Accurate and up-to-date population forecasts are needed in a variety of fields including endangered species management, disease epidemiology, and invasive species control. However, we still lack many tools necessary for operationalizing ecological forecasts. To help address that gap, we are developing a generalized population forecasting framework based on particle filtration. The forecasting framework allows for a dynamic set of hierarchical models that can include both phenomenological and mechanistic approaches to time series as well as ensemble-based forecasts. Iterative filtering provides a means to maximize model likelihoods, ratios of particle filter-based likelihood enable model comparison based on full time-series evaluation, and Bayesian model averaging facilitates ensemble building.
Here, we leverage a 40-year study of desert rodent communities outside of Portal, AZ to develop and test the frameworkâs capacity with a realistic observational multi-species dataset. Approximately monthly sampling (censuses centered on new moons) has generated 438 sets of observations (of 500 possible, 87.6% coverage). Specifically, we focus on four study plots that have been functionally unmanipulated for the duration of the study. We first build observation and sampling models to match the data and then explore both phenomenological and mechanistic time series process models.
Results/Conclusions
To date, 13,895 observations of rodents representing 21 species have been made in the four plots, with census-specific counts between 0 and 30 (mean: 7.96, variance: 24.61). Counts were equidispersed among plots and plot-level residuals were only marginally biased, indicating that the rodent populations are well-mixed at the site-level and that the sampling process can be modeled using a common Poisson distribution with a plot-level offset. The data also support a binomial observation process representing the detection of rodents present within each plot.
Total rodent counts show complex dynamics indicative of a combination of phase-remembering, phase-forgetting, and acyclic fluctuations, and these patterns were consistent across the four plots (lag-0 cross correlations ranged between 0.696 and 0.772). Phenomenological process models that include yearly and decadal cycles, long-term non-linear trends, and first-order autocorrelation were able to capture and reproduce the observed dynamics, setting the stage for predictive testing of forecasts.
By leveraging generalized mathematics and plug-and-play coding, our framework provides a flexible and powerful means to forecast dynamic populations, including hierarchical sampling and observation modeling as well as a smooth transition from phenomenological to mechanistic process modeling as data become available.