Copyright
P.(Paul) Hendricks.

Surveys for grassland birds of the Malta Field Office-BLM, including a seven-year study in north Valley County (Volume 2008) online

. (page 5 of 10)
Online LibraryP.(Paul) HendricksSurveys for grassland birds of the Malta Field Office-BLM, including a seven-year study in north Valley County (Volume 2008) → online text (page 5 of 10)
Font size
QR-code for this ebook


with significantly taller (Kruskal-Wallis ANOVA pair-wise comparisons, P < 0.05) grassland vegetation
(40.8 ± 9.2 cm and 43.7 ± 9.8 cm, respectively), than McCown's Longspur (37.4 ± 8.2 cm), which is also
consistent with their patterns in distribution across the northeastern Montana landscape. Lark Bunting,
another species occurring on a greater percentage of point counts in Valley County, also tended to occur
on points where grassland vegetation was significantly (P < 0.05) taller (41.2 ± 10.9 cm) than those occu-
pied by McCown's Longspur during 2001-2007, but it is sometimes nomadic, and greater numbers in Val-
ley County could as well be a function of spatial location relative to over-wintering areas. Assuming the
tendencies in vegetation structure associations we saw by the respective species in north Valley County
apply across all three counties, then that further supports the suggestion that spatial differences in relative
abundance of SOC birds is tied to vegetation structure.



Discussion

BLM lands in northern Blaine and Phillips counties continue to support a diversity of breeding birds,
including most or all of those we detected every year of our study to the east in north Valley County.
Bobolink was the only SOC bird present in Valley County every year but not detected on our Blaine and
Phillips counties points in 2007. However, even in Valley County Bobolink occurred on very few points
(no more than 5.3% in seven years of survey), so we may have missed them if it was limited to just a few
sites with taller and denser vegetation, which is quite possible. That aside, our survey results indicate
that the public lands administered by the BLM in the north half of Blaine and Phillips counties provide a
mosaic of grassland conditions necessary to support breeding populations of SOC birds that use sites with
relatively short grass of low density near the ground, to other species associated with tall and dense grass.

Appendix B - 4



The absence of Brown-headed Cowbird on our points in Blaine and Phillips counties may be tied to a
larger expanse of agricultural lands within the region, but the reasons for their absence remain unknown.

As our seven-year study in north Valley County shows, vegetation conditions can vary significantly from
year-to-year, so extrapolating results or patterns from a single year of counts to multiple years would be
a mistake. The primary conclusions that can be made from our single year of point counts are 1) a large
number of breeding SOC grassland birds that deserve additional conservation attention occur in northern
Blaine and Phillips counties, and 2) they comprise a significant proportion, in terms of species richness
and relative abundance, of the total grassland avifauna on public lands.



Appendix B - 5



Appendix C. Predictive Distribution Models for Grassland

Birds



Introduction

Despite a proliferation of bird point count monitoring stations across Montana and the Malta Field Office
of the BLM in recent years, baseline coverage in some areas is still poor. Furthermore, existing predicted
distribution models for birds and other vertebrates is limited to deductive or "rule based" models that do
not make use of the existing positive data available in statewide databases. Predicted distribution models
were therefore created for 11 grassland bird species in order to: (1) better understand the potential ranges
of the species in Montana and the Malta Field Office of the BLM; (2) identify areas of predicted distri-
bution that should be targeted for future surveys; (3) determine which environmental variables have the
highest predictive association with the presence of each species; and (4) identify spatial patterns in the
suitability of habitat for these species.

Modeling Approach

Program Maxent Version 2.3 (Jaynes 1957; Miroslav et al. 2004, 2005, 2007; Phillips et al. 2004, 2006;
www.cs.princeton.edu/~schapire/maxent/ ) was used to create statewide predicted distribution models
based on locations of observations for individual species in the statewide Point Observation Database and
6 continuous (elevation, slope, landsurface curvature, annual precipitation, maximum July temperature,
minimum January temperature) and 5 categorical (aspect, geology, 1992 National Land Cover Data, soil
temperature class, STATSGO soils) environmental layers. Observation locations for each species were
input in comma delimited format and environmental layers were fitted to 90 m grid cell pixels in ASCII
format. Both species observation locations and environmental layers were projected to North American
Datum 1983 in Montana State Plane.

Program Maxent Version 2.3 (Jaynes 1957; Miroslav et al. 2004, 2005, 2007; Phillips et al. 2004, 2006;
www.cs.princeton.edu/~schapire/maxent/ ) generates probability distributions for all environmental layers
over all pixels that are associated with observations of the species. These environmental "features" are
used to constrain the probability distribution computed to predict the statewide distribution of species so
that estimated distributions match characteristics of the empirical distributions resulting from the positive
data. Depending on the amount of positive data available the estimated distribution is automatically con-
strained to match the empirical average (linear feature), average and variance (quadratic feature), covari-
ance (product feature), proportional occurrence (threshold feature), or average below and constant above
a certain point (hinge feature). Categorical environmental variables have a discrete feature for every
possible value of the variable so that the estimated distributions have the same proportional representa-
tion of each categorical value. In practice Maxent 2.3 avoids overfitting of models to the training data by
"regularizing" or relaxing the feature constraints so that feature expectations only have to be close to the
empirical distributions rather than exactly equal to them. A regularization parameter can be entered in
Maxent 2.3 in order to delineate just how close of a fit is needed between empirical features and estimated
distributions.

Maxent Version 2.3 makes use of the Gibbs distribution which takes the form:

P(x) = exp(cl * fl(x) + c2 * f2(x) + c3 * f3(x). . .) / Z

Where cl, c2, c3, ... are weighted constants, fl, f2, f3, ... are the constrained environmental features, and
Z is a scaling constant that ensures that P sums to 1 over all grid cells. The Maxent 2.3 algorithm is guar-
anteed to converge to values of cl, c2, c3, ... that give the unique optimum distribution P, and, therefore,
the outputs are deterministic. The program begins with a uniform distribution (the distribution with maxi-
mum entropy) and successively modifies each weighted constant on each iteration until either the change
in the "gain" (the log of the number of grid cells minus the average of the negative log probabilities of the

Appendix C - 1



sample locations) falls below a set threshold or a set maximum number of iterations are performed. Thus,
the Maxent 2.3 algorithm finds the most uniform distribution (the one with the maximum entropy) subject
to the constraints of the environmental features. Maxent 2.3 can use every grid cell that has values for all
the environmental variables to calculate the distribution. However, because there are a large number of
90-m grid cells for Montana (10,204 columns and 5,892 rows), and the modeling performance of Maxent
2.3 does not improve significantly with very large numbers of pixels, 60,000 background pixels (added to
those with positive species data) were used to represent the variety of environmental conditions present
in the data. This limitation also serves to generalize or "regularize" the model to make sure that it is not
overfit to the limited number of locations where species were observed.

The main output is in the form of a cumulative function with grid cells having values of 100 representing
the most suitable habitat (best fit to the positive data) while those with values closer to are less suitable
habitat (worst fit to the positive data). Note that Maxent Version 2.3 output is not interpreted as prob-
ability of occupancy. Other standard model output includes: (1) an evaluation of omission error rates for
training and test data as a function of the cumulative threshold and overall predicted area; (2) a Receiver
Operating Characteristic (ROC) curve which evaluates the overall predictive power of the model; (3) a
table of potential thresholds which could be used to categorize the continuous output into a binomial pre-
dicted or not predicted output; (4) a statewide map showing the continuous probability function as well as
the training and test data points; (5) response curves for individual environmental variables showing how
each affects the output while all other environmental variables are held constant at their average sample
value; and (6) jackknife charts showing the relative importance of all environmental variables as a func-
tion of the fit of the model to the data with their exclusion or sole inclusion in the model.

Below we provide an overview of bird observations used in the modeling effort as well as comments on
the model output. We then describe the environmental input layers and provide complete model output
for individual species.



Appendix C - 2



Table CI. Overview of Bird Observations Used in Modeling Effort. The predicted distributions of 11 grassland bird Species of
Concern on the Malta Field Office of the BLMwere modeled. The table below summarizes observation records for these species,
the number of records used to train the models, and the number of records used to test the models. In general models performed
well in terms of omission and commission error rates where positive data is present in a region. However, there is reason to
believe that the models may perform less well in regions that lack surveys (i.e., they are probably over fit to regions that have pos-
itive data from surveys). All model output should be regarded as a first iteration with additional modeling needed (e.g., the exact
characteristics of the soils and geology classes need to be identified when these variables are found to be driving the models).



Scientific Name

(Bold = Species of Concern)


G Rank /
SRank


Total

No.
Records


No. Records

< 400 Meters

Locational

Uncertainty


No. Spatially

Unique Records

Used to Train

Model


No. Spatially

Unique Records

Used to Test

Model


Greater Sage Grouse
(Centrocercus urophasianus)


G4/S3


7896


7181


1177


392


Mountain Plover

(Charadrius montanus)


G2/S2B


3244


467


187


62


Long -billed Curlew

{Numenius americanus)


G5/S2B


860


510


279


92


Sprague's Pipit

(Anthus spragueii)


G4/S2B


2007


1418


412


137


Brewer's Sparrow

{Spizella breweri)


G5/S2B


1670


1327


633


210


Lark Bunting

{Calamospiza melanocorys)


G5/S3B


1545


1185


657


218


Baird's Sparrow

{Ammodramus bairdii)


G4/S2B


1737


954


335


112


Grasshopper Sparrow

{Ammodramus savannarum)


G5/S3B


996


643


356


119


McCown's Longspur
{Calcarius ornatus)


G4/S2B


684


487


209


69


Chestnut-collared Longspur

{Calcarius ornatus)


G5/S3B


2628


1739


519


173


Bobolink

{Colichonyx oryzivorus)


G5/S2B


398


135


80


26



Appendix C - 3



Descriptions of Environmental Input Layers

Input environmental layers consisted of 6 continuous (elevation, slope, curvature of land surface, an-
nual precipitation, maximum July temperature, minimum January temperature) and 5 categorical (aspect,
geology, 1992 National Land Cover Data, soil temperature class, STATSGO soils) variables. These layers
were tiled together, resampled, and converted to a statewide coverage of 90 m grid cells in ASCII format
with 10,204 columns and 5,892 rows projected to North American Datum 1983 in Montana State Plane.
Each source environmental layer is described below and, where appropriate, links to metadata are pro-
vided.

Aspect (Categorical)

Calculated using the Aspect function in ArcMap 9.2 Spatial Analyst from the 10-m National Elevation
Dataset (NED) and resampled to 90 m. See description of the elevation layer and the associated NED
metadata link below. A brief summary of raster cell value descriptions for grid bearing values follows
below.

= Flat

1 = North (337.5-22.5)

2 = Northeast (22.5-67.5)
3= East (67.5-112.5)

4 = Southeast (112.5-157.5)

5 = South (157.5-202.5)

6 = Southwest (202.5-247.5)

7 = West (247.5-292.5)

8 = Northwest (292.5-337.5)

Curvature of Land Surface (Continuous)

Calculated using the Curvature function in ArcMap 9.2 Spatial Analyst from the 90-m resampled grid
resulting from the 10-m National Elevation Dataset (NED). See description of the elevation layer and
the associated NED metadata link below. Values are continuous from -50 (concave land surface) to +50
(convex land surface).

Elevation (Continuous)

The National Elevation Dataset (NED) is a 1/3 arc-second (10-m) raster grid of decimal meter values
assembled by the U.S. Geological Survey. Metadata on the 1/3 arc-second (10-m) NED is at: http ^/seam-
less. usgs.gov/products/3arc.php

Geology (Categorical)

A polygonal coverage of surficial geology available in a mixture of 1:100,000 and 1:250,000 scales from
the Montana State Geologic Mapping Program at the Montana Bureau of Mines and Geology. Metadata
on the state geology map is at: http://www.mbmg.mtech.edu/gmr/gmr-statemap.asp

Maximum July Temperature (Continuous)

A polygonal coverage of estimated average maximum daily temperatures for July in degrees Fahrenheit,
for the climatological period 1971-2000. Estimates are based on Parameter-elevation Regressions on
Independent Slopes Model (PRISM) derived raster data which uses known point temperature data and a
digital elevation model (DEM) to generate gridded estimates of annual, monthly and event-based climatic
parameters. General information on the underlying PRISM data and the source data itself can be down-

Appendix C - 4



loaded from the Oregon Climate Service website at: http://www.ocs.orst.edu/prism/ . The Montana data
reprojected to Montana State Plane and resampled to a resolution of 600 m representing 33 temperature
ranges in degrees Fahrenheit is available at: http://nris.mt.gov/nsdi/nris/tmax71_00.html

Minimum January Temperature (Continuous)

A polygonal coverage of estimated average minimum daily temperatures for January in degrees Fahren-
heit, for the climatological period 1971-2000. Estimates are based on Parameter-elevation Regressions on
Independent Slopes Model (PRISM) derived raster data which uses known point temperature data and a
digital elevation model (DEM) to generate gridded estimates of annual, monthly and event-based climatic
parameters. General information on the underlying PRISM data and the source data itself can be down-
loaded from the Oregon Climate Service website at: http://www.ocs.orst.edu/prism/ . The Montana data
reprojected to Montana State Plane and resampled to a resolution of 600 m representing 33 temperature
ranges in degrees Fahrenheit is available at: http://nris.mt.gov/nsdi/nris/tmin71_00.html

National Landcover Data (Categorical)

The 1992 National Land Cover Data Set is based on 30-m Landsat Thematic Mapper imagery. A brief
summary of raster cell value descriptions follows below. Metadata on this layer can be found at: http://
nris.mt.gov/nsdi/nris/nlcdgrid.html

1 1 Open Water

12 Perennial Ice/Snow

21 Low Intensity Residential

22 High Intensity Residential

23 Commercial/Industrial/Transportation

31 Bare Rock/Sand/Clay

32 Quarries/Strip Mines/Gravel Pits

33 Transitional

41 Deciduous Forest

42 Evergreen Forest

43 Mixed Forest
51 Shrubland

61 Orchards/Vineyards/Other

71 Grasslands/Herbaceous

81 Pasture/Hay

82 Row Crops

83 Small Grains

84 Fallow

85 Urban/Recreational Grasses

91 Woody Wetlands

92 Emergent Herbaceous Wetlands

Annual Precipitation (Continuous)

A polygonal coverage of annual precipitation in inches for the climatological period 1961-1990 based
on source data from the National Weather Service Cooperative stations. Natural Resources Conservation
Service SNOTEL stations, and local networks. Metadata on this layer can be found at: http://nris.mt.gov/
nsdi/nris/precip.html



Appendix C - 5



STATSGO Soils (Categorical)

State Soil Geographic data (STATSGO) is a polygonal coverage of general soil associations developed by
the National Cooperative Soil Survey. The soil maps for STATSGO are compiled by generalizing more
detailed soil survey maps. Map unit composition for a STATSGO map is determined by transecting or
sampling areas and expanding the data statistically to characterize the whole map unit. Therefore, map
units depict the dominant soils making up the landscape and often contain dissimilar soil types. The ap-
proximate minimum area delineated is 625 hectares (1,544 acres). Background information and metadata
on STATSGO is available at: http://nris.mt.gov/nsdi/statsgo.pdf , http://nris.mt.gov/nsdi/nris/SS19.htmk
and http://dbwww.essc.psu.edu/doc/statsgo/statsgo_info.html . Definitions for the 694 map units used in
the input can be downloaded as .dbf files along with a STATSGO shapefile for Montana at: http://nris.
mt. gov/gi s/gi sdatalib/gi sDataLi st. aspx? datagroup=statewide-regional& search Term s=statsgo

Slope (Continuous)

Percent slope (rise over run multiplied by 100) calculated using the Slope function in ArcMap 9.2 Spatial
Analyst from the 10-m National Elevation Dataset (NED) and resampled to 90 m. See description of the
elevation layer and the associated NED metadata link above.

Soil Temperature Regime (Categorical)

A generalized polygonal coverage of soil temperature regimes. A brief summary of raster cell value
descriptions follows below. A glossary of relevant soil terminology can be found at: https://www. soils,
org/sssagloss/ . Metadata on this layer can be found at: http://soils.usda.gov/use/worldsoils/mapindex/str.
html

1 Cryic/Udic

2 Frigid/Udic

3 Frigid/Typic Ustic

4 Cryic/Typic Ustic

5 Frigid/ Aridic Ustic

6 Frigid Aquic

7 Frigid/Typic Xeric

8 Water

9 Cryic/Typic Xeric

10 Cryic/ Aridic Ustic

1 1 Mesic/Ustic Aridic

12 Cryic/Udic Ustic



Appendix C - 6



JxEFERENCES (most can be downloaded at: www.cs.princeton.edu/~schapire/maxent/ )

Jaynes, E. T. 1957. Information theory and statistical mechanics. Physics Reviews 106:620-630.

Miroslav, D., S. J. Phillips, and R. E. Schapire. 2004. Performance guarantees for regularized maximum
entropy density estimation. Proceedings of the Seventeenth Annual Conference on Computational
Learning Theory pp. 472-486.

Miroslav, D., S. J. Phillips, and R. E. Schapire. 2007. Maximum entropy density estimation with gener-
alized regularization and an application to species distribution modeling. Journal of Machine Learn-
ing Research 8 (2007): 1217-1260.

Miroslav, D., R. E. Schapire, and S. J. Phillips. 2005. Correcting sample selection bias in maximum
entropy density estimation, pp 323-330. In: Advances in Neural Information Processing Systems 18.
MIT Press.

Phillips, S. J., R. P. Anderson, and R. E. Schapire. 2006. Maximum entropy modeling of species geo-
graphic distributions. Ecological Modelling 190(3-4):23 1-259.

Phillips, S. J., D. Miroslav, and R. E. Schapire. 2004. A maximum entropy approach to species distribu-
tion modeling. Proceedings of the Twenty -First International Conference on Machine Learning pp.
655-662.



Appendix C - 7



Model Output

Map and chart outputs for the predicted modeling effort are included below for each of the 11
grassland bird Species of Concern that were modeled. The individual species models highlight
the relative importance of Valley, Phillips, and Blaine Counties to the conservation of these
grassland bird Species of Concern in Montana. This is further highlighted in the composite im-
age of predicted distributions for all species shown below.




Figure CI. Composite Image of Predicted Distributions for 11 Grassland Bird Species of Concern. The hot-to-cold color map
indicates the composite suitability of each grid cell for all 11 grassland bird Species of Concern as a function of the environmen-
tal variables at that grid cell. Hotter colors indicate areas that are predicted to have more suitable habitat for all 11 species.
BLM Field Office boundaries and county lines are included for reference.



Appendix C - 8



Greater Sage-Grouse {Centrocercus urophasianus)



TO
S

S-

X

n




Figure C2. The hot-to-cold color map indicates the suitability of each grid cell as a function of the environmental variables at that grid cell. Hotter colors indicate areas that are
predicted to have more suitable habitat for the species. Black dots are positive data used to build the model A shaded relief map, BLM Field Office boundaries, and county lines
are included for reference.



1.D

□ .7
di

I 0.5
■^0.4

□ .3

□ .2

□ .1

□ .□





Omission


vs. Predicted Area for GreaterSage-grouse
















































/






















^


/




















4


/




















A


'/










1










X


^












\








^


^














V








^
















\


s^


'^






















t^


^




















A


r








- ^


. —

































Fraction of background predicted

Omission on training sampies

Omission on test sampies

Predicted omission



10 20 30 40 50 60

Cumuiative tiiresiioi



70



80



90



100



Figure C3. An evaluation of omission error rates for training (dark blue line) and test (light blue line) data as a function of the
cumulative threshold and overall predicted area. The red line indicates the overall fraction of the map area fitting each value of
the cumulative threshold. The black line is the predicted omission rate for each cumulative threshold.



Appendix C - 10







Sensitivity vs.


1 - Specificity for GreaterSage-grouse




o.s

I0.7

£=


SO.6

E


:eo.4

■1— '

I0.3

□ .2

□ .1

□ .□


















1 1








y


^


^*










/








/


r












/








i


7












/










1












/












1










/














/








/
















/






/






















/






















/




















V






















\























Training data [AUC = 0.931)

Test data (AUC= 0.91 7)

Random Prediction (AUG = 0.5)



0.0



0.1



0.9



1.0



0.2 0.3 0.4 0.5 0.6 0.7 O.S
1 - S|3ecifi:ity (Fractional Predicted Area)

Figure C4. Receiver Operating Characteristic (ROC) curve evaluating the overall predictive power of the model with the Area
Under the Curve (AUC). The AUC value indicates that when two random locations are chosen the model has that probability
of assigning a higher cumulative threshold value to the location with more suitable habitat. The light blue line indicates how a
neutral or random model would perform (i.e., it only has a 50% probability of assigning a higher cumulative threshold value to
a random location with more suitable habitat than a random location with less suitable habitat). The further toward the top left
of the graph the training (red) and test (blue) data lines are, the better the model is at predicting the presences contained in the
training data. Sensitivity (plotted on the y-axis) is the proportion of positive locations that were correctly classified by the model.
Sensitivity is also known as the true positive rate and can be thought of as the degree of absence of omission errors. Specificity is
the proportion of random locations chosen from the background (these pseudo-absences are used instead of true negative loca-
tions) that were correctly classified by the model as negative. One minus the Specificity (plotted on the x-axis) is known as the
false positive rate and represents the commission error rate.



Appendix C - 11



TO

s

n







Respo


nse


of Create


Sage


n


ro


use


to


aspectSOm mtS3cllp




0.305


! 1 1 1 1 1
















































□ 79D
□.785

[1.7T5

anu

□ 766

□ 76^


























































































=11=










III






1 ■


J


1


1


1


T


1


1


1


1


1


1


1


1


1 1


L





















as


ec19am.


mt83tli|


1 2 3 5 7 8 9 10

Online LibraryP.(Paul) HendricksSurveys for grassland birds of the Malta Field Office-BLM, including a seven-year study in north Valley County (Volume 2008) → online text (page 5 of 10)