Multi-observer methods for estimating uncertain species identification

Abstract

Species identification typically is assumed certain in population ecology. However, partial identification (to lowest certain taxon) and misidentification (collectively “uncertain identification”) have pervaded surveys. Consequent bias in occupancy and abundance estimates impairs inference and decision-making, but methods for estimating uncertain identification and remedying its effects have required restrictive assumptions or utilized artificial cues. I present likelihood-based multi-observer methods for estimating uncertain identification during surveys enumerating individuals, assess performance via simulation study, demonstrate application to line transects, and provide recommendations for use and further development. Methods typically employ three or more independent observers using one or more observation methods. Species identity is a latent state, and discrepancies in species classifications of the same individuals or groups enable estimation of uncertain identification and true species composition of detected individuals. Because complete or certain classifications are not required for any observer, methods uniquely befit surveys for free-ranging, unmarked animals. Methods permit arbitrary numbers of species and observation states, multi-species groups, covariates, and distinct uncertain identification for each species, observer, and observation state. Innovative technologies for acquiring recorded observations (e.g., digital media of survey observations) complement methods by enabling post hoc classifications by independent observers and enhanced species identification (e.g., replays, magnification, spectrograms). Using multi-observation methods (MOM), classifications by observers at sampling sites are supplemented with classifications from recorded observations by additional observers. Alternatively, recorded observations can be acquired solely via technology (e.g., unmanned aerial systems, autonomous recording units). Multi-observation methods enable supplementing automated classifications (e.g., machine learning) with classifications by humans, while single-observation methods allow classifications by multiple human observers. Simulations demonstrated good model performance and robustness, except to high un-modeled heterogeneity in misidentification. Minimizing misidentification and other practical steps enhance precision and accuracy of results. I applied MOM to line transects for two genus Brachyramphus murrelet species by photographing a sample of detected groups. Post hoc classifications from photographs by three independent observers had lower misidentification and partial identification (0.6%, 9%) vs. experienced observers at survey sites (3.2%, 23%). Integrating multi-observer methods with other methods (e.g., distance sampling, N-mixture models) enables estimation of occupancy or abundance accounting for uncertain detection and identification.