物种分布模型被广泛应用于评估气候变化对物种分布的影响。随着计算机和统计学的发展, 模拟物种分布的模型层出不穷, 但对这些模型的相对表现知之甚少, 因此需要对其进行对比分析, 以便更可靠地评估气候变化的影响。该文采用3个比较新颖的组合集成学习(ensemble learning)模型(随机森林(random forest, RF)、广义助推法和NeuralEnsembles)、3个常规模型(广义线性模型、广义加法模型和分类回归树)、3个大气环流模型(global circulation model, GCM) (MIROC32_medres, JP; CCCMA_CGCM3, CA; BCCR-BCM2.0, NW)和一个气体排放情景(SRES_A2), 模拟分析了马尾松(Pinus massoniana)历史基准气候(1961-1990)和未来3个不同时期(2010-2039, 2020s; 2040-2069, 2050s; 2070-2099, 2080s)的潜在分布。基于环境阈值方法选择物种不发生区, 依据ClimateChina软件进行当前和未来气候数据的降尺度处理, 采用接收机工作特征曲线(receiver operator characteristic, ROC)下的面积(area under the curve, AUC)、Kappa值和真实技巧统计法(true skill statistic, TSS)以及马尾松种子区划范围来评价模型的预测精度。结果表明: 6个物种分布模型都具有较高的预测精度, 但组合集成学习模型的预测精度稍高于其他常规模型, 其中RF的预测精度最高。3个GCM和6个模型模拟条件下, 马尾松对气候变化的响应格局既有一致性也有异同性。一致性表现在: 随着时间的推移, 马尾松分布区将逐渐向北迁移, 未来潜在分布区的面积将逐渐增加; 异同性表现在: 在不同模型和不同气候情景下, 马尾松潜在分布区的迁移距离和面积变化幅度不同, 其中NW模式下预测的变化幅度小于CA和JP模式; RF模型预测的分布区迁移距离和面积变化幅度最大。随着时间的推移, 未来马尾松的18个潜在分布空间预测图(6个模型 × 3 GCM)之间的差异也逐渐增大, 其中空间不一致性地区主要集中发生在马尾松潜在分布区的北部和西部边缘地带。模型本身不同的构建原理以及GCM之间的差异是导致预测结果存在差异的主要原因。

Aims New, powerful statistical techniques and GIS tools have resulted in a plethora of methods for modelling species distribution. However, little is known about the relative performance of different models in simulating and projecting species’ distributions under future climate. Our objective is to compare novel ensemble learning models with other conventional models by modelling the potential distribution of Masson pine (Pinus massoniana) and identifying and quantifying differences in model outputs.
Methods We simulated Masson pine potential distribution (baseline 1961-1990) and projected future potential distributions for three time periods (2010-2039, 2040-2069 and 2070-2099) using three global circulation models (GCM) (MIROC32_medres, JP; CCCMA_CGCM3, CA, and BCCR-BCM2.0, NW), one pessimistic SRES emissions scenario (A2), three ensemble learning models (random forest, RF; generalized boosted model, GBM, and NeuralEnsembles) and three conventional models (generalized linear model, GLM; generalized additive models, GAM, and classification and regression model, CART). The environmental envelope method was used to select absence of species. The area under the curve (AUC) values of receiver operator characteristic (ROC) curve, Kappa and true skill statistic (TSS) were used to objectively assess the predictive accuracy of each model. National standards for seed zone of Masson pine (GB 8822.6-1988) was employed to intuitively assess model performance. We developed ClimateChina software to downscale current and future GCM climate data and calculate seasonal and annual climate variables for specific locations based on latitude, longitude and elevation.
Important findings Ensemble learning models (GBM, NeuralEnsembles and RF) achieved a higher predictive success in simulating the distribution of Masson pine compared to other conventional models (CART, GAM and GLM). RF had the highest predictive accuracy, and CART had the lowest. Masson pine shows a globally consistent pattern in response to climate change for the three GCMs and six models, i.e., Masson pine will likely gradually shift northward and expand its distribution under altered future climate, with the magnitude of range changes dependent on model classes and GCMs. RF predicts a greater magnitude of range changes than other models. Projections of Masson pine distribution by NW are more conservative than JP and CA climate scenarios. In the case of Masson pine, range changes are mainly attributed to the colonization of newly available suitable habitat in high-latitudes and unchanging habitat suitability in the south-central part of its baseline range. Differences among 18 projections (6 models × 3 GCMs) increase with increasing time, and the greatest spatial uncertainty in projections is mainly in the north and west borders of the potential distribution range.