spearman’s ρ is defined as the Pearson correlation coefficient (r) of two ranked variables, is bounded between −1 and +1 (same interpretation as Pearson’s r), and is less sensitive than Pearson’s r to outliers. To investigate temporal trends in LAImax, we calculate linear trends (1982–2008) using three methods: a best fit ordinary least squares linear regression, Spearman’s ρ, and Kendall’s τ. Kendall’s τ is a non-parametric test (as is Spearman’s ρ) for rank correlation between two variables, calculated as the difference between the number of concordant and discordant pairs in the ranked bivariate data (i.e., LAImax versus time), divided by the total number of possible pair combinations of the ranked data. As with the Pearson’s and Spearman correlation coefficients, Kendall’s τ is bounded between −1 and +1. A value of +1 indicates perfect agreement (i.e., correlation) among the two rankings; a value of −1 indicates perfect disagreement (i.e., anti-correlation). We consider temporal trends in LAImax to be robust if trend estimates from all three tests pass a significance threshold of p ≤ 0.05. To minimize the influence of cultivated crop areas, we ignore any cells containing greater than 30% crop area, using crop fractions from the HYDE dataset. To compare trends spatially, we additionally stratify the data based on latitudinal bands: the Southern Hemisphere extra-tropics (60°S–20°S), the tropics (20°S–20°N), and the Northern Hemisphere extra-tropics (20°N–60°N).

Precipitation is the dominant climate control on inter-annual variability of LAImax in pasture cells, showing widespread and significant positive correlations, especially in semi-arid regions (Figure 4(a)). Significant correlations with temperature are less extensive and mostly negative (Figure 4(b)). Regions exhibiting negative temperature correlations generally overlap with areas of significant positive precipitation correlation, suggesting that warmer temperatures may cause increased moisture stress. Additionally, the regions correlated with temperature are generally not core tropical regions where plants are more likely near their temperature optimum, i.e., areas where we might expect a negative relationship with temperature because of increases in respiration or photorespiration [51]. Inter-annual variability in pasture LAImax therefore appears to be limited primarily by moisture, rather than by temperature.

Globally, 23.4% of cells with ≥10% pasture cover have significant temporal trends for 1982–2008, based on our three trend test significance criteria. Similar results are found if we calculate trends on the residuals after the influence of annual precipitation on LAImax has been removed using a simple linear regression (21.7% of pasture cells). Removing the influence of temperature has little effect, consistent with the generally weaker and less widespread temperature correlations (not shown). The absolute magnitude of the trends (both negative and positive) is reduced when the precipitation signal is statistically removed (Figure 5), implying some contributing effect of precipitation toward long-term greening and browning in pastures. Averaged across cells, however, the trend is only slightly diminished: +0.0162 and +0.0155 m²/m² per year before and after removal of the precipitation signal, respectively. For cells with significant LAImax trends, the precipitation detrending decreases the magnitude of the LAImax trends by an average of 12.9%. Given the similarity in magnitude and direction of the LAImax trends with and without removal of the precipitation signal, this suggests that recent precipitation trends are only a minor contributor to productivity trends in pasture areas over the last 30 years. Comparisons between LAImax trends and fractional pasture coverage show no significant global relationship between the magnitude of the LAImax trends and time averaged (1982–2008) pasture intensity. We also found no relationship between LAImax trends and temporal changes in pasture cover intensity over this interval.

Pasture and natural vegetation regions show a mixture of browning and greening trends (Figure 6, no significance masking). For natural vegetation (Figure 6(a)), the most cohesive regions of greening are found in tropical regions of South America and Africa, while the high northern latitudes show a mix of greening and browning trends. For pasture areas (Figure 6(b)), general trends are towards increasing LAImax, with localized areas of browning in southeast Australia, Mongolia and central China, western North America, east Africa, southern Africa, and southern South America. However, much of the browning in both the natural and pasture vegetation cells disappears when we apply our three significance test filtering (Figure 7), and greening dominates the significant trends in both natural vegetation and pasture systems. Strongest trends towards greening in pasture areas are found in West Africa and South America, while productivity declines are most apparent in Australia, western North America, and central China and Mongolia.

Changes in LAImax vary across latitudinal bands and between natural vegetation, low-intensity pasture, and high-intensity pasture areas. This can be seen in the cumulative distribution functions (CDFs) for each latitudinal band and vegetation classification (in Figure 8). Consistent with the global analysis, greening trends dominate among all three vegetation types and across all three latitudinal bands. In the Southern Hemisphere extra-tropics (60°S–20°S), greening trends are higher in pasture cells versus regions with natural vegetation. This may reflect a sample bias, since natural vegetation makes up a relatively small fraction of the vegetated areas in the Southern Hemisphere extra-tropics (Figure 3(b)). The same general pattern holds for the tropics (20°S–20°N), where pasture cells are greening much more than natural vegetated areas. A look at the cdfs (Figure 8(c,d)) suggests that differences in the tropics are driven primarily by differences among the most rapidly greening cells. The situation reverses in the Northern Hemisphere extra-tropics (20°N–60°N), where cells with natural vegetation are increasing their greenness at a much higher rate than pastures. Except for the tropics, differences in the CDFs are generally reflected in the median trends for each vegetation type (Figure 9). Greatest differences across vegetation types are seen in the Northern Hemisphere extra-tropics, where cells with natural vegetation and significant trends are greening at 2–4 times the rate of pasture cells.

These differences can be confirmed quantitatively by comparing the trend distributions using a two sample Kolmogorov–Smirnov (KS) test (a non-parametric goodness of fit test applied to the cdfs) and a two sample Wilcoxon rank-sum (WR) test for differences in the median. For the tropics and Southern Hemisphere extra-tropics, the KS and WR tests confirm that low and high intensity pasture areas are greening at a greater rate than natural vegetated areas (p < 0.001). For the Northern Hemisphere extra-tropics, the KS and WR tests indicate that natural vegetated areas are greening at a higher rate than low or high intensity pasture areas (p < 0.001). These results and their significance levels are similar whether these tests are applied to the cdfs with or without the three trend test significance masking.