Predicting spatiotemporal abundance of breeding waterfowl across Canada: A Bayesian hierarchical modelling approach

• DOI: 10.1111/ddi.13129
• Publication Date: 2020-08-13T13:42:21Z
• Authors: Antoine Adde; Marcel Darveau; Nicole Barker; Steven Cumming
• Keywords: Generalized additive model

Abstract Aim Our aim was to develop predictive statistical models for mapping the abundance of 18 waterfowl species at a pan‐Canadian level. We refined previous generation national by (a) developing new, more interpretable that (b) explicitly account spatiotemporal variations in abundance, while (c) testing associations with an updated suite habitat covariates. Location All Canada, excluding Northern Arctic ecozone. Methods response variables were annual counts on 2,227 aerial‐survey segments over period 25 years (1990–2015). Combining machine‐learning and hierarchical regression modelling, we devised innovative covariate selection strategy select each best subset panel 232 candidate With selected covariates, implemented generalized linear Bayesian framework, using integrated nested Laplace approximation stochastic partial differential equation approaches. Results On average, our explained 47% observed variance predictions 74% temporally averaged spatial predictions. The included 94 significant waterfowl‐habitat involving 42 distinct average 5.3 covariates per model. Covariates forest attributes most represented models. proportional biomass Populus tremuloides frequently (10/94 10/18 species). Model generated maps abundances almost all Canada. Main conclusions showed it is possible efficiently combine machine‐learning, variable methods exploit high‐dimensional spaces. approach yielded powerful easily distribution very few accounting residual autocorrelation. Possible applications resulting include development biodiversity indicators, evaluation execution conservation planning strategies, ecosystem services monitoring.