Monitoring animal populations with cameras using open, multistate, N-mixture models

Remote cameras have become a mainstream tool for studying wildlife populations. For species whose developmental stages or states are identifiable in photographs, there are opportunities for tracking population changes and estimating demographic rates. Recent developments in hierarchical models allow for the estimation of ecological states and rates over time for unmarked animals whose states are known. However, this powerful class of models has been underutilized because they are computationally intensive, and model outputs can be difficult to interpret. Here, we use simulation to show how camera data can be analyzed with multistate, Dail-Madsen (hereafter multistate DM) models to estimate abundance, survival, and recruitment. We evaluated 4 commonly encountered scenarios arising from camera trap data (low and high abundance and 25% and 50% missing data) each with 18 different sample size combinations (camera sites = 40, 250; surveys = 4, 8, 12; and years = 2, 5, 10) and evaluated the bias and precision of abundance, survival, and recruitment estimates. We also analyzed our empirical camera data on moose (Alces alces) with multistate DM models and compared inference with telemetry studies from the same time and region to assess the accuracy of camera studies in tracking moose populations. Most scenarios recovered the known parameters from our simulated data with higher accuracy and increased precision for scenarios with more sites, surveys, and/or years. Large amounts of missing data and fewer camera sites, especially at higher abundances, reduced the accuracy and precision of survival and recruitment. Our empirical analysis provided biologically realistic estimates of moose survival and recruitment and recovered the pattern of moose abundance across the region. Multistate DM models can be used for estimating demographic parameters from camera data when developmental states are clearly identifiable. We discuss several avenues for future research and caveats for using multistate DM models for large-scale population monitoring.

Methods
For the simulation analysis, data were generated using base simulation functions in R (see code) and there are no traditional field data associated with this part of the manuscript.

The dataset (moose_data.rds) accompanies the manuscript: "Monitoring animal populations with cameras using open, multistate, N-mixture models". It is an rds file that includes counts of adult female and juvenile moose (Alces alces) captured on remote cameras. The file is a 4-dimensional table that includes sites (n = 225 [indexed as 257]), years (n = 6), age classes (n = 2; adult and juveniles), and surveys per year (n = 4). We have also included another file (moose_data_bulls.rds) that includes counts of adult female and male moose as well as juveniles. These data were not formally analyzed but mentioned in the discussion as a dataset for readers to explore using multistate DM models. The data were collected by Dr. Alexej Siren and the other co-authors (see dataset authors) in Vermont and New Hampshire, USA, from 9 January 2014 to 9 August 2019. Data were examined for logical errors in processing step 1. Logical errors that could not be reconciled by observer agreement (majority), were assigned as 0s in the database, which is consistent with camera data standards. All missing data (not recorded in the field) was assigned an NA  value. Data were collected at random sites in Vermont and New Hampshire. Care should be taken when extrapolating values outside of the geographical domain. The counts of adult female and juvenile moose are affected by local weather conditions and seasonal phenology, therefore raw counts do not represent a census or absolute measure of absence or abundance.