Reducing the uncertainty of wildlife population abundance: model-based versus design-based estimates
Résumé
Knowledge of population sizes in delimited spatial regions is crucial for most ecological research. Data from population surveys are collected with strip, line, or point transects sampling. These data are positively skewed and spatially autocorrelated, which makes estimation of uncertainty in the population size difficult. Thus, we propose a novel spatial-based estimator from a hierarchical spatial model for count data where the inhomogeneous animal density is decomposed into a deterministic trend related to potential habitat and a stationary latent field modeled by geostatistics. An empirical estimate of the latent variable is obtained including corrective terms for non-stationarity and variance resulting from a Poisson distribution of sightings. A novel block kriging estimator that takes into account both non-stationarity and the nature of count data is derived to obtain a spatial estimate of animal total abundance and variance of errors of prediction. From spatial simulated count data and real count data of common guillemot wintering in the Bay of Biscay (France), we compare mean population size and variance estimates obtained from our model-based approach to the design-based estimator (i.e., block bootstrap). The novel Poisson block kriging estimates greatly reduces uncertainty of population size estimates while block bootstrap provides larger uncertainties.