全球变化知识资源中心(globalchange): Life history and spatial traits predict extinction risk due to climate change

There is an urgent need to develop effective vulnerability assessments for evaluating the conservation status of species in a changing climate¹. Several new assessment approaches have been proposed for evaluating the vulnerability of species to climate change^{2, 3, 4, 5} based on the expectation that established assessments such as the IUCN Red List⁶ need revising or superseding in light of the threat that climate change brings. However, although previous studies have identified ecological and life history attributes that characterize declining species or those listed as threatened^{7, 8, 9}, no study so far has undertaken a quantitative analysis of the attributes that cause species to be at high risk of extinction specifically due to climate change. We developed a simulation approach based on generic life history types to show here that extinction risk due to climate change can be predicted using a mixture of spatial and demographic variables that can be measured in the present day without the need for complex forecasting models. Most of the variables we found to be important for predicting extinction risk, including occupied area and population size, are already used in species conservation assessments, indicating that present systems may be better able to identify species vulnerable to climate change than previously thought. Therefore, although climate change brings many new conservation challenges, we find that it may not be fundamentally different from other threats in terms of assessing extinction risks.

Attempts to quantify the threat that climate change poses to species’ survival commonly infer extinction risk from changes in the area of climatically suitable habitat (the bioclimate envelope)^{10, 11}, but this approach ignores important aspects of species’ biology such as population dynamics, vital rates and dispersal^{12, 13, 14, 15, 16}, leading to high uncertainty^{1, 17}. To address this challenge, we coupled ecological niche models (ENMs) with demographic models^{13, 14, 15, 18, 19, 20} and expanded this approach by developing a generic life history (GLH) method. The coupled modelling approach estimates extinction risk as the probability of abundance falling to zero by the year 2100, rather than as the proportion of species committed to extinction due to contraction of bioclimate envelopes¹⁰ (Methods).

By matching ENMs for 36 amphibian and reptile species endemic to the US with corresponding GLH models (Supplementary Table 1), we estimate mean extinction risk by 2100 to be 28 ± 7% under a high CO₂ concentration Reference climate scenario²¹ and 23 ± 7% under a Policy climate scenario that assumes substantive intervention²² (Methods). In contrast, extinction risk is estimated by the same models to be <1% without climate change, showing that the methods are not biased towards predicting high risks. The contrast between predicted extinction risk with and without climate change suggests that climate change will cause a pronounced increase in extinction risk for these taxonomic groups over the coming century. Contrary to other studies²³, the relatively small difference in extinction risk that we predict between Policy and Reference scenarios indicates that conservation actions that incorporate climate adaptation²⁴ will be necessary to substantially reduce extinctions due to climate change regardless of whether mitigation measures that decrease CO₂ emissions are implemented.

We next used sensitivity analysis to test how reliably extinction risk due to climate change can be predicted from present information on life history and spatial traits. Our goal was to assess whether commonly available variables can be used to effectively estimate extinction risk due to climate change without necessitating the application of complex forecasting models that are impractical to run for most species. On the basis of the simulated period 2000–2010, we extracted 21 variables that could, in practice, be measured for conservation assessment purposes (Table 1). Application of machine-learning methods—Random Forests (RF) and Boosted Regression Trees (BRT)—revealed good ability to predict extinction risk due to climate change from these variables: AUC =0.80–0.86 based on cross-validation partitioned by species so as to test using independent data (Methods).

Table 1: Predictor variables used to test whether extinction risk due to climate change can be predicted from present information on life history and spatial traits.

Full table

We used an ensemble of five atmosphere–ocean general circulation models to generate an annual time series of climate anomalies to 2100 based on two strongly contrasting greenhouse gas emission scenarios: a Reference scenario with CO₂ concentration of 750 ppm (WRE750; ref.21) and a Policy scenario with CO₂ stabilization at 450 ppm (MiniCAM LEV1; ref.22). Climate anomalies were downscaled to an ecologically relevant spatial resolution (~800 m×800 m) and 19 bioclimate variables were generated, from which 7 variables were selected on the basis of reasoning as to the physiological and life history requirements of the study species and analysis of correlations between variables.

We then combined the seven climate variables with other environmental variables (including land cover, hydrography and land surface form) and species’ occurrence records to generate annual maps of suitable habitat using ENMs. Occurrence data were obtained from NatureServe and we used the maximum entropy ENM method (Maxent). Maxent regularization was set for each species individually so as to avoid over-fitting, and the most relevant land cover, hydrography and land surface form variables were selected for each species to avoid fitting models with an unnecessarily large number of variables (Supplementary Table 2). The extent of the study region for ENM calibration was selected for each species on the basis of occupied ecoregions, and for each species the final ENM was a consensus from 50 replicates so as to account for uncertainty in the species’ occurrence data.

The dynamic spatial structures generated by ENMs were then combined with GLH models, which incorporate demographic structure, density dependence and stochasticity, to create metapopulation models (spatially structured models of multiple populations, with partially correlated dynamics, that may exchange individuals through dispersal). We constructed age- and stage-structured, density-dependent, stochastic models for 6 life history types: small salamanders, large salamanders, turtles, tortoises, snakes and lizards. From these GLH models we created 9,720 population models (3,240 for each climate scenario) by sampling a standard set of life history parameters between upper and lower bounds. Using a generic approach for modelling demographic processes prevents us from making species-specific projections or ranking these 36 species in terms of their vulnerability to climate change. However, it also avoids the need to obtain species-specific parameters, which are rarely known, and extends our scope of inference beyond a limited set of extant species to all possible trait combinations that may increase risk of extinction due to climate change. To estimate the risk of extinction we ran each of the 9,720 population models for 1,000 replicates, and each replicate with annual time steps to 2100. We also estimated extinction risk without climate change and we did not model non-climatic threats, such as habitat destruction or exploitation, enabling us to isolate the degree to which climate change increases extinction risk.

We then used machine-learning methods, RF and BRT, to identify important variables and interactions for predicting extinction risk due to climate change. We assessed RF and BRT model predictive ability using leave-one-out cross-validation in which each of the 36 species was treated in turn as an independent validation data set. By using species as a data partition instead of random sampling methods (for example, standard tenfold cross-validation), we were able to challenge the modelling algorithm against truly independent data. Thus, the predictive performance metrics we calculated (AUC =0.86 for RF under Reference scenario; for other results, see Supplementary Methods) could be expected to hold for predictions of climate-related risk to additional North American species not included in this study. For RF, importance of each predictor variable was determined by computing the prediction error of each tree for the out-of-bag sample (the set of observations set aside for validation and not used in constructing the trees) and assessing the degree to which out-of-bag prediction error increases when the values of that predictor variable are randomly shuffled. Univariate relationships between variables and extinction risk due to climate change (Fig. 1b) were derived post-hoc by predicting across the parameter space for each variable while holding all other variables constant at mean values. Two-way interaction strengths were computed post hoc following three steps: for each variable pair, predictions were made across the full two-dimensional slice of parameter space (holding all other predictor variables constant at mean values); predictions from step 1 were modelled assuming additivity (but the relationship is not constrained to be linear and could take any shape); and the root mean squared residual error under the additive model from step 2 (multiplied by 100 to convert to per cent risk) was calculated as an index of the strength of interaction³¹. For BRT, importance of each predictor variable was computed as the total reduction in deviance associated with that variable for the full model.

Supplementary Fig. 4 presents a flowchart detailing processing steps and data inputs/outputs. Further explanation and justification of data and methods is provided in Supplementary Methods.

Climate data are available through the NASA Center for Climate Simulation (www.nccs.nasa.gov; http://dx.doi.org/10.7917/D7WD3XH5).


Email:
Passwd
验证:
	换一张
Have you forgotten your password? Stay signed in Log In

全球变化知识资源中心

用户服务

站点统计

相关链接

联系我们