Interaction between classification detail and prediction of community types: implications for predictive modelling of benthic biotopes

Marine Ecology Progress Series

Per Bergstr#xf6m¹  Genoveva Gonzalez-Mirelis¹  Mats Lindegarth¹ 

关键词: Classification;    Benthic communities;    Predictive mapping;   

DOI  :  10.3354/meps09129

学科分类：海洋学与技术

来源: Inter-Research

【 摘 要 】

ABSTRACT: Ecological classifications lie, explicitly or implicitly, at the foundation of biotic maps, which are in turn an essential tool in ecology and conservation. Techniques are being developed that make efficient use of scarce biological data to create those maps, such as those based on modelling, or predictive mapping. However, little attention has been paid to just how sensitive models are to the pattern they are designed to model. We have examined the predictive accuracy of classification models (Random Forests) relative to the classification detail of community types. The data are from a recently established Marine National Park off the west coast of Sweden where a map of benthic communities has been commissioned. A total of 447 georeferenced, underwater video sequences constituted the sample base for this study. Samples were classified according to increasing values of a similarity threshold, based on faunal composition, in a hierarchical cluster analysis framework. A random forest was fitted at each level of classification detail to predict the class membership of an independent set of samples, based on environmental (terrain and substrate-related) variables. Predictive accuracy was high (multiclass area under the curve, M-AUC = 0.79 to 0.81) across intermediate levels of classification detail (similarity cut off, 40 to 50%). Predictive accuracy was moderate (M-AUC = 0.73 to 0.78) at both ends of the spectrum of classificatory resolutions tested (cut offs: 30%, i.e. very coarse; 60%, very fine). These findings have ramifications in the ‘classify first, then predict’ approach to predictive mapping at the community level, as they show that calibrating classes before modelling them greatly enhances model reliability.

【 授权许可】

Unknown