We analyze temporal patterns in various metrics of the terrestrial carbon cycle based on three independent data sources. First, we analyze data-driven NPP fields based on simulations with the satellite-constrained biogeochemical Carnegie–Ames–Stanford Approach (CASA) model (van der Werf et al., 2006) for the period of available satellite vegetation data 1982–2011 (Zhu et al., 2013). This updated and extensively validated model runs at a 0.5∘ spatial resolution on a monthly time step. NPP is a measure of the amount of carbon fixed by plants during photosynthesis and accumulated as biomass. The CASA model is conceptually relatively simple, and a number of potentially important processes and mechanisms, such as ones related to nutrients (e.g., carbon and nitrogen), are not considered explicitly (van der Werf et al., 2006). However, factors that influence vegetation productivity may be indirectly captured through the satellite-based fraction of available photosynthetically active radiation absorbed by plants (fAPAR), a key driver in the NPP CASA light use efficiency parameterization (van der Werf et al., 2006). This is demonstrated by two recent studies showing that trends in satellite-based vegetation cover are consistent with expectations of growth enhancement via the CO2 fertilization effect (Donohue et al., 2013; Los, 2013). Yet there is also new evidence suggesting that in densely forested ecosystems (tropical rainforests) fAPAR may not be fully responsive to the CO2 fertilization effect (Forkel et al., 2015). Temporally varying driver data used for the CASA simulations include, in addition to satellite-based fAPAR (fAPAR3g) (Zhu et al., 2013), also land temperature (CRU TS3.21) (Harris et al., 2014) as well as precipitation and surface solar radiation. We pay particular attention to uncertainties in these observationally based data sets and corresponding effects on NPP estimates. While land surface temperature data are considered to be relatively robust, a substantial limitation is that presently only one consistent satellite fAPAR data set exists that covers the last 3 decades (Zhu et al., 2013), which we consider a minimum record length for meaningful change point analysis. Nonetheless, to account at least partially for observational uncertainties, we evaluated different data sources for precipitation and surface solar radiation, which are known to have substantial uncertainties (Wild, 2009; Greve et al., 2014) (see Methods in the Supplement). Based on a final selection of driver data sets (one fAPAR, one temperature, three precipitation and three solar radiation; see Table S4 in the Supplement) we performed multiple CASA simulations with all possible input combinations (a total of 9 simulations) and analyzed the NPP ensemble mean with our change point methodology. We used the spread in these simulations as a measure of “observationally based” uncertainty. In a complementary analysis, we also analyzed the heterotrophic respiration (Rh) and net ecosystem production (NEP; estimated as NPP – Rh) ensemble means from these sets of CASA simulations. Owing to the lack of consistent fire observations for our study period 1982–2011 (e.g., the Global Fire Emissions Database (GFED) era starts in 1997; see van der Werf et al., 2006), corresponding effects on variability in NPP and NEP were not considered. To ensure that the carbon pools are at steady state, the CASA model was spun up for 250 years using a driver climatology based on the study period 1982–2011.

Second, we analyze a GCB for the period 1959–2011 and consisting of CO2 emissions from fossil fuel burning and cement production as well as net LUC emissions, atmospheric CO2 growth rates and oceanic uptake (Le Quéré et al., 2013). Uncertainties in these budget terms are also provided and utilized to estimate uncertainties in net land uptake and the residual land sink (through the sum of squared errors).

Third, we analyze process-based NPP, Rh and NEP data based on ensembles of nine single terrestrial biosphere models that participated in the recent TRENDY model intercomparison project (Sitch et al., 2015). Compared to CASA, the TRENDY models are substantially more complex and also run at significantly shorter time steps to resolve the diurnal cycle needed when coupled within Earth system climate models (Sitch et al., 2015). An important distinction is that in the TRENDY models vegetation characteristics (e.g., fAPAR) are simulated prognostically (unlike in CASA, where such information is inferred from satellite observations). In the TRENDY experiments (Sitch et al., 2015) the models were driven with observed climate and atmospheric CO2 data (S2 experiments) as well as with observed atmospheric CO2 data only (S1), and, in order to isolate the variability due to climate, the difference between these two experiments (S2–S1) was taken. We analyze the NPP, Rh and NEP ensemble means (based on anomalies) from the nine participating models (Community Land Model 4CN, Hyland, Lund–Potsdam–Jena (LPJ), LPJ-GUESS, ORCHIDEE, ORCHIDEE-CN, Sheffield-DGVM, TRIFFID and VEGAS; for model details see Sitch et al., 2015) and used the spread among them as a measure of “model-based” uncertainty in our change point framework. Based on a recent comprehensive evaluation against observations, it was found that most of the TRENDY models are capable of simulating the short- and long-term first-order dynamics of the terrestrial carbon cycle (Piao et al., 2013).

We apply a consistent change point methodology on the various metrics of the terrestrial carbon cycle to identify patterns of regime shifts (characterized as abrupt, substantial and sustained changes) and to contrast them with patterns showing either no or more gradual changes. We thus determine in a first step the statistical model that best fits the time series under investigation based on three options: (1) a constant mean, (2) a shift in the mean and (3) a linear trend. While there are numerous alternative statistical models (e.g., shifting trends as seen in satellite vegetation data at local to regional scales; Piao et al., 2011), our choice of these three models is based on our primary objective to identify large-scale patterns in global and continental carbon fluxes that would be consistent with the recently observed regime shift in the land carbon sink (Sarmiento et al., 2010; Beaulieu et al., 2012a).

In the “shift-in-the-mean” model, the shift is located through a change point detection algorithm that includes discrimination against a trend and the background autocorrelation (red noise) by considering all positions in a time series as a potential change point from 5 to n-5, with n being the record length (Beaulieu et al., 2012b). In a previous study, we found that, by restricting the search for change points in this manner, detection of spurious shifts at the beginning or end of a series can be avoided (Beaulieu et al., 2012a). In the change point method applied here, we also further developed the Beaulieu et al. (2012b) methodology to account for known explicit uncertainties in the time series under investigation. One important limitation here is that this statistical change point model cannot distinguish between a rather drastic shift (e.g., change from one year to the next) and a smoother shift over the span of several years. Adding additional parameters could in principle provide more information on the nature of the shift (e.g., smooth vs. abrupt), but this would also make the model prone to overfitting given the rather short time series in this study.

The most likely model among the three statistical models fitted is determined based on the Schwarz information criterion (SIC), which compares their likelihoods with a penalty for the number of parameters fitted. If the shift-in-the-mean model seems the most likely, we calculate in a second step the direction and magnitude of the shifts (subtracting means prior to and after the shift) and the corresponding P value by integrating the full uncertainty of the data using Monte Carlo simulations. To perform the Monte Carlo simulation, we draw 1000 normally distributed synthetic series having the same statistical properties as the time series of interest. A new feature is that the series are simulated with uncertainty additivity: the squared variance of each data point is added to the overall time series squared variance, and the square root of this sum provides the synthetic series variance. This therefore takes into account the explicit uncertainties in the various time series under investigation. The change point method is applied to the synthetic time series, and a SIC difference between the model with a shift in the mean and no shift is calculated for each time series. This provides a distribution for the SIC difference under the hypothesis of no shift in the mean. The P value is the estimated probability to find a SIC difference at least as extreme as the one observed, under the hypothesis of no change. This methodology assumes that the errors of the model are independent and normally distributed with a constant variance. We test normality of the residuals using the Lilliefors test, the independence is verified using the Durbin–Watson test and the constant variance is verified using a Fisher test. All tests are available and performed using MATLAB. If independence is not respected, we generate synthetic series with the same first-order autocorrelation as observed in the respective time series residuals.

Applying our change point methodology (Sect. 2.2) on data-driven global NPP fields reveals a marked spatial clustering of abrupt and sustained increases in NPP across northern Eurasia and northern Africa in the late 1980s (Fig. 1). At more regional levels, the impact of severe disturbance events such as the mountain pine beetle outbreak in the late 1990s in the temperate and boreal forests of western North America (Kurz et al., 2008) is also disclosed (via rapid and sustained decreases in NPP). A similar analysis without constraining to only statistically significant results at the grid point level implies that the coherent pattern of abrupt and sustained NPP shifts across northern Eurasia and northern Africa is spatially even more extensive (Fig. S1 in the Supplement).

This point is further illustrated when we apply our change point framework on data-driven NPP time series representative of large land regions and highlights the important role of the northern extratropics (magnitude of NPP shift: ∼ 0.7 PgC yr-1), and the northern Eurasian continent (∼ 0.5 PgC yr-1) in particular, in the regime shift in carbon uptake by terrestrial plants in the late 1980s (Table 1 and Fig. S2). While the northern African region also exhibits a robust albeit smaller increase in data-driven NPP (∼ 0.2 PgC yr-1) in the late 1980s, no corresponding NPP shifts are discernable for tropical/southern and global land areas (Table 1). It is well known that factors like the El Niño–Southern Oscillation (ENSO; van der Werf et al., 2004) and volcanic aerosols (Lucht et al., 2002) have a large influence on variability in the carbon balance of terrestrial ecosystems (particularly at interannual timescales), and these phenomena may have also played a role in the prominent late-1980s NPP regime shifts. An analysis that explicitly accounts for such effects, however, suggests that these two factors are not the causes of the shift but indicates that a shift in data-driven NPP in the late 1980s emerges also for tropical/southern and global land regions, albeit with limited statistical significance (Table 1). In a complementary analysis, we also re-assessed the robustness of the earlier reported abrupt shift in the residual global land carbon sink in the late 1980s (Sarmiento et al., 2010; Beaulieu et al., 2012a) by analyzing a GCB for the period 1959–2011 (Le Quéré et al., 2013) with our change point methodology that also accounts for explicit uncertainties in the individual budget terms (Sect. 2.2). Results confirm the presence of a regime shift in the global residual land sink (∼ 1–1.3 PgC yr-1 depending on statistical treatment) in the late 1980s (Table 1 and Fig. S3). Taken together, the good agreement in the timing of regime shifts in the global residual land carbon sink and continental data-driven NPP may imply that the latter is a significant driver of the increased terrestrial carbon uptake in the late 1980s.

In order to unravel the mechanisms leading to the continental shifts in data-driven NPP, we focus on the two target regions of northern Eurasia and northern Africa that predominantly contributed to this late-1980s shift (see Fig. 1a). A factorial analysis for specific seasons shows that the northern Eurasian continent experienced a marked increase in spring temperatures and spring satellite vegetation activity (fAPAR) in the late 1980s that together produced a substantial increase in spring NPP (Fig. 2a). This relatively sudden springtime warming was also associated with a markedly earlier spring onset (∼ 5 days; see Fig. S4), and the enhanced productivity in the early part of the growing season appears to have also benefited plant productivity in subsequent summers (Figs. 2a and S4). A closer inspection shows that the increase in summer productivity stemmed solely from satellite fAPAR contributions (Fig. 2a), which is consistent with the notion of a more developed canopy during the peak of the growing season (as captured trough fAPAR) that carries over the early spring onset signal (Richardson et al., 2010). Increased plant productivity in the spring and summer seasons contributed predominantly to the pronounced and sustained increases in annual NPP in the late 1980s (Fig. 2a).

Over northern Africa including the dry Sahel, marked increases in data-driven NPP during wet and dry seasons that are driven both by increases in rainfall and by satellite fAPAR triggered a pronounced increase in annual NPP in the late 1980s (Fig. 2b). A closer inspection shows that in the period after this shift rainfall increased specifically during the later portion of the rainy season, which effectively lengthened the more productive growing season (Fig. S4).

The exploited biogeochemical model (CASA) for data-driven NPP simulations has a relatively simple structure and provides an integrated view (via satellite fAPAR) of the many interacting factors that influence NPP variability. Further, data-driven NPP estimates are also influenced by observational uncertainties in both satellite (e.g., volcanic aerosols, cloud cover, signal saturation) and key climate driver data that are only partially accounted for in our data-driven NPP simulations (see Sect. 2.1). We thus explored if process-based terrestrial biosphere models driven by climate and atmospheric CO2 observations also show evidence of a marked shift in NPP in the late 1980s. Results based on the TRENDY ensembles (Sect. 2) show that for the satellite period (∼ last 3 decades) a NPP shift in the late 1980s emerges as a prominent feature, but only in experiments that capture variability due to climate exclusively (Figs. 3, S5 and Table S1). Regional attributions associated with the shift are similar to those in the case for data-driven NPP, but differences in NPP sensitivities to climate (inferred from differences in the magnitude of the shifts) are evident (Tables 1 and S1). For example, the magnitude of the late-1980s NPP shift in northern Eurasia based on TRENDY is only about half the size of the corresponding shift in data-driven NPP (Tables 1 and S1). One reason for this marked difference may be that seasonal carry-over effects in NPP are largely absent in the TRENDY models; e.g., the late-1980s northern Eurasian annual NPP shift in TRENDY is solely due to spring contributions (coincident with the time of the climate forcing), whereas for data-driven NPP it is comprised of about equal contributions from spring and summer, with satellite fAPAR providing the link through which the early spring onset signal is propagated (Figs. 2a and 3a). It should be noted that such seasonal carry-over effects due to an earlier spring onset are observed at multiple eddy covariance flux sites across temperate and boreal ecosystems (Richardson et al., 2010) and that the phenology response (at seasonal and interannual timescales) in data-driven approaches is generally considered more robust (Raczka et al., 2013).

The TRENDY simulations are not restricted to the satellite period, allowing us to assess whether the identified late-1980s NPP shifts also emerge as the dominant pattern when the study period is extended to the last 5 decades (to be consistent with the time frame of the GCB). Results show that the late-1980s shift over northern Eurasia is a stable pattern. For northern Africa, however, an even more prominent shift is identified in the late 1960s (Fig. S5 and Table S1). This may suggest that this region by itself is not important enough to influence the global land sink (since there is no evidence for a corresponding shift in the global residual land sink; see Table 1). Further, in TRENDY experiments in which atmospheric CO2 and climate drivers are varied, the shift appears to be masked by an increasing trend in NPP associated largely with the CO2 fertilization effect (Figs. 3, S6 and Table S1). In fact, the high NPP sensitivity to changes in atmospheric CO2 concentrations in many of the current generation of terrestrial biosphere models (Arora et al., 2013) and the potential role of nutrient limitations (Zaehle, 2013) and/or climate feedbacks (Smith et al., 2016) in mitigating this sensitivity are presently a subject of intense research.

The extent to which the identified climate-driven regime shifts in NPP in the late 1980s translate into a sustained carbon sink (consistent with the shift seen in the residual land carbon sink from the GCB; Table 1) depends in part on associated responses in key carbon loss fluxes such as Rh, which (apart from its dependence on substrate supply from NPP) often depends on climatic factors in a similar fashion to NPP (Lucht et al., 2002). A limitation, however, is that currently no data-driven analog for Rh estimation exists, and one has to revert to alternative methods including more uncertain process-based simulations. Nevertheless to estimate the degree at which shifts in NPP may be potentially offset by corresponding shifts in Rh, we apply our change point framework also on the Rh as well as NEP fluxes from the CASA and TRENDY simulations (see Sect. 2). Results show that, for the large land regions of interest, regime shifts in NPP are repeatedly accompanied by substantial shifts in Rh (Tables 1, S1–S2; Figs. S2, S5–S6) often with 1–2-year lags (seen most clearly when ENSO and volcanic influences are accounted for). A consequence is that corresponding shifts in NEP are often less robust or not detectable (Tables 1 and S3). For example, in the case of the two focal regions northern Eurasia and northern Africa, the late-1980s shifts in data-driven NPP are offset by corresponding shifts in Rh at levels of 64–66 and 80–90 % (depending on statistical treatment), respectively (Table 1). As stated, the estimated shifts in the Rh fluxes are more uncertain and may represent more upper-bound estimates as a new study suggests that carbon models have a tendency to transfer carbon too quickly through the plant–soil systems because of severe biases in simulated soil carbon and/or too-high Rh sensitivities to climate (Carvalhais et al., 2014).

At global scale, a shift in Rh can also be estimated as the residual between a NPP shift and a corresponding shift in the residual land sink based on the GCB (Anderegg et al., 2015). Using methods that include such residual calculations as well as our direct model estimates (see Methods in the Supplement), we estimate a shift in global NPP in the late 1980s of 1.14 ± 0.34 PgC yr-1 and a corresponding shift in Rh of 0.36 ± 0.48 PgC yr-1. Our best estimate for an associated global shift in NEP is 0.63 ± 0.30 PgC yr-1, which amounts to roughly 60 % of the magnitude of the late-1980s shift in the residual land sink from the GCB (1.12 ± 0.14 PgC yr-1; based on the three estimates shown in Table 1).