Modelling habitat selection at multiple scales with multivariate geostatistics: an application to seabirds in open sea

Abstract

Modelling habitat of species necessitates robust identification of relevant environmental variables linked to species dis tribution. To achieve this, we connect hierarchical patch theory and habitat modelling at multiple scales. We suggest discriminating between ‘circumstancial variables’ and ‘process variables’ on the basis of temporal evolution of the spatial links between species distribution and their environment at different scales. ‘Process variables’ are informative of the eco logical processes driving the distribution of organisms at multiple scales. By opposition, ‘circumstantial variable’ provide little insight because their relationship with animal spatial distribution is subject to great variations through time. As a real case study, we investigate the relationships between auk distribution (mainly Uria aalge) and oceanographic landscapes over two scales (i.e. large ~ 200 km and medium ~ 50 km) during the wintering season in the Bay of Biscay. Surface salinity, mixed layer depth and chlorophyll a are identified as ‘process variables’ as they are invariably correlated with the spatial distribution of auks, whereas bottom temperature can be viewed as a ‘circumstantial variable’ since the correlation is nonconstant through time at large scale. The process variables at large scale are used to model the potential habitat of auks in the Bay of Biscay during the wintering season. At medium scale, only the chlorophyll a is identified as ‘process variable’ and used to model preferential habitat of wintering auks in the Bay of Biscay. The analytical approach proposed here (i.e. mul tivariate factorial kriging in a temporal context) is an effective framework to model the potential and preferential habitat of a species and can be related to the ecological niche concept and by focusing explicitly on scale dependence, the distinction between the variables that can be used as niche descriptors into species distribution models. Then our method lead to the identification of variables that should be used to define the Grinnellian niche which is defined by environmental conditions on broad scales and the Eltonian niche which focuses on biotic interactions and resource–consumer dynamics that can be measured principally at local scales.