The uncertainty of climate change and uncertainty of species-environment relationship cause great variability in the studies of climate change biology. To reduce such uncertainties, scientists started to use ensemble models in this field. Our objective is to introduce the approach of the ensemble models, and predict the future range shift of one endangered species, the crested ibis (Nipponia nippon) as an example. The crested ibis had been critically endangered, and currently its population is rapidly recovering. The range of the crested ibis is still small after its recovery from the critical endangered status, so that climate change might be a threat to its long term survival. We used the locations of nest site to represent the distribution of the crested ibis, which have a high accuracy level and has being accumulated from 1981 to 2010. We applied nine modes in BIOMOD (a package of R software) to predict the current (1950-2000) and future (i.e. 2020, 2050, and 2080) distribution ranges of the crested ibis using five climate variables (i.e. annual minimum temperature, annual maximum temperature, seasonal variance of temperature, annual total precipitation, and seasonal variance of precipitation) based on CGCM2 climate model A2a emission scenario in WorldClim database. The nice models are Generalized Linear Models, Generalized Additive Models, Classification Tree Analysis, Artificial Neural Networks, Mixture Discriminant Analysis, Multivariate Adaptive Regression Splines, Generalized Boosting Models, Random Forest, and Surface Range Envelope.
We compared the current climate conditions with those in 2080, and found that the current habitat of the crested ibis would become warmer and wetter in the future. All nine models indicated that the crested ibis would have a northward range shift (actually a higher elevation shift), and the distribution center would be out of the current nature reserve. Therefore, it is necessary to develop a long term conservation plan for the crested ibis, e.g. adjusting the nature reserve border or design a new nature reserve. The nine models showed differences in predicted ranges, weights of explanatory variables, and goodness-of-fit (based on ROC curves and Cohen's Kappa indices). Among five climate variables, the seasonal variance of precipitation is the most important variable that associated with distribution of the crested ibis; and seasonal variance of temperature is the secondly important variable. The overall performance of all models are very high, indicated the distribution of the crested ibis had a strong pattern (The crested ibis is well constrained by environmental variables, not scattered randomly). The Random Forest has the highest model performance, and the Artificial Neural Networks ranks the second. The high performance of the two models is partly due to their high complexity.
We should be cautious whenever using species distribution models to predict the effect of climate change, because such models are based on the assumption that climate variables are the limiting variables restricting the range of the species, and the current population is in its favorite climate niche. As to the crested ibis, the assumption can hardly be satisfied, because other environmental variables such as human disturbance, wetland, and vegetation are also important to the crested ibis. As a result, the predicted range shift of the crested ibis is only a trend or potential distribution pattern in the future.
Because of the difference in model prediction and variability of model performance, we suggest to use ensemble models to deal with complex problem such as biological consequences of climate change to decrease the errors from models.