Spatial Gaussian processes improve multi-species occupancy models when range boundaries are uncertain and nonoverlapping

Abstract

Species distribution models enable practitioners to analyze large datasets of encounter records and make predictions about species occurrence at unsurveyed locations. In omnibus surveys that record data on multiple species simultaneously, species ranges are often nonoverlapping and misaligned with the administrative unit defining the spatial domain of interest (e.g., a state or province). Consequently, some species display differentially restricted extents within a study area. Assuming hard boundaries based on expert opinion or published range maps to restrict species occurrence predictions implies a false sense of certainty in model-based inferences. We propose a multi-species occupancy model with a spatial Gaussian process on site-specific effects for each species as a model-based solution. Specifying informative Bayesian hyperpriors on the spatial hyperparameters encapsulates broad-scale correlation among site occupancy probabilities for each species. We fit this model to acoustic detection/nondetection data collected with autonomous recording units during summer of 2016–2019 throughout Oregon and Washington, USA, on 15 bat species. We found vast improvements in spatial predictions of spotted bat ( Euderma maculatum ), canyon bat ( Parastrellus hesperus ), and Brazilian free-tailed bat ( Tadarida brasiliensis ) when the available environmental predictors were insufficient for characterizing their restricted ranges within the region. In contrast, widespread species ( Lasionycteris noctivagans , Myotis californicus , Myotis evotis , Myotis volans ) were appropriately modeled using only environmental predictors, such as percentage forest cover and cliff and canyon cover. Utilizing spatial Gaussian processes within a community or multi-species model incorporates uncertainty in range boundaries and allows for simultaneous predictions for the entire faunal assemblage even if species have nonoverlapping or restricted ranges within a spatial domain of interest. Such modeling improvements are essential if species distribution models are to accurately inform monitoring, species recovery plans, and other conservation efforts.