Using artificial neural networks to model the suitability of coastline for breeding by New Zealand fur seals

Abstract

previous termNew Zealand fur sealnext term (Arctocephalus forsteri) numbers and distribution were reduced by human exploitation but the species is now re-colonizing much of its former range. Pinnipeds occupy two different habitat media: the marine (feeding) and terrestrial (previous termbreeding)next term environments. Measures of geographic variation in both these environments can be modelled together to predict previous termcoastline suitabilitynext term for colonization (i.e. potential availability of previous termbreedingnext term sites). To avoid problems of non-linear modelling, we used an previous termartificial neural networknext term (ANN) to: (1) predict the previous termsuitability of coastlinenext term in South Island, previous termNew Zealandnext term to support previous termbreedingnext term A. forsteri colonies by creating a previous termmodelnext term using pup condition (measured from 20 previous termbreedingnext term colonies during 1996–98), prey distribution and abundance, bathymetry, and the type of coastal substrate; (2) compare the predicted distribution of suitable previous termcoastlinenext term for colonization from the previous termmodelnext term to the current distribution of A. forsteri colonies (n=198 colonies); and (3) using ANN inference rule extraction, determine which factors are the most influential in predicting previous termcoastline suitability.next term ANN previous termmodelnext term predictions overlapped current distributions of A. forsteri colonies in South Island. Inference rule extraction gave good predictions of colony performance (i.e. the ability to predict observed pup condition); however, they were not consistent among years in terms of the prey species constituting the rules or in the direction of the relationships. Arrow squid and octopus were important previous termmodelnext term terms in 1996 and 1997, but the direction of their coefficients in the inference rules were opposite between years. Hoki was an important term in 1997 and 1998, but it also varied in direction between years. Terms of secondary importance include the distance from sample colonies to 250 m-, 500 m- and 1000 m-isobaths. Variation in previous termmodelnext term predictions may result from climatic variation, the constant index of prey availability that was used and the potential for A. forsteri to switch main prey species among year. Resource availability appears to be a good predictor of the potential distribution of A. forsteri colonies, but future previous termmodelsnext term should attempt to incorporate indices of temporal variation in resource availability as well as population density to better predict the colonization process and understand the ecological mechanisms operating within.