We discovered a decreasing trend of E from 1957–1990 which is in line with the previous finding that the proportion of carbon emissions remaining in the atmosphere (airborne fraction) was increasing during this period . After the Pinatubo eruption period (after 1994), E increased again and then decreased after 2009. To verify the robustness of this breakpoint, we did the following sensitivity tests: 1) using L from other three data sources (H&N, DGVMs and OSCAR), B from DGVMs and B from the residual of different ocean flux estimates; 2) calculating E using all the combinations (28 in total) of B from seven estimates and L from four estimates; 3) removing adjustment of fossil fuel emissions on Jena data; 4) applying different moving average methods; and 5) masking the latest El Niño event.
In general, our breakpoint is robust regardless of different choices of datasets and methods. Although there are large differences in L among different datasets (Figure S1c), the absolute values of L are small, and the breakpoint detection in E is robust regardless of different L choices (p < 0.01, Figure S11). This breakpoint is, however, not reflected using B from DGVMs, mainly due to the large interannual variability of B before 2009 (Figure S12). On the other hand, the decreasing trend of E after 2009 in DGVMs (Figure S12) is consistent with the trend detected by inversion data. A previous study also found the global difference in land sink between DGVMs and inversion datasets agreed well with the budget imbalance in the global carbon budget , indicating a possible bias in the land sink simulated by DGVMs. We also test the ocean sink from each individual ocean model estimate and pCO2-based product of the global ocean sink reported in the global carbon budget, instead of the ensemble mean, to calculate the residual B in the global carbon budget, and the breakpoint of global E still exists (p < 0.01, Figure S13). Using the mean E from 28 combinations of B and L estimates, the detection of breakpoint remains significant (p < 0.01, Figure S14). With more observation data constraint, the 1-σ uncertainty range of E is smaller after the Pinatubo eruption (Figure S14). Without the adjustment of fossil fuel emissions on Jena data (see Sect. 2.1), the global pattern remains consistent (p < 0.01, Figure S15). To evaluate the influence of moving average methods, we applied 3-yr moving average on the time series, and significant global breakpoint around 2009 in E is found in each dataset (p < 0.1, Table S2). Even using the original annual values without moving averages, the breakpoint around 2009 is still detectable although not significant (Table S2). After replacing the B in the strong El Niño years (2015 and 2016) with the averaged B of 2014 and 2017, the breakpoint detection is still significant after 5-yr moving average in CAMS and 3 Jena datasets (p < 0.1, Table S2), indicating that this reversal is not completely caused by the latest strong El Niño event during 2015–2016. These sensitivity tests further confirm the robustness of the breakpoint and support that the smaller increasing trend of B is the dominant factor determining the reversed trend in land sink efficiency after 2009.
Compared to the period before 2009, the acceleration of L growth in Africa and East Asia together with the weakening or relatively stable trend in B in the tropics, especially in Latin America, result in the weakening of E after 2009 (Fig. 3). This is consistent with the dominant role of tropical regions in the interannual variability of the global carbon cycle [43–46].
Climate variations play an important role in the reversed trend in land sink efficiency after 2009. In fact, 5-yr moving averages of MEI and PDO index both show a strong increasing trend since 2009 (Figure S16), which influences the land carbon sink. However, the impacts of ENSO and PDO on E seem to be small (Fig. 1) during the increasing phase of ENSO and PDO in the previous cycle (2000–2004), probably because of reduced intensity and shorter duration (Figure S16). In fact, if we apply multiple linear regression using annual precipitation, temperature and CO2 concentration to predict B in tropics and use the predicted B to calculate tropical E, the same breakpoint in the predicted E trend around 2009 is still significant (Figure S17a), indicating these factors can largely explain the trend reversal. Note that MEI is strongly correlated with temperature and precipitation in the tropics (Figure S17b). Previous studies also found that climate variations caused by e.g. El Niño may influence the tropical carbon cycle through different processes in different regions .
Climate impacts on land sink may need to be accurately quantified in the future. It has been proved that extreme El Niño events that strongly reduce tropical land carbon sinks are expected to be more frequent due to future greenhouse warming , but the trend of El Niño/La Niña intensity still remains unclear. While El Niño/La Niña cycles affect E, El Niño impacts are probably compensated by recovery fluxes in the subsequent years . If they can fully offset each other, ENSO impacts on E may be negligible in the long term. On the other hand, if there is an anthropogenic fingerprint in trends in El Niño/La Niña intensity , climate change impact on natural land sink need to be considered to define mitigation goals compatible with the Paris agreement. In addition, PDO and other low frequency variability patterns might affect climate at the time scales these mitigation policies should be acting (the next few decades). They could either result in additional CO2 in the atmosphere (e.g. by imposing more drought/higher temperatures) amplifying the impacts or reduce it temporarily (if they would lead to some cooling and wetter conditions). Their impacts on the natural sink thus need to be accurately quantified to avoid a false sense of implementation progress when assessing climate mitigation policies. Currently, there is not enough evidence to identify the most likely of these two possibilities. Nevertheless, it is necessary to track the efficiency of natural sink dynamics.
Saturation of land carbon sinks could also contribute to the reversal of trends in land sink efficiency after 2009. Processes that regulate land sink driven by atmospheric composition change (e.g. CO2 fertilization, nitrogen deposition), climate change (e.g. rising temperature) and LUC (e.g. forest regrowth, woody encroachment) are unlikely to be sustained permanently. Land carbon sinks are thus likely to decrease as the terrestrial carbon storage saturates. The saturation of the land sink is already indicated by a network of Amazonian forest plots  and modelling studies . However, some of this potential saturation may be offset by secondary forest regrowth in the tropics [16, 51]. Forest area gain in the tropics during 2009–2013 is lower than over the period before 2009 but became higher in the recent period after 2013 (Figure S18). Linking forest area gain to the net land sink, however, remains challenging due to the uncertainties in the forest gain detection from remote sensing, biomass growth and soil carbon dynamics. Therefore, the contribution of legacy land sink from forest gain needs to be further investigated with emerging evidence.