Transcript Evaluating Habitat Suitability Index (HSI) Models for

Evaluating Habitat Suitability Index (HSI) Models for Landbird Conservation Planning:
Challenges & Opportunities
Todd Jones-Farrand1, John Tirpak2, Charles Baxter2, Jane Fitzgerald3, Frank Thompson4, Dan Twedt5, and Bill Uihlein2
1USGS
Missouri Cooperative Fish and Wildlife Research Unit, University of Missouri, Columbia, Missouri, USA; 2Lower Mississippi Valley Joint Venture, US Fish and Wildlife Service, Vicksburg, Mississippi, USA;
3Central Hardwoods Joint Venture, American Bird Conservancy, St. Louis, Missouri, USA; 4USDA Forest Service, North Central Research Station, Columbia, Missouri, USA; 5USGS Patuxent Wildlife Research Center, Vicksburg, Mississippi, USA
Introduction
Verification Methods
Results:
Purpose: Link habitat conditions to priority bird numbers to assist conservation
partners translate population goals into on-the-ground objectives.
Compare average HSI values to average BBS counts across
ecological subsections using Spearman’s Rank Correlation.
Table 1 shows model selection and parameter estimate results for 4 species that represent
a range of habitat associations (forest-interior, early-successional, bottomland
hardwoods, and generalist).
Objective: Assess performance of multi-scale Habitat Suitability Index (HSI) models
for 40 species of forest and shrubland birds.
• Assess model outputs to ensure high HSI values for subsections
with high counts and low values for areas with low counts.
• Revised models as necessary.
Approach: Compare HSI predictions to existing population monitoring data sets to
(1)verify model performance and (2) determine proper scale of model application.
• Three of the 4 species could be evaluated by each method; the prairie warbler model
could not be evaluated using the R8Bird data because information was lacking.
• BBS analyses could be performed for 38 of 40 HSI models.
• R8 Bird analyses could only be performed for 20 HSI models due to either a lack of
detections for calculating density or insufficient habitat information.
Validation Methods:
BBS: Compare average HSI values to average BBS counts at 2
scales using log-linear regression in SAS. AIC model selection
used to chose appropriate link function for regression (Poisson,
Negative Binomial, or Zero-inflated Poisson).
Study Area
We focused our research in the Central Hardwoods (CH) and the West Gulf Coastal
Plain (WGCP) Bird Conservations Regions (BCR) (Figures 1 & 2).
• Ecological subsections (n=88).
• BBS = area-weighted average count per route (1994-2003)
for subsection from smoothed BBS grid.
• HSI = subsection average (Zonal Statistics tool in ArcGIS).
• Predicted count = e a + b1*HSI + b2*BCR.
• Assess model AIC compared to Null model.
• Assess sign of coefficient on HSI parameter.
Challenges
• We consider these HSI models validated because they outperform a null model and
showed a positive relationship between HSI value and the population measure in each
analysis.
• Originally, we expected our models to have better predictive power in the R8Bird
analysis due to the spatial exactness of the site-level habitat data. This was true of
some models (e.g., Acadian Flycatcher) but not others. This is due to inefficiencies in
translating habitat data collection between FIA and R8Bird (e.g. continuous canopy
cover versus 4 classes).
Table 1. Validation results for selected species using 3 evaluation data sets.
Model Performance
Figure 1. Study area Bird Conservation Regions
(BCRs) and location of public lands in the
R8Bird database.
Figure 2. Study area Bird Conservation Regions
(BCRs) and location of Breeding Bird Survey
(BBS) routes used in the analyses.
Data Sources
HSI models were built from 6 national datasets: Bailey’s Ecoregions, Forest
Inventory and Analysis (FIA) data, the National Land Cover Dataset (NLCD), the
National Elevation Dataset (NED), U.S. General Soil Map (STATSGO) data, and the
National Hydrography Dataset (NHD).
Avian population data for the evaluation came from the Breeding Bird Survey (BBS)
and the U.S. Forest Service Region 8 Bird Monitoring Protocol (R8Bird). BBS data
included smoothed maps (21,475 km2 grid cells) of average counts per route (19942003), as well as route-level average counts (1990-1995; Figure 2). R8Bird data
included point count data from 1997-2006 on 4 National Forests and Land Between
the Lakes (Figure 1). R8Bird also provided spatially exact habitat data for most
points, which we substituted for FIA in calculating HSI values.
• BBS routes (n=147)
• BBS = average count per route (1990-1994).
• HSI = average within 3 km of route (Zonal Statistics).
• Predicted count = e a + b1*HSI + b2*BCR .
• Assess model AIC compared to Null model.
• Assess sign of coefficient on HSI parameter.
Species
Acadian
Flycatcher
R8Bird: Compare HSI value to bird density at individual survey
points. using repeated measures log-linear regression in SAS.
AIC model selection used to chose appropriate distribution
(Poisson or Negative Binomial).
• R8Bird monitoring points (n=species-specific, range 5126-8521).
• HSI calculated from locally collected data and landscape
statistics (NLCD, NHD, DEM). Some site-level model
variables approximated with R8 data, some unavailable.
• Density calculated with Distance 5.0 based on 3 distance
bands (25, 50, infinity). Data stratified by Site*Year, Site, or
Year depending on sample size and best model fit.
• Predicted density = e a + b1*HSI + b2*BCR .
• Assess model AIC compared to Null model .
• Assess sign of coefficient on HSI parameter.
BBS Methods:
• No detection adjustment for counts.
• BBS grids assume abundance
declines with distance from route.
• BBS grids are model results not
intended for rigorous analyses.
• FIA uncertainty in route-level analysis
(scale mismatch).
Black-andwhite
Warbler
Blue-gray
Gnatcatcher
Prairie
Warbler
Parameter Estimates
Analysis
Dist. a
N
rb
AIC
Null DAIC c
Gen. R2 d
a
b1
b2
BBS Subsection
NB
88
0.47
83.9
2.7
0.07
0.30
2.16
0.31
BBS Route
NB
147
-287.0
6.5
0.07
0.4
1.85
0.23
R8Bird
NB
7275
9069.2
249.5
0.24
-1.59
1.24
0.16
BBS Subsection
NB
88
122.2
42.7
0.41
-1.97
6.81
0.58
BBS Route
NB
147
112.5
23.9
0.17
0.12
2.17
-1.92
R8Bird
NB
7275
8815.9
19.9
0.03
-0.99
0.33
-0.26
BBS Subsection
NB
88
-1187.0
20.0
0.24
1.23
2.99
-0.29
BBS Route
NB
147
-3813.0
1.0
0.03
1.90
1.39
-0.19
R8Bird
NB
7792
7717.7
108.0
0.11
-0.58
0.62
0.53
BBS Subsection
P
88
262.7
7.2
0.12
-1.04
8.91
0.33
BBS Route
NB
147
-11.6
11.5
0.10
-0.31
11.08
0.94
R8Bird
n/a
n/a
n/a
n/a
n/a
n/a
n/a
n/a
0.54
0.58
0.41
The distribution chosen to model the data based on AIC score and goodness-of-fit as measured by Pearson’s chi square / degrees of freedom. NB=negative
binomial, P=Poisson.
b Spearman’s rank correlation coefficient for model show were all significant at P < 0.0001.
c The difference between the AIC value for the model of interest and the null (intercept-only) model.
d The Generalized R2 from Allison (1999). It provides a measure of predictive ability ranging from 0-1 based on the likelihood ratio chi square.
a
R8Bird Method:
• Limited geographic coverage in CH
BCR
• Limited number of species with
sufficient detections
• Vegetation measurements difficult to
equate with FIA measurements
• Frequency of vegetation measurements
inconsistent across sites
• Fairly consistent habitat conditions
across points
Opportunities
BBS Methods:
• Able to assess models for most
species
• Good geographic coverage in both
BCRs
• Sampled a wider range of landscape
types (e.g., agricultural)
R8Bird Method:
• Detection-corrected densities
• Site-specific forest structure information
Conclusions
• Both evaluation data sets present us with challenges in implementation and
interpretation.
• Each method provided some indication of the usefulness the HSI models have for
conservation planning.
• Neither dataset is capable of fully testing the underlying relationships &
assumptions (i.e., hypotheses) in the models.
• Validation of habitat models is best accomplished with surveys specifically
designed for that purpose
Acknowledgments
• Thanks to M. Nelson & M. Hatfield for assistance using FIA data, F. La Sorte for
assistance with Program Distance and the R8Bird database, M. Trani for
assistance with R8Bird point locations, and W. Thogmartin & S. Sheriff for statistical
advice.