Modeling Habitat Relationships using Point Counts Tim Jones Atlantic Coast Joint Venture

Use of Point Counts Investigate responses of avian populations to management treatments or to environmental disturbances Estimate spatial distribution of species Model bird-habitat relationships Monitor population trends

Study Design Considerations Pure trend estimation –Systematic sampling Habitat-specific population estimate –Stratified by habitat type Bird-habitat modeling –Stratify by habitat type –Avoid edges/boundaries

Numerous good sources of information for technique

Minnesota’s Forest Bird Diversity Initiative

What’s the Problem? Timber harvesting in Minnesota began to significantly increase Forest songbirds have received little management attention

Objectives Monitor relative abundance of common bird species to assess annual changes, Define avian habitat relationships, Determine how forest management activities influence breeding bird abundance and distribution, and Provide a product that a regional wildlife biologist could use to plan forest management activities to accommodate a variety of bird species, especially those with specific habitat needs or declining populations in a region.

Monitoring Program Design Integrate with each National Forest's method of describing vegetation cover types forest stand that was > 40 acres, the minimum size needed for three point counts Fixed radius counts (100m) - although all birds detected noted 10-minute counts (3, 3-5, 5+)

Study Area

> 500 Stands

12-year Data Summary > 250,000 individuals observed 182 species detected (note about 150 forest-dependent bird species in region)

Trend Analysis Statistical analysis –Non-parametric route regression (James et al. 1996) –Uses untransformed counts –Does not assume functional form –Data for each stand smoothed (LOESS) –Fitted values averaged across stands for each year –Bootstrap 95% confidence interval (1,000 reps)

Disclaimer Counts not corrected for detectability Assumed all birds within 100m were always detected –Based on previous work in Upper Midwest Cost of double observer would have resulted in effort costing > $90,000 (> $120,000 in 2006)

Forest Number of Species Tested Number of stands Chequamegon NF50133 Chippewa NF49135 Superior NF41168 St Croix39171 Southeast40211 Regional35436

Ovenbird Regional

White-throated Sparrow Regional

Superior NF Decreasing Eastern Wood-Pewee Winter Wren Ruby-crowned Kinglet Golden-winged Warbler Black-throated Green Warbler Black-and-white Warbler Common Yellowthroat Canada Warbler Chipping Sparrow White-throated Sparrow Rose-breasted Grosbeak Increasing Black-capped Chickadee Red-breasted Nuthatch Northern Parula Magnolia Warbler Pine Warbler Swamp Sparrow

Regional Summary Yellow-bellied Flycatcher Red-breasted Nuthatch Northern Parula American Redstart Eastern Wood-Pewee Brown Creeper Winter Wren Hermit Thrush Black-and-white Warbler Ovenbird Common Yellowthroat Canada Warbler Scarlet Tanager Song Sparrow White-throated Sparrow IncreasingDecreasing

Bird-Habitat Relationship Modeling

Developing Models to Describe How Birds Respond to Forest Habitat

Species Habitat Models Range Limits Population Demographics

Habitat Characteristics Local site variables –dominant tree species, relative density estimates, foliage height diversity (fhd), percent canopy closure Landscape variables –derived from Landsat TM satellite imagery –metrics computed using FRAGSTATS –patch size, cv patch size, patch richness, Simpson’s diversity index, contagion, edge density

How species response to: Landscape variables Land cover Age of forest stand Climatic factors

100m

Habitat Relationship Models Statistical Models –Forest composition –Landscape pattern –82 species Probabilistic approach –Empirical relationship to specific habitat types –Allow unified approach for all 129 species

Statistical Methods Multiple Linear Regression – Widely used, assumes normal distribution Logistic Regression –generalized linear model (GLIM), widely used, assumes binomial distribution, loss of information Classification & Regression Trees –adaptive, but data intensive Poisson Regression –GLIM, assumes Poisson distribution, predicts either probability of occurrence or count

Common Issues in Analyzing Survey Data Small sample size Counts do not meet underlying assumptions of multiple linear regression (e.g., large spike of zero counts) Predictions not constrained by zero (i.e., negative abundance) Loss of information by converting counts to presence/absence

Blackburnian Warbler

Poisson Frequency Distributions

Evaluating Poisson Regression Simulated bird counts on real 1.1 million ha landscape Randomly sampled conifer patches to obtain counts Used Poisson and logistic regression models to fit data Compared performance of the 2 regression techniques

Simulation Details Artificial species which uses only conifer patches Assumed individuals were evenly distributed in each patch Simple functional relationship between number of individuals and patch size Did not explicitly model spatial autocorrelation between patches

Model Performance HighLow Density

Poisson Regression Poisson regression generally performed well as compared to logistic regression –except when the density is high (i.e., small territory size); underlying data approximates normal distribution –At small means (i.e., low density) Poisson regression performed as well as logistic regression without loss of abundance information

Lack of Fit and Poisson Regression Often attributed to overdisperson, which indicates that the variance and mean are not equal Or because the rate of the count variable varies between individuals (i.e., heterogeneity)

Nashville Warbler % Correctly Classified = 0.762

Summary of Explanatory Variables # Composition Patch Climate4 Landscape11 Geographic2

For more information on wide array of statistical approaches to modeling species occurrence and/or abundance:

Practical Considerations Only 30 – 45% of deviance explained Difficult to implement for: –Multiple species (with different responses) –Multiple management scenarios –Within a Monte Carlo framework - typically run 1,000 simulations to bootstrap confidence intervals

Optimal Solution Uniform approach for all 129 species of interest Easily updated with new information (i.e., new years of data collectoin) Easily linked to predictions of future habitat conditions Directly related to forest management practices

Probabilistic Modeling Concept Use 10 years of field data to generate probabilities of observing X number of individuals in sampled area (6.4ha) Probabilities are cover type specific Updated annually to reflect additional data Avoid issue of how to scale density to a given area

Sample Design Sampling unit = 6.4 ha Proportional allocation based on amount of each USFS forest type Subsample - 2 points per stand, 10 minute point count

Land Cover Classification not used jack pine red pine white pine upland mixed lowland conifer oak lowland decid aspen/birch northern hardwoods regen conifer regen decid non-forested wetland non-forested upland developed water

Observed Probability Matrix

Simulation Methods

Step 1: Subdivide Patches

Draw number from random number generator Compare to cumulative probability from field data Determine number of individuals “observed” for each “sample” area Step 2: Populate Subdivisions

Step 3: Patch Estimate For subdivisions that are not completely contained in patch, proportionally reduce estimated number of individuals Sum number of individuals across all subdivisions of a patch

Evaluation of Modeling Approach

PlotSpearman’s rho Bandana0.81 Blandin0.77 Boise0.81 Boulder Lake0.80 Clover0.69 Erin0.55 Pine0.77 Potlatch0.77 Wolf Ridge0.60 Correlation between Observed and Predicted Species Abundance

Conclusions Model approximates reality Incorporates observed variability Appears to have no systematic bias Easily implemented Easily updated as additional data become available Does not violate statistical assumptions

Summary Point counts are applicable to questions at a variety of spatial scales and geographic extents Point counts can relate habitat quantity to a measure of species’ density or relative abundance Point counts do not necessarily relate density estimates to habitat quality

Summary (cont) Point counts good for assessing adequacy of bird-habitat modeling Require long-term commitment of resources to realize adequate sample size If designed correctly allow use to assess cause of trend

Acknowledgements Gerald J. Niemi, JoAnn Hanowski, Nick Danz and Jim Lind Natural Resources Research Institute, University of Minnesota Duluth

Cooperators Funded By Legislative Commission for Minnesota’s Natural Resources