A joint distribution framework to improve presence-only species distribution models by exploiting opportunistic surveys

Aim: The availability of data related to species occurrences has favoured the development of species distribution models using only observations of presence. These data are intrinsically biased by the sampling effort. Presence-only (PO) species distribution models (SDM) typically account for this effect by introducing additional data considered to be related with the sampling. This approach, however, does not allow the characterisation of the sampling effort and hinders the interpretation of the model. Here, we propose a Bayesian framework for PO SDMs that can explicitly model the sampling effect. Location: Mexico. Taxon: Pines, flycatchers (family Tyranidae), birds and plants. Methods: The framework defines a bivariate process separable into ecological and sampling effort processes. PO data are conceived of incomplete observations where some presences have been filtered out. A choosing principle is used to separate out presences, missing data and absences relative to the species of interest and the sampling observations. The framework provides three modelling alternatives to account for a spatial autocorrelation structure: independent latent variables (model I); common latent spatial random effect (model II) and correlated latent spatial random effects (model III). The framework was compared against the Maximum Entropy (MaxEnt) algorithm in two case studies: one for the prediction of pines and another for the prediction of flycatchers. Results: In both case studies, at least one of the proposed models achieved higher predictive accuracy than MaxEnt. The model III fit best when the sampling effort was informative, while model II was more suitable in cases with a high proportion of non-sampled sites. Main Conclusions: Our approach provides a flexible framework for PO SDMs aided by a sampling effort process informed by the accumulated observations of independent and heterogeneous surveys. For the two case studies, the framework provided a model with a higher predictive accuracy than an optimised version of MaxEnt.