Discrimination capacity in species distribution models depends on the representativeness of the environmental domain

摘要: Aim When faced with dichotomous events, such as the presence or absence of a species, discrimination capacity (the ability to separate instances from absence) is usually only characteristic that assessed in evaluation performance predictive models. Although neglected, calibration reliability (how well estimated probability represents observed proportion presences) another aspect models provides important information. In this study, we explore how changes distribution make context-dependent For first time, explain implications ignoring context dependence can have interpretation species models. Innovation In paper corroborate that, under uniform presence, well-calibrated model will not attain high power and value area curve be 0.83. Under non-uniform distributions simulations show broad range values. These results illustrate property, i.e. it gives information about certain algorithm data population. Main conclusions In modelling, meaningful for given geographic temporal snapshot. This because representativeness environmental domain geographical context, which unavoidably entails presence. Comparative studies intend generalize their based on may broadly extrapolated. Assessment especially recommended when are intended transferred time space.