Spatial capture-recapture density estimation with latent, partial, and erroneous individual identity

Ben Augustine (USGS)

Friday 29th April, 2022 15:00-16:00 Zoom

Abstract

Capture-recapture study designs and statistical models are commonly used to estimate wildlife population parameters such as density, survival, and population growth rate, which inform conservation and management decisions. One of the key assumptions of capture-recapture models is that all individuals are identified without error upon capture or detection, and error rates as low as 1-5% can introduce substantial bias into population parameter estimates. The assumption of perfect individual identity has traditionally been difficult to meet in some sampling scenarios due to tag loss or misidentification, and uncertainty in individual identity is a common feature of newer, noninvasive methods of collecting capture-recapture data, such as genetic fingerprinting from hair or scat samples, camera traps, and bioacoustic recorders. Further, deep learning is increasingly being used to assign individual identities to camera trap photos or bioacoustic spectrograms and these methods are subject to classification error.

Several approaches have been used to account for partial or erroneous individual identity in classical capture-recapture models; however, spatial capture-recapture (SCR) models are particularly promising for resolving the uncertainty in individual identity since they use the spatial information of each detection event. SCR specifies a model for the number and spatial distribution of individual “activity centers” and conditioned on the activity centers is a model for the individual by trap detection probability, which is a decreasing function of the distance between an activity center and trap (i.e. a detection function). This spatial structure in population distribution and detection implies that the spatial location of a detection event is informative of its individual identity. In fact, density can be estimated using the spatial locations of samples without individual identity (unmarked SCR), though these density estimates are usually very imprecise and subject to bias.

In this talk, I will survey some recent examples of estimating population parameters from capture-recapture data with fully latent, partial, and erroneous individual identity within an SCR framework. Of particular focus will be a model for genetic fingerprinting that propagates the uncertainty in inferring individual identity from genetic samples observed with error to the population parameters of interest and allows all samples to be used regardless of reliability, thereby increasing precision and removing bias. This model replaces the subjective and deterministic identity assignments made by a geneticist with a spatially-explicit unsupervised classification model that leverages information from the ecological, capture, and genotype observation processes to probabilistically assign individual identities. In application to a fisher hair snare study, this model identifies probable errors made by a geneticist and increases the precision of the density estimate by 25%.

Add to your calendar

Download event information as iCalendar file (only this event)