Predictive modeling techniques with application to the Cerulean Warbler (Dendroica cerulea) in the Appalachian Mountains Bird Conservation Region

Abstract

Many statistical approaches have been used for developing predictive models for wildlife presence/absence and abundance, each with varying levels of accuracy and complexity. As concerns for declining species intensify and anthropogenic impacts on habitats increase, the ability to quickly quantify and map species distributions and abundances over large regions will become increasingly important. To date, there is no set of best practices for modeling specific wildlife groups. My primary objectives with this thesis were to 1) compare model techniques for ease of use and accuracy, and 2) compare resolution of species occurrence data and its effect on model accuracy. For the first objective, I compared two modeling techniques that range from moderately quick and simplistic (decision trees) to conceptually and computationally complex (hierarchical spatial models). I used North American Breeding Bird Survey counts with a suite of explanatory variables to predict presence and abundance of cerulean warblers (Dendroica cerulea) in the Appalachian Mountains Bird Conservation Region. Of the decision tree methods, cerulean warbler occurrence was most accurately described by presence/absence models. Regression tree abundance models under-predicted counts and had low accuracy. Hierarchical spatial models predicted abundance of cerulean warblers similar to actual counts, and with better overall accuracy than regression trees. All techniques produced models using similar variables; interior forest and percent forest were most important for identifying areas with cerulean warblers. For the second objective, I compared two model types, differing in the resolution of the species distribution data. I used North American Breeding Bird Survey (NABBS) counts with a suite of explanatory variables to predict presence and abundance of cerulean warblers (Dendroica cerulea) in the Appalachian Mountains Bird Conservation Region (BCR28). Decision trees were created for route-level and stop-level analyses of presence and abundance. Additionally, output maps have typically been resolved to the resolution of the environmental spatial datasets with little attention given to the scale at which the predictions represent. Using the modeling results, predictive distribution maps were created for cerulean warblers with appropriate resolutions for each model group. Route-level decision trees performed better than stop-level models for predicting both presence and abundance of cerulean warblers. Similar to raw NABBS distribution data, cerulean warblers were predicted to occur in highest concentrations in the central portions of the BCR. Poor performance of stop-level models may result from a mismatch of resolution of environmental data to species survey data, or lack of important environmental covariates at the stop-level scale. The results of this study highlight the importance of correctly matching the resolution of the species distribution data to the resolution of environmental covariates and the extent of analysis. The results and relationships hig...

CASE STUDY: Prioritization strategies for reforestation of minelands to benefit cerulean warblers

“…Fine-scale species models have been successful at small extents (Dettmers and Bart, 1999;Buehler et al, 2006) but often do not perform well when tested in other geographic regions or over broader extents (Heglund, 2002). Poor performance may result from a mismatch in the resolution of environmental data to species survey data, or lack of important environmental covariates at the point-level scale (Heglund, 2002;Shumar, 2009).…”

“…To identify potential survey areas, we developed a map of surface-mined lands in the core breeding range of the Cerulean Warbler using remotely-sensed landcover data and digitized mine permit boundaries. This was overlaid on the map of predicted Cerulean Warbler abundance developed from statistical models that related abundance to environmental characteristics (Shumar, 2009). Mines were identified from landcover layers ( The final map was used to identify mined lands that lie adjacent to forested habitat predicted to be suitable for Cerulean Warblers.…”

“…We evaluated accuracy of the Cerulean Warbler predicted abundance map developed by Shumar (2009) at two scales (the point-level and the route-level) with Pearson product-moment correlation. We report the correlation coefficient (R) which ranges 1 to -1 and values close to ±1 Because Cerulean Warblers become less vocal later in the breeding season, we divided the survey data in half and defined two periods: early (19 May-2 June) and late (6-16 June).…”

“…region, relative to Cerulean Warbler breeding distribution and relative abundance based on Breeding Bird Survey (BBS) data (Sauer et al, 2011). (B) Cerulean Warbler male abundance from transect surveys conducted in 2009 (circles) was higher within areas predicted to have higher abundance by Shumar (2009).…”

“…The FRA accelerates natural forest succession by 1creating a suitable rooting medium for tree growth, (2) limiting soil compaction by loosely grading topsoil, (3) using ground covers that encourage tree growth, (4) planting early-successional tree species and commercially-valuable crop trees, and (5) using proper tree planting techniques (Burger et al, 2005). Shumar, 2009). Additionally, insight gained from single-species models can be applied to the management of multiple species (Sheehan et al, 2013).…”