Evidence of Amazon rainforest dieback in CMIP6 models

Climate models that incorporate dynamic vegetation and are from the 6th Phase of the Coupled Model Intercomparison Project CMIP6 were utilised in this study. See Table 1 for the corresponding seven CMIP6 climate models. For the purpose of this study we wanted to focus on the climatic drivers alone impacting vegetation and therefore make use of the 1pctCO2 runs. Data from the PIControl runs were also used to determine each model’s internal variability. Primarily, we use model output data of the vegetation carbon and surface temperature for the seven climate models in the NSA region. The amplitude of the temperature seasonal cycle in this study is defined as the difference between the maximum and minimum monthly mean for each year. All anomalies presented correspond to the yearly mean relative to the mean of the first ten years, aside from temperature anomalies which correspond to the ten year running mean relative to the mean of the first ten years.

The algorithm used to detect abrupt shifts is relatively simple by design. Three criteria must be be fulfilled for a grid point to be identified as containing an abrupt shift in the vegetation carbon. Namely, the vegetation carbon must change by at least 2 kgC/m² over a 15-year period and that this must contribute to at least 25% of the overall change in vegetation carbon. Finally, to remove detected abrupt shifts that might be due to a model’s high internal variability the mean annual rate of change of the abrupt shift must be at least three times larger than the variability of the rates of change in the unforced control run.

Grid points where abrupt shifts were detected are subsequently sorted based on the direction of the abrupt shift (positive or negative) and the direction of overall trend (positive or negative). This results in four classifications of abrupt shifts, where analysis focused on the dieback abrupt shifts corresponding to the overall trend and direction of the abrupt shift being negative. This type of abrupt shift can be used as an analogy for a tipping event where a region changes equilibrium state from rainforest to savannah.

Our system specific EWS for Amazon dieback is to observe high sensitivities of the amplitude of the seasonal cycle to global warming. This is defined as the gradient of a linear regression fit to the amplitude of the temperature seasonal cycle against global warming. For a comparison between grid points with abrupt shifts and those that do not, the regression is fitted against the first 73 years of data corresponding to when CO₂ has doubled from pre-industrial levels (noting that most abrupt shifts occur after a doubling of CO₂). Area weighted histograms of the sensitivities are calculated to compare the distributions for areas with abrupt vegetation dieback (red) and areas with no abrupt shifts (purple). The mean of the seven model histograms was taken to produce the compiled models histogram in Fig. 4h. The percentage risk of an abrupt dieback shift occurring (Fig. 4i) is calculated as the percentage of abrupt dieback grid points to all grid points for the specified sensitivity.

We focus our analysis on detecting Amazon dieback abrupt shifts in seven state-of-the-art climate models, which all enable dynamic vegetation, from the 6th Phase of the Climate Model Intercomparison Project (CMIP6). Specifically, we are interested in climate change induced dieback (rather than direct deforestation) and therefore consider the idealised scenario of $CO_{2}$ increasing 1% per year starting from pre-industrial levels. We have defined that for an abrupt shift to be detected a $2kgC/m^{2}$ change must be observed within a 15-year period and this change must contribute to over a quarter of the change observed in the entire simulation run. We additionally require that the mean annual rate of change is more than 3 times as large as the variability in the rates of change seen in the unforced control runs (see Methods for further details).

The algorithm categorises abrupt shifts into four types dependent on the directions of the abrupt shift (AS) itself and the overall trend (T; change in 5-year means between start and end of the run). Figure 1a–g depicts the types of abrupt shifts detected spatially for the CMIP6 models. Conventional dieback abrupt shifts (red; $T<0$, $AS<0$), which indicate a move towards a savannah state, are predominantly detected. Three models; NorCPM1, TaiESM1 and SAM0-UNICON all show a clustering of dieback abrupt shifts in the north of the Amazon. GFDL-EMS4 also presents a coherent structure with abrupt shifts clustered in central-west Amazonia. Contrast to this, EC-Earth3-Veg shows many abrupt shifts scattered across the Amazon basin. This may be due to the high natural variability that is inherent in this model and which is difficult to distinguish from climate forced abrupt shifts. Interestingly, UKESM1-0-LL displays no dieback events, despite showing large scale dieback in previous CMIP generations (Cox et al., 2004). Similarly, very few abrupt shifts are detected in the MPI-ESM1-2-LR model.

Some sample time series of detected dieback abrupt shifts across the models are shown in Fig. 1h. Between the models there is some variation in the general shape of abrupt shift time series, however, most exhibit a change of state from one equilibrium to another that would be expected of a tipping event.

Henceforth, we restrict our analysis to the IPCC AR6 defined North South America (NSA) region, which contains the majority of the Amazon basin (see Fig. 2a), and features many of the detected abrupt dieback shifts (red points in Fig. 1). Figure 2b shows how the fractional area of the NSA region to experience an abrupt dieback shift evolves for increasing global warming. Some models show clear jumps in the NSA area to experience an abrupt shift reflecting multiple gird points featuring an abrupt shift at a similar level of warming. Strikingly, TaiESM1 shows about 20% of the NSA region suffering an abrupt dieback event at about $1.7^{o}C$ warming and by $3^{o}C$ global warming about 40% of the NSA region would experience an abrupt shift. The bold black line represents the mean behaviour and the plume variability of all seven models. Although some of the large jumps from individual models can still be identified, the model mean shows a smoother increase in fractional NSA area to undergo an abrupt shift under global warming. There is not a singular temperature threshold, instead the risk of tipping increases (approximately linearly) between $1.3^{o}C$ and $3^{o}C$ of warming and reaches approximately 20% of the NSA region to undergo an abrupt dieback shift.

Interestingly, when evaluating at the regional scale many of the abrupt shifts do not appear to materialise, despite identifying a significant number of local abrupt shifts, (see Fig. 2c). Only the cluster of abrupt shifts at approximately $1.3^{o}C$ warming for GFDL-ESM4 appear in the total vegetation carbon anomaly for the NSA region. Furthermore, the CMIP6 models do not even agree on the sign change in vegetation carbon.

Three identified abrupt shifts, which all occur around a doubling of $CO_{2}$ in different models, are shown in Fig. 3a–c and all show a change of equilibrium state after the abrupt shift. Initially, vegetation carbon may increase due to the $CO_{2}$ fertilization effect (c.f. Fig. 3a,b), however there exists a critical threshold in the $CO_{2}$ concentration at which increased temperature and drying overwhelm the positive effect of $CO_{2}$ and results in an abrupt dieback shift. Fig. 3d–f show the trend in the temperature seasonal cycle associated with these three grid points. An increasing trend is observed in the amplitude of the seasonal cycle in the lead up to an abrupt shift for each point, suggesting that this increase in variability may be an EWS for abrupt dieback events. This behaviour can be expected in the lead up to a vegetation dieback shift because the length and intensity of the dry season has been shown to increase due to drying in the Amazon (Vogel et al., 2020; Malhi et al., 2009).

Figure 4 investigates the robustness of using the temperature seasonal cycle amplitude as an EWS for an impending dieback event. Specifically, we compare the distributions of grid points possessing an abrupt shift and those without for the temperature seasonal cycle amplitude sensitivity to global warming. The sensitivity is calculated up to a doubling of $CO_{2}$ (most abrupt shifts are detected after doubling of $CO_{2}$) for all grid points and models (Fig. 4a-g). The grid point are colour coded according to whether an abrupt shift occurs and its type. Four of the five models that contain abrupt shifts within the NSA region (GFDL-ESM4, NorCPM1, SAM0-UNICON, TaiESM1; Fig. 4b,d,e,f) display clear thresholds in the sensitivity, such that above the thresholds only grid points with an abrupt shift feature. EC-Earth3-Veg (Fig. 4a) provides an exception, however due to the high stochasticity of the model some of the detected abrupt shifts are likely to be simply natural variation. Promisingly, MPI-ESM1-2-LR and UKESM1-0-LL (Fig. 4c,g) do not possess grid points with high sensitivities and could therefore offer an explanation for not exhibiting any abrupt shifts in the NSA region.

Taking an ensemble mean of all the models, shows that grid points without an abrupt shift tend to have sensitivities centred around zero, whereas abrupt shift grid points are positively skewed to higher sensitivities (see Fig. 4h). This means that the risk of a grid point having an abrupt shift (defined as the ratio of gird points with an abrupt dieback shift to all grid points for each sensitivity) generally increases for grid points with higher sensitivities to global warming as shown in Fig. 4i. The minimum risk of a grid point experiencing an abrupt shift is for a sensitivity close to 0, where the seasonal cycle amplitude is unaffected by global warming. As the sensitivity increases from 0.5 to 1.0$K/K$ the risk of a grid point containing an abrupt shift increases approximately linearly from 10% to 60%. For sensitivities greater than 1.0$K/K$ the risk remains between 60% and 80% providing a potential EWS.The risk also increases to 35% for negative sensitivities, however this is largely from the EC-Earth3-Veg model in which it is not clear how many of the detected shifts are indeed abrupt.