ROC曲线形状在生态位模型评价中的重要性——以美国白蛾为例

The importance of the shape of receiver operating characteristic(ROC) curve in ecological niche model evaluation——case study of Hlyphantria cunea

【目的】生态位模型在生物地理学、入侵生物学和保护生物学中具有广泛的应用,被越来越多地用于预测物种潜在分布和现实分布的研究中。本文以美国白蛾为例介绍pROC方案在生态位模型评价中的应用及其注意事项,以期对物种潜在分布预测进行合理的评价,促进生态位模型在我国的合理运用和发展。【方法】介绍ROC曲线和AUC值基本原理,总结其在生态位模型评价中的应用,从物种存在分布点和不存在分布点的可信度出发,分析AUC值用于模型评价的优点和不足,最后介绍局部受试者工作特征曲线的线下面积方案(pROC方案)来弥补传统AUC值的不足。【结果】AUC值虽独立于阈值,但因其综合灵敏度和特异度,而屏蔽这2个指标各自的特征,不能分别评估预测结果的灵敏度和特异度,同时对遗漏率和记账错率不能进行权衡,会误导使用者对模型的评价。与AUC值相比,ROC曲线的形状更具有价值,蕴含丰富的模型评价信息。【结论】模型评价需要将灵敏度和特异度区别对待,ROC曲线形状比AUC值在生态位模型评价中更为重要,pROC方案相对于传统AUC值具有优势,但容易对过度模拟做出不当判断。模型评价与作者研究目的密切相关:当以预测物种潜在分布为目的时(如入侵物种潜在分布、气候变化对物种分布的影响和谱系生物地理学),模型评价应当给予灵敏度(或者遗漏率)更多的权重;当以预测物种现实分布为目的时(如保护区界定和濒危物种引入),模型评价应当给予灵敏度和特异度同等的权重。

[Aim] Ecological niche modeling （ENM） is increasingly used to estimate the potential and realized distributions of species in studies of biological invasion and conservation. We present the pROC approach for the evaluation of ENM of Hyphantria cunea, as a case study.[Method] We first introduced the ROC curve and AUC value in niche model evaluation. We then presented the shortcomings of AUC value based on different reliability of presence and absence records. Finally, we introduced the partial area under the receiver operating characteristic curve （pROC） approach to backup traditional AUC value in niche model evaluation.[Result] Model evaluation using AUC misleading although it independent of threshold. The AUC combined sensitivity and specificity but blanked the information of individual sensitivity and specificity, and weighted omission and commission error equally. We argued that the shape of ROC curve led to valuable information and was more important than AUC value in ENM evaluation.[Conclusion] Niche model evaluation should treat sensitivity and specificity separately. The shape of ROC curve was more important than AUC value. The pROC approach was found more powerful than traditional AUC value in model evaluation, but cautions are warrant when it was used to evaluate the model output of over prediction. Niche model evaluation should take the purpose of study into account, when the aim of study was to estimate potential distribution （e.g. biological invasion, climate change, phylogeography）, model evaluation should give higher weight on the sensitivity or omission error, whereas if the aim were to estimate realized distribution （e.g. conservation and reintroduction program）, model evaluation should weight sensitivity and specificity equally.