Addressing widespread detection heterogeneity in avian occupancy modeling using passive acoustic surveys

Abstract Ornithologists have embraced occupancy models to account for imperfect species detection. These models misestimate occupancy when they fail to model detection heterogeneity, variability in the probability of species detection. Practitioners typically account for detection heterogeneity using site and survey covariates, but generally ignore heterogeneity arising from ubiquitous processes like within-territory movement and territory position relative to the observer. Furthermore, while such heterogeneity may be widespread, traditional in-person monitoring projects usually lack the temporal resolution to quantify it. Passive acoustic monitoring (PAM) provides an alternative, enabling collection of more visits per site for less field effort. We investigated how spatial processes create detection heterogeneity, how the number of visits per site impacts diagnosis of heterogeneity, and how heterogeneity biases occupancy estimates for 3 occupancy modeling approaches. We simulated territory position, within-territory movement, and auditory detections of individual birds to generate single-season detection histories with up to 20 visits per site for varying avian occupancy probabilities, densities, and territory sizes. We also collected 3 large-scale single-season PAM datasets containing 10–21 visits per site. We modeled occupancy using the “basic” single-season, zero-inflated binomial occupancy model, the Royle-Nichols (RN) abundance model, and the zero-inflated beta-binomial (ZIBB) model, which draws site detection probability from a beta distribution. Detection heterogeneity was common in simulations across realistic ranges of density and territory size. Diagnosing heterogeneity required 3–20 visits per site, depending on occupancy probability and severity of heterogeneity. Heterogeneity caused occupancy misestimations from 94% underestimates to 400% overestimates. The RN model’s estimates were the most biased. The basic and ZIBB models produced negligibly biased estimates with increasing numbers of visits, the ZIBB model requiring fewer visits. Our results suggest widespread spatial processes may bias occupancy estimation from point-count and PAM surveys, but many-visit detection histories generated by PAM improve diagnosis of detection heterogeneity and accuracy of occupancy estimates.