A large-scale model for the at-sea distribution and abundance of Marbled Murrelets ( Brachyramphus marmoratus) during the breeding season in coastal British Columbia, Canada

The role that the marine environment plays in the distribution and abundance of Marbled Murrelets ( Brachyramphus marmoratus), a seabird which nests in old-growth forests, is not well understood. Therefore, we investigated how Marbled Murrelet marine distribution and abundance is related to the abiotic and biotic components of the marine environment. Data on the marine distribution of Marbled Murrelets in British Columbia (BC), densities (birds/km 2; 1972–1993), counts (number of birds per survey; 1922–1989), and pertinent environmental variables as identified from the literature were compiled and then organized in a Geographic Information System (GIS). On a 10 km scale, count surveys were not correlated with density surveys ( r 2=0.01, P=0.46). This suggests the interpretation of count survey data (relative abundance) should be done with care; and it is not further used in this study. We built a parsimonious model to explain marine densities with marine predictors. First, significant predictors were identified with multivariate Generalized Linear Models (GLMs) by evaluating the shortest distances from survey locations to predictor variables. Murrelet density is higher close to sandy substrate, estuaries and cooler sea temperatures, and lower close to glaciers and herring spawn areas. Model predictors selected by using P-values and AIC include sea surface temperature, herring spawn index, estuary locations, distribution of sand and fine gravel substrates (as a proxy for sand lance distribution), and proximity to glaciers. Secondly, spatially explicit large-scale distribution model algorithms use this set of significant predictors to predict Marbled Murrelet abundance (density), distribution and populations in coastal BC. The modelling algorithms used include GLM, Classification and Regression Trees (CART) [Classification and Regression Trees, Wadsworth & Brooks, Pacific Grove, CA, 368 pp.; Software CART and MARS, San Diego, CA] and Tree (SPLUS) [Modern Applied Statistics with S-Plus, Statistics and Computing, 2nd ed., Springer, New York, 462 pp.], Multivariate Adaptive Regression Splines (MARS) [Software CART and MARS, San Diego, CA], and Artificial Neural Networks (ANNs) (SPLUS) [Modern Applied Statistics with S-Plus, Statistics and Computing, 2nd ed., Springer, New York 462 pp.]. Model performances were evaluated by backfitting, and by standardizing models. Tree-SPLUS was identified as the best performing model, and therefore used to predict the maximum