So far the remoteness of the Arctic has limited the number of biological surveys in the region and thus, estimates about distributions and abundances of Arctic marine mammals (AMMs) are missing in many areas. A better knowledge about distributions of AMMs would improve the assessment of their sensitivities to the impacts of climate change and increasing human actions. We present how this data shortage can be tackled by combining several (heterogeneous) data sets within a single spatiotemporal Poisson point process framework. We demonstrate our approach with a study on distributions of polar bears, walruses and ringed seals in the Kara Sea. We combined species observations from multiple studies which had differing survey methods. Based on the data we estimated how species respond to the habitat covariates and created a hindcast of species relative densities in the study area.
Our data set was mostly based on survey cruises where researchers had made species sightings with varying effort. The novelty of our modelling methodology is in taking into account the survey bias and spatiotemporal autocorrelation, which come as downsides of utilizing an extensive but poorly controlled data (Fithian et al. 2015). We built a hierarchical Bayesian framework, which allowed us to model observations as a Poisson point process and to formulate an additive regression model for the species density process (Warton & Shepherd 2010). In the additive model we assigned fixed effects for covariates and random effects for survey specific observation bias and spatiotemporal autocorrelation.
According to our results, the (relative) density of polar bears was mostly explained by the relative density of seals. As apex predators polar bears are dependent on prey abundance, which has not been considered in earlier estimates of polar bears’ habitat suitability. Hence, the response of seals to shrinking ice cover may be an important feature for the future of the polar bear distribution. Seal density was highest in areas with ice cover around 70 % and walrus density was highest relatively near coastal regions (shallow water areas), which support that they are both dependent on access to prey. Moreover, there was strong variation assigned to both random effects. The spatiotemporal effect explained variation caused by unmeasured environmental covariates and possibly by spatiotemporally structured survey bias.
Our model structure could treat the heterogenic sampling protocols, which came with the cost of predicting only the relative densities instead of absolute ones. Anyhow, this did not have an effect on the estimates of species’ habitat characteristics. The novel methods in SDM field proved their efficiency in our study and created quantitative knowledge and new understanding about the Arctic ecosystem.
Fithian, et al. (2015). Methods in ecology and evolution, 6, 424-438.
Warton & Shepherd (2010). The Annals of Applied Statistics, 4, 1383-1402.