Greening of the land surface in the world’s cold regions consistent with recent warming

Abstract

Global ecosystem function is highly dependent on climate and atmospheric composition, yet ecosystem responses to environmental changes remain uncertain. Cold, high-latitude ecosystems in particular have experienced rapid warming¹, with poorly understood consequences^2–4. Here, we use a satellite observed proxy for vegetation cover – the fraction of absorbed photosynthetically active radiation⁵ – to identify a decline in the temperature limitation of vegetation in global ecosystems between 1982 and 2012. We quantify the spatial functional response of maximum annual vegetation cover to temperature and show that the observed temporal decline in temperature limitation is consistent with expectations based on observed recent warming. An ensemble of Earth system models from the Coupled Model Intercomparison Project (CMIP5) mischaracterized the functional response to temperature, leading to a large overestimation of vegetation cover in cold regions. We identify a 16.4% decline in the area of vegetated land that is limited by temperature over the past three decades, and suggest an expected large decline in temperature limitation under future warming scenarios. This rapid observed and expected decline in temperature limitation highlights the need for an improved understanding of other limitations to vegetation growth in cold regions^3,4,6, such as soil characteristics, species migration, recruitment, establishment, competition, and community dynamics.

Main

A global increase in green vegetation has been observed over recent decades^7,8 and widely attributed to both direct and indirect anthropogenic influences, primarily elevated atmospheric CO₂, but also to changes in climate, nitrogen deposition, and land use change ^8–11. The global greening has contributed to changes in biophysical feedbacks such as evapotranspiration and albedo¹², along with an increased terrestrial carbon sink¹³, which together have served to slow the rate of global warming². Uncertainties remain, however, regarding the drivers of the reported greening, and their spatial variation, and thus the likelihood of continued greening.

Ecosystems in cold regions in particular have exhibited rapid increases in green vegetation ^3,7. Plot scale evidence suggests a link to warming^4,6, but direct attribution of observed regional trends to environmental changes has been elusive. Regional studies have relied on multi-factorial simulation experiments with global vegetation models^8,11,14. But vegetation models are known to perform poorly for high-latitude and cold-limited ecosystems^15–17, overestimating both the extent of green vegetation and the trend in recent decades¹⁷. Model results suggest a role of climate change and CO₂ in the observed greening trend^{8,10,11,14,18}, but direct attribution is hindered by model spread and uncertainty¹⁴.

Here, we develop a data-based approach^19,20, using three decades of remotely sensed estimates of the fraction of absorbed photosynthetically active radiation (fAPAR), a proxy for productive foliage cover²¹, to characterize the relationship between maximum annual foliage cover (F_max) and the summer warmth index (SWI₅, the sum of the monthly mean temperatures above 5°C). We use the approach, which was previously developed for water limited ecosystems^19,20, to identify the spatial distribution of ecosystems in which F_max is limited by temperature, and thus to track changes in temperature limitation over time. By estimating the temperature sensitivity of F_max (${\gamma }_F^T$) using spatial gradients, and comparing with expected changes in F_max due to observed changes in temperatures, we show a long-term increase in foliage cover that is consistent with the expected influence of recent warming. The observed greening occurred together with a large decline in the spatial extent of temperature limitation over recent decades, in particular for northern high-latitude ecosystems. Finally, we use an ensemble of global Earth System Models (ESMs) from the recent Coupled Model Intercomparison Project (CMIP5) to examine the impact of future projected temperature changes on the spatial extent of temperature limitation, and to assess ESM skill in estimating the sensitivity of maximum foliage cover to changes in temperature.

Our functional responses analysis of global satellite observations from the Global Inventory Modeling and Mapping Studies (GIMMS3g⁵) shows the spatial dependence of vegetation cover on temperature (Fig. 1), with low maximum foliage cover in colder regions (low SWI₅) and high maximum foliage cover in warmer regions. We examined the 95^th percentile of the distribution of annual F_max (F₉₅, %), which characterizes the maximum F_max attained globally for a given annual SWI₅. We delineated the regions where temperature strongly affected F_max, and found it depended linearly on SWI₅ under colder conditions (Fig. 1). The slope of the F₉₅ edge quantifies the temperature sensitivity of F_max (${\gamma }_F^T$, % °${GDD}_5^{-1}$), and the temperature at saturation (SWI₅ = 43.2 ± 1.36 °C, mean ± st. dev.) delineates the region where F_max is limited by temperature (Fig. 1, Fig. S1).

Fig. 1 |. The spatial relationship between temperature and vegetation cover.

Spatial relationship between the summer warmth index (SWI₅, °C) and the annual maximum fraction of absorbed photosynthetically active radiation (F_max, %), for the three-year period from 1982–1984. The 95th percentile of the distribution of F_max (F₉₅) in each 5°C bin represents the maximum attainable F_max for a given annual SWI₅ (black dashed line). A breakpoint regression applied to the 95th percentile approximates the sensitivity of F_max to spatial changes in SWI5 (red dashed line), and delineates regions where F_max is temperature limited and those that are not, as either below or above the breakpoint (vertical dashed line).

We examined changes in the slope and intercept of F₉₅ during the GIMMS3g observational record (1982–2012) and found both an increasing intercept (4.16 ± 1.47 % decade⁻¹, mean ± 95% CI) and a declining slope (−0.10 ± 0.05 % decade⁻¹) (Fig. 2) for F₉₅ relative to the baseline SWI₅ from 1982–1984. An increasing intercept indicates an increase in F₉₅ in regions that were temperature limited at the start of the measurement record, with a 26.5 ± 1.9% increase in F₉₅ in the most temperature limited regions (Fig. 2, Fig. S2). A decreasing slope implies that regions that were more temperature limited at the start of the observational record are greening faster than regions that were less temperature limited. The changes in the relationship between F_max and SWI₅ in temperature-limited regions are consistent with accelerated warming in colder regions¹ and could indicate a change in temperature limitation over time.

Fig. 2 |. Changes in temperature limitation.

a, The slope of the relationship between 95^th percentile (F₉₅) of the maximum annual fraction of absorbed photosynthetically active radiation (F_max) and the summer warmth index (SWI₅). b, The intercept of the relationship between F₉₅ and SWI₅. Error bars represent 95% confidence intervals. The red dashed lines show fitted linear regressions, with slope m (a, % °SWI₅⁻¹ yr⁻¹; b, % yr⁻¹) and statistical significance p.

We used a space for time substitution to examine whether the observed changes in F_max in temperature-limited regions are consistent with observed long-term (1982–2012) changes in temperature. To do so, we predicted the expected change in F₉₅ (F₉₅ = I + ${\gamma }_F^T$SWI₅), where I and ${\gamma }_F^T$ are estimated as the intercept and slope of the relationship between F₉₅ and SWI₅ from the start of the observational record (1982–1984), and SWI₅ is temporally dynamic throughout the observational record. The change in F₉₅ over time, predicted based on temporally-static spatial changes in F₉₅, was statistically equivalent to the observed long-term change (p < 0.01, Chow test, Fig. 3). Observed temperature changes implied a somewhat lower increase in F₉₅ in the most temperature-limited regions (19.12 ± 2.27%, Fig. 3), but with a smaller decline in the F₉₅ slope over time than observed. The concordance of the observed and expected change in F₉₅ suggests that the observed long-term increase in F_max is consistent with a response to long-term warming, though other factors such as CO₂ fertilization and nitrogen deposition could also play a role.

Fig. 3 |. Observed and predicted changes in F₉₅.

Each solid line represents the observed F₉₅ for a specific three-year period, starting from 1982–1984 (black) and ending in 2010–2012 (blue). Gray lines show F₉₅ for intermediate three-year periods. The red dashed line represents the predicted temporal change in F₉₅ based on the observed spatial sensitivity to SWI₅ and observed long-term changes in SWI₅.

Ecosystems below the temperature limitation threshold in the relationship between F₉₅ and SWI₅ (Fig. 1) represented 19.87 ± 0.67% of the total vegetated area of the extra-tropical northern latitudes at the start of the measurement record (1982–1984), and are primarily located at high-latitudes and on the Tibetan Plateau (Fig. 4). Our results indicate that the spatial extent of temperature limitation of F_max has declined by 16.35 ± 0.64% over the observational record. The release from temperature limitations was largely experienced at the southern edge of high-latitude ecosystems. We estimate 45% and 85% reductions in the temperature-limited area by 2100 for CMIP5 ensemble warming projections under Representative Concentration Pathway (RCP) 4.5 and RCP8.5, respectively (Fig. 4).

Fig. 4 |. Current and predicted changes in the relative spatial extent of temperature limited area of vegetated land.

Shaded areas in a and b represent areas where F_max indicated temperature limitation at the start of the observational record (1982–1986) but did not by the corresponding year. Bright green areas represent the change in the observational record to 2010, and other shades represent projected changes based on temperature projections from the CMIP5 models under RCP 4.5 (a) and RCP 8.5 (b). The proportion of vegetated areas that are temperature limited over time is shown in c, relative to the extent of vegetated areas that were temperature limited at the end of the measurement record (2010–2012), for the mean (dashed lines) and standard error (shaded areas) of an ensemble of 10 ESMs (Table S1) from the CMIP5 ensemble under RCP 4.5 (blue) and RCP 8.5 (green).

By focusing on the 95^th percentile, the space for time substitution utilized here minimizes the influence of other limitations to growth^19,20, thus allowing for the identification of the independent temperature sensitivity. In reality, however, multiple factors limit maximum vegetation cover in cold regions, including nutrient availability, rooting depth, permafrost dynamics, and soil moisture. We estimate that 44% of the vegetated land surface identified as temperature-limited is primarily temperature-limited, defined as being within ±10% of the potential F_max (Fig. S3, S4). This proportion suggests that the vegetation cover response to warming of the remaining 56% will likely be mediated by other factors. Indeed, the mean long-term increase in maximum vegetation cover for vegetated areas that are primarily temperature-limited was 1.48 ± 0.05 % decade⁻¹ (mean ± std. error), whereas vegetated areas limited by both temperature and other factors had smaller increases of 0.8 ± 0.03 % decade⁻¹. The predicted declines in the temperature-limited area therefore do not necessarily translate into a uniform increase in vegetation cover. For instance, warming also extends the growing season length, which may lead to earlier snow-melt, increase peak season water stress, and thus reduce vegetation cover²². In addition, the space for time approach inherently assumes a climate-vegetation equilibrium, which may lead to overestimated rates of change due to the inability of species range shifts to keep pace with warming^23,24. Finally, we note that long-term satellite records are subject to uncertainty related to orbital effects and platform changes, though such issues are expected to be lower at high latitudes²⁵.

The temperature sensitivity of F₉₅ in the examined CMIP5 ESMs spanned a large range (Fig. 5), from a positive ${\gamma }_F^T$ of 1.56 (% $°{GDD}_5^{-1}$) in the Canadian Earth System model (CanESM2), to a relatively flat temperature sensitivity in the CCSM4 and NorESM models, which both use the CLM4 land surface model (Table S1), and are the only models to include an explicit nitrogen cycle. On average, the models underestimated the observed ${\gamma }_F^T$ by 63.53 ± 50.8% (mean, standard deviation), with only CanESM2 giving a temperature sensitivity larger than that observed. The underestimated sensitivity was reflected in an overestimated F₉₅ at low temperatures (Fig. 5b, Fig. S5) of 77.48 ± 41.45%. The overestimated F₉₅ and underestimated ${\gamma }_F^T$ in the CMIP5 models are consistent with, and shed light on, previous reports that vegetation cover is consistently overestimated at high latitudes in CMIP5 models^15,16 and dynamic vegetation models¹⁷. Combined with reports of a persistent underestimation of photosynthetic capacity at high latitudes in terrestrial biosphere models²⁶, these results suggest that models of cold limited ecosystems need improvement, and call into question their utility for attribution^8,14. Our results provide a benchmark for model development, though further analysis is needed to identify the responsible processes that govern the relationship between temperature and vegetation cover.

Fig. 5 |. Earth System Model (ESM) estimates of the relationship between maximum vegetation cover and temperature.

The sensitivity (${\gamma }_F^T$) was estimated as the slope of the relationship of the 95th percentile (F95) of maximum modeled fAPAR (Fmax) and the summer warmth index (SWI5) based on ESM modeled temperatures. The observed ${\gamma }_F^T$ and intercept (horizontal bars) were estimated from the relationship between GIMMS3g Fmax observations and CRU monthly temperatures. The dashed blue line represents the mean observed value.The shaded blue area and model error bars represent the standard deviation about the mean. Model details are provided in Table S1, and individual model functional responses are presented in Fig. S5.

The greening of the Earth is a widespread phenomenon, one that models have primarily attributed to changes in atmospheric CO₂ and climate^8,11. Here we use direct observations to isolate the functional response of vegetation cover to temperature in temperature-limited regions and report an observed greening consistent with the effect of long-term temperature changes. The identified temperature sensitivity shows that growing season warmth is a dominant factor for vegetation production in cold regions, confirming previous reports of temperature controls on both spatial and temporal vegetation dynamics^27–29. Our analysis also suggests a large reduction in ecosystem temperature limitation under future warming, though other limitations will likely play a large role in mediating the extent to which high-latitude ecosystems green. For example, arctic tundra soils are nutrient poor and, through stoichiometric requirements, impose limits on potential biomass, although mineralization of previous frozen soil N may³⁰ or may not³¹ offset those limitations. Similarly, the waterlogged soils of extensive northern wetlands are unsuitable for dense vegetation, regardless of the temperature limitation. Other limitations include the limited ability of species to migrate and adapt to the distinct environmental conditions of high-latitude ecosystems^23,24. Current models need to accurately reproduce the effect of temperature on vegetation cover. That said, the expected release from temperature limitation under future warming highlights the importance of non-temperature limitations in mediating ecosystem responses to future climate change.

Methods

Satellite data

We used estimates of the fraction of absorbed photosynthetically active radiation (fAPAR) from the Global Inventory Modelling and Mapping Studies third-generation fAPAR data product (GIMMS3g⁵), available for the period 1982–2012. The data set is provided biweekly at 0.083° spatial resolution and was regridded to match the spatial resolution of the climate data. Maximum annual fAPAR (F_max) was calculated as the maximum recorded value during each year. fAPAR is closely related to the photosynthetic activity of plants, and therefore constitutes an indicator of the presence and productivity of live vegetation.

Climatic variables

Monthly fields of air temperature at a 0.5° spatial resolution were obtained from the Climate Research Unit (CRU) high-resolution gridded datasets version 3.24³². The monthly mean air temperature values were converted to annual values of summer warmth by summing all monthly temperature values above a baseline of 5 °C (SWI₅). This approach is designed to account for changes in temperatures that effect vegetation growth, whilst minimizing changes in temperatures that are too low to influence vegetation. Using a base temperature of 0 °C (SWI₀) led to the inclusion of low temperature and low fAPAR pixels that were relatively insensitive to temperature change, but did not affect the overall results (Fig. S6).

Breakpoint regression analysis

Three-year running mean F_max values were binned according to their corresponding temperature values for 5° temperature bins^19,20. For each bin, the upper and lower 95^th, 90^th and 75^th percentiles were determined for each running mean block. Breakpoint regression was applied to the 95^th percentile values (F₉₅) using multi-phase linear regression. We estimated uncertainties of fit parameters through Monte Carlo simulations of zero-mean deviates based on the Cholesky decomposition of the covariance matrix. The regression of the 95^th percentile of F_max represents the maximum F_max attainable for a given SWI₅, thus minimizing the influence of other factors, such as precipitation or aridity, on the derived responses (Fig. S8). The breakpoint of the regression identifies the region where the vegetation–temperature relationship plateaus and vegetation ceases to be temperature-limited. Note that the breakpoint temperature was relatively insensitive to the percentile used (Fig S7). We constructed time series of the slopes and intercepts of the breakpoint regression and determined linear trends for both variables, using changes in the F_max - SWI₅ relationship but keeping SWI₅ fixed to that experienced in the first three-year window of the observational record (1982–1984). As running means were used to construct the time series, non-independent running mean blocks were removed before determining the strength and significance of trends.

CMIP5 simulations

We analyzed output from 10 Coupled Model Intercomparison Project Phase 5 (CMIP5) coupled carbon-climate models (Table S1) obtained from the Program for Climate Model Diagnosis and Inter-comparison (PCMDI) Earth System Grid (ESG) (http://pcmdi9.llnl.gov/esgf-web-fe/)). The land components of these ESMs differ in their representations of vegetation types, soil properties, human disturbances, and carbon and nitrogen pools. We used model output of leaf area and air temperature at native spatial resolution, and converted to the fraction of absorbed photosynthetically active radiation using the standard conversion of Beer’s law, which accounts for the exponential decline in absorbed radiation with increasing leaf area. Values of F_max, F₉₅ and SWI₅ were calculated through a functional response analysis as with the remote sensing observations. We used historical simulations (1980–1990) for the comparison of spatial responses of F_max to SWI₅, and projections of future monthly temperatures from 2010 to 2100 under two representative concentration pathways (RCPs), 4.5 and 8.5.

Data Availability

The data that support the findings of this study are publicly available. The satellite fAPAR data are hosted on NASA NEX (see instructions at http://sites.bu.edu/cliveg/datacodes/). The CMIP5 simulation outputs are available from the Program for Climate Model Diagnosis and Inter-comparison (PCMDI) Earth System Grid (ESG) (http://pcmdi9.llnl.gov/esgf-web-fe/). The climate data used (CRU3.24) can be downloaded from https://crudata.uea.ac.uk/cru/data/hrg/. All code is available from the corresponding author upon request.