Species Distribution Modelling: Contrasting presence-only models with plot abundance data

Abstract

Species distribution models (SDMs) are widely used in ecology and conservation. Presence-only SDMs such as MaxEnt frequently use natural history collections (NHCs) as occurrence data, given their huge numbers and accessibility. NHCs are often spatially biased which may generate inaccuracies in SDMs. Here, we test how the distribution of NHCs and MaxEnt predictions relates to a spatial abundance model, based on a large plot dataset for Amazonian tree species, using inverse distance weighting (IDW). We also propose a new pipeline to deal with inconsistencies in NHCs and to limit the area of occupancy of the species. We found a significant but weak positive relationship between the distribution of NHCs and IDW for 66% of the species. The relationship between SDMs and IDW was also significant but weakly positive for 95% of the species, and sensitivity for both analyses was high. Furthermore, the pipeline removed half of the NHCs records. Presence-only SDM applications should consider this limitation, especially for large biodiversity assessments projects, when they are automatically generated without subsequent checking. Our pipeline provides a conservative estimate of a species’ area of occupancy, within an area slightly larger than its extent of occurrence, compatible to e.g. IUCN red list assessments.

Evaluating presence‐only species distribution models with discrimination accuracy is uninformative for many applications

“…These include choice of modelling algorithm, required sample size, optimal model complexity, choice of study area from which data are drawn, the exclusion of outliers and selection of environmental predictors, among others (Acevedo, Jimenez-Valverde, Lobo, & Real, 2012;Boria, Olson, Goodman, & Anderson, 2014;Domisch, Kuemmerlen, Jahnig, & Haase, 2013;Garcia-Callejas & Araujo, 2016;Guisan, Graham, Elith, Huettmann, & Distri, 2007;van Proosdij, Sosef, Wieringa, & Raes, 2016;Soley-Guardia et al, 2016;Varela, Anderson, Garcia-Valdes, & Fernandez-Gonzalez, 2014;Wisz et al, 2008). Many of the measurable phenomena that are potentially related to suitability (e.g., population density [Carrascal, Aragon, Palomino, & Lobo, 2015], upper limit of local abundance [VanDerWal, Shoo, Johnson, & Williams, 2009, Gomes et al, 2018 have not been quantified in detail for many real species and as such are unavailable for model validation. Decisions about how best to model species are typically made using metrics that test discrimination accuracy on subsets of species occurrence data that have been withheld during model construction (Elith et al, 2006;Radosavljevic & Anderson, 2014); for a recent literature review and summary see Appendices S1 and S2.…”

“…However, it is often not clear which (if any) measurable biological phenomena should be correlated with suitability estimates from SDMs. Many of the measurable phenomena that are potentially related to suitability (e.g., population density [Carrascal, Aragon, Palomino, & Lobo, 2015], upper limit of local abundance [VanDerWal, Shoo, Johnson, & Williams, 2009, Gomes et al, 2018 have not been quantified in detail for many real species and as such are unavailable for model validation.…”

Disentangling the processes driving tree community assembly in a tropical biodiversity hotspot (New Caledonia)

“…This assumption can be flawed for two reasons. First, habitat suitability predicted by pseudoabsence SDMs differ from true occupancy probability that only presence–absence SDMs can provide (Gomez et al., ). Here, thresholds used to convert continuous habitat suitability maps into binary presence/absence maps based on sensitivity and specificity were found to vary with species prevalence ( r = 0.10; p < 0.01).…”

Identifying climate refugia and its potential impact on Tibetan brown bear (Ursus arctos pruinosus) in Sanjiangyuan National Park, China

“…MaxEnt model is considered one of the most efficient tools to predict species distribution with presence-only data, leading to its widespread use (Aryal eT al., 2016;Gomes et al, 2018;Lamsal, Kumar, Aryal, & Atreya, 2018;Ma & Sun, 2018;Phillips, Anderson, & Schapire, 2006). The parameters of MaxEnt model were set to: 25% for random test percentage and 1 regularization multiplier.…”