Creating spatially explicit predictions of bird presences in Maine: Evaluating input data, model performance, and model output

Breeding Bird Survey (BBS) data and selected environmental variables were used to predict the spatial occurrences of land bird species in Maine. First, however I tested the representativeness of BBS routes by comparing geomorphology and vegetation surrounding BBS routes to statewide patterns and random samples. Routes run in 1993 over-sampled lower elevations and 1--10% slopes, and under-sampled or did not sample elevations over 200 m or steeper slopes. Routes over-sampled grasslands and sparse residential areas and under-sampled deciduous forests, coniferous forests, coniferous forested wetlands, and open water, but biases had minimal effects on the associations between vegetation types and species survey data.;Next, Bayesian probability modeling was used to predict species occurrences for 28 species of birds from the relationships between species survey data and six environmental data layers. The associations between BBS stop data (point counts) from 1990 and unclassified 1991 Landsat Thematic Mapper (TM) satellite imagery, a 1993 vegetation map, and two measures of landscape heterogeneity formed conditional probabilities that were input into Bayes' Theorem. Forty-seven model permutations (based on all allowable combinations of the explanatory data layers) were calculated for each species. Model performance was evaluated using three criteria: difference from a priori; agreement with BBS data; and difference from a random model. Model predictions were validated against independent field data; twenty-five species were modeled satisfactorily.;The effects of species niche (Grinnellian) width on the agreement and accuracy of models were assessed for 28 land bird species using Bayesian predictions and Maine Gap Analysis Project (ME-GAP) predictions. ME-GAP species predictions were based mainly on species-habitat associations with vegetation types. Model predictions from both methods were compared with field data. Species using more types, with presumably wider niches, had higher positive agreement with field data and much lower overall agreement for ME-GAP predictions. Comparisons of the predictions for the 28 species showed that ME-GAP consistently (n = 25) overestimated species occurrence when compared to Bayesian predictions (n = 3), indicating that statistical models represent a more conservative method for predicting species presence than habitat-association methods.