Assessing the accuracy of species distribution models: prevalence, kappa and the true skill statistic (TSS)

By Omri Allouche, Asaf Tsoar, R. Kadmon. 2006

5,269 citations recorded by the source corpus.

Subfields: Ecological Modeling, Ecology, Nature and Landscape Conservation

View Paper: https://www.semanticscholar.org/paper/8519821d5c27722bf18728b3783340ad86b6cde3

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Ecological Modeling

In Ecology

In Nature and Landscape Conservation

Citations in the corpus

Species distribution models and mitochondrial DNA phylogeography suggest an extensive biogeographical shift in the high‐intertidal seaweed Pelvetia canaliculata (2014, methodology): “Models were evaluated in terms of discriminatory power (i.e. the capacity to differentiate presences from absences) against testing data by the area under the receiver operating characteristic curve (AUC; Allouche et al., 2006).”
Combining ensemble modeling and remote sensing for mapping individual tree species at high spatial resolution (2013, methodology): “A model was trained on 70% of the data (chosen randomly) and evaluated on the remaining 30% using the true skill statistic (TSS; Allouche et al., 2006).”
From introduction to the establishment of alien species: bioclimatic differences between presence and reproduction localities in the slider turtle (2009, background): “This further confirms that reproductive populations have a small but well-defined niche, because distribution models of generalist species often have lower AUC (Brotons et al ., 2004; Allouche et al ., 2006).”
Role of climate and invasive species in structuring trout distributions in the interior Columbia River Basin, USA (2011, methodology): “We determined the optimal cutoff separately for each model as the value between 0.01 and 0.99 (in 0.01 increments) that maximized the true skill statistic (TSS; Allouche et al. 2006), which is defined as sensitivity + specificity – 1.”
Drawbacks to palaeodistribution modelling: the case of South American seasonally dry forests (2013, methodology): “The models were originally evaluated based on true skill statistics (TSS; Allouche et al., 2006), and models with poor performance (i.e. TSS < 0.7) were eliminated from the ensembling process.”
Tropicalization of fish assemblages in temperate biogeographic transition zones (2014, methodology): “…of the short- and long-term predictive models was evaluated by measuring the agreement between the observed and modelled assemblages from the common years (1993 to 2002, 2010) through the mean area under the receiveroperated characteristic curve (AUC), as recommended by Allouche et al. (2006).”
Modelling invasion for a habitat generalist and a specialist plant species (2008, background): “Kappa statistic values range from –1 to +1, where +1 would be perfect agreement and any values less than 0 would indicate a performance no better than random (Cohen, 1960; Allouche et al ., 2006).”
Comparing species distribution models constructed with different subsets of environmental predictors (2015, background): “The TSS metric quantifies a model’s ability to correctly classify presences and absences (calculated as sensitivity + specificity — 1) for a defined threshold, and has been shown to be independent of species’ prevalence (Allouche et al., 2006).”
Exploring consensus in 21st century projections of climatically suitable areas for African vertebrates (2011, methodology): “Projections of the probability of occurrence of species in each site were converted to binary format (presence/absence) using a threshold maximizing the True Skill Statistic (TSS, Allouche et al. , 2006).”
Bioclimatic velocity: the pace of species exposure to climate change (2014, background): “We selected the threshold that maximized the true skill statistic (TSS) metric because this metric is unaffected by prevalence (Allouche et al., 2006).”
Measuring the accuracy of species distribution models: a review (2009, methodology): “Though statistical properties for many measures are known (see Couto 2003, Obuchowski 2005, Allouche et al. 2006, and the other references cited) and can be used to calculate the variance or standard deviation or confidence interval for the calculated accuracy, such reporting is unusual in modelin”
Deep Convolutional Neural Networks Enable Discrimination of Heterogeneous Digital Pathology Images (2017, methodology): “We also address other measures such as Cohen’s kappa (Cohen, 1960) which is a popular way of measuring the accuracy of presence and absence predictions because of its simplicity and its tolerance to zero values in the confusion matrix (Allouche et al., 2006).”