summer habitat identification of an endangered bat ... · 55 weber, sparks / journal of...

53

Summer habitat identification of an endangered bat, Myotis sodalis, across its eastern range of the USA

Theodore C. Weber; Dale W. Sparks

*Theodore C. WeberThe Conservation Fund410 Severn Avenue, Suite 204Annapolis, MD 21403USAPhone: (410) 990-0175Email: [email protected]

*corresponding author

Dale W. SparksEnvironmental Solutions and Innovations, Inc.4525 Este AvenueCincinnati, OH 45232USAPhone: (513) 451-1777Email: [email protected]

ABSTRACT: Myotis sodalis (Indiana bat) is an endangered bat in the Midwest and eastern U.S.A. that hibernates in caves

or mines during the winter, and roosts primarily under the bark of trees during the summer. To help identify mitigation

and other conservation opportunities for M. sodalis, we modeled summer habitat in six regions east of the Mississippi

River. We utilized existing capture data and supplemented these records by placing acoustic detectors in areas that had

previously been undersampled. We used maximum entropy modeling (Maxent) to compare summer location points to

environmental data. We then combined Maxent’s predicted distributions with home range data to model habitat patches

using the program FunConn. Modeled habitat patches captured 81-98% of recorded M. sodalis locations. Pennsylvania

had the lowest percentage of suitable habitat (14.7%), while Indiana, Ohio, and New York-New Jersey had the highest

(35.0-38.7%). In agricultural areas, riparian forest and adjacent edges with fields appeared to provide the best habitat. In

forested areas, distance to hibernacula appeared more important. We used modeled M. sodalis summer habitat to help

identify suitable mitigation opportunities and suggest locations for future survey work.

Keywords: Indiana bat, Myotis sodalis, maximum entropy, Maxent, FunConn, habitat modeling

Journal of Conservation Planning Vol 9 (2013) 53 – 68

54

Weber, Sparks / Journal of Conservation Planning Vol 9 (2013) 53 - 68

INTRODUCTION

The Indiana bat (Myotis sodalis) was among the first cadre of species listed as endangered under the Endangered Species Act, primarily because of marked declines in hibernating populations and the fact that the majority of the species could be found in a handful of caves and mines in Missouri, Indiana, and Kentucky (USFWS 2007). Prior to the emergence of the fatal fungal disease, White Nose Syndrome (WNS), populations had declined across the southern and western portions of the range (especially Missouri), and had increased across the eastern and northern portions of the species range (Thogmartin et al. 2012). The well-defined winter range has allowed conservation efforts to be focused on caverns with either large numbers of bats or with physical characteristics that would support large numbers of bats if other problems (usually anthropogenic disturbance) were eliminated. As of this writing, Myotis sodalis showed marked declines in the northeast as it succumbed to WNS (Thogmartin et al. 2012).

Conservation efforts aimed at the summer range of the species have been limited in part because the species has a large range that includes much of the forestlands between the Appalachian Mountains and Midwest prairies. Summer occurrences are more frequent in Indiana, northern Missouri, and southern portions of Iowa, Michigan, and Illinois (USFWS 2007). Historically, these areas were vegetated in a mix of prairies, forest, and savannas (Küchler 1964). At the eastern end of the distribution, tree densities are greater, but the bat appears to be less abundant (Brack et al. 2002), likely a result of unsuitable climatic conditions (Loeb and Winters 2013). Given the large range of M. sodalis, it is likely that critical landscape elements vary across the range of the species.

Efforts to determine the summer distribution of Indiana bats have long relied on two sources: (a) recaptures of banded bats that link summer and winter areas, and (b) the distribution of suitable habitat. For example, Gardner and Cook (2002) examined county-level vegetation patterns in light of known records of summer roosts and the maximum documented migration distance available at the time. The resulting geographic model identified all counties containing at least one percent oak-hickory forest and 38 percent open land within migration distance of the priority

one hibernacula. This study provided a critical first step toward developing an understanding of the link between suitable summer habitat and migration distance from hibernacula. It has served as the basis for many additional efforts to understand this relationship, including the ones used herein.

Multiple other habitat suitability models are available for Indiana bats (Carter 2004). Some of these models addressed very local scales, like individual forest patches (Rommé et al. 1995, Farmer et al. 2002). Other efforts have focused on the existence of suitable forest type over larger landscapes (Rittenhouse et al. 2007). Only two efforts (Duchamp and Swihart 2008, Watrous et al. 2006) provide insight into habitat selection at a regional scale. A recent contribution (Loeb and Winters 2013) modeled suitable climatic conditions of the species. The current study aims to examine habitat selection at a series of scales ranging from one to multiple contiguous states.

This project is a direct outgrowth of an effort to develop a multi-state habitat conservation plan covering future construction, operation, and maintenance of natural gas facilities owned by NiSource, Inc. Other partners in this effort included the U.S. Fish and Wildlife Service (USFWS), The Conservation Fund (TCF), and 13 states. In order to address potential variability in habitat requirements for Indiana bats, we modeled habitat use considering species occurrence (a combination of existing records and new samples obtained by using AnaBat detectors for areas where records were absent) and eighteen environmental variables. We based our model on a maximum entropy algorithm with the aim of predicting M. sodalis occurrences in regions of the species’ eastern range. Finally, we used a software program called FunConn to delineate potential habitat patches.

METHODS

Study Area

We confined our analyses to the states covered by the habitat conservation plan. Rather than model each state separately, we combined them into six regions (Figure 1; Table 1). We attempted to group areas with similar physiography and climate, and also sought a minimum

55


of 50 occurrence points per region. It is also noteworthy that many of the records of Indiana bats are a direct result of studies aimed at regulatory compliance with the Endangered Species Act, which can result in adjacent

states having markedly different levels of sampling. Thus, we also combined areas with similar regulatory environments. We could not model the entire study area at once because of computer limitations.

Figure 1: The six regions where we modeled M. sodalis summer habitat. There was some overlap between regions. The states covered by the NiSource habitat conservation plan are labeled.

56


This project encompassed portions of five physiographic provinces (Fenneman 1938). Western portions of Kentucky and Tennessee are part of the Gulf Coastal Plain, while the central parts of the states belong to the Interior Lowland Plateau. Western Ohio is within the Central Lowland. Portions of eastern Tennessee, Ohio, and Kentucky, as well as areas of Pennsylvania, New York, West Virginia, Virginia, and Maryland, belong to the Appalachian Plateau. Finally, portions of the project in West Virginia, Virginia, and Tennessee, fall within the Valley and Ridge Province. Forest cover differed dramatically across the project area, with the flatter areas dominated by agriculture and the Appalachian Mountains dominated by forest.

Occurrence Data

Based on Hernandez et al. (2006), we sought at least 50 occurrences for modeling, and judged this number to be adequate (Wilson VanVoorhis and Morgan, 2007). We used existing M. sodalis summer locational data obtained from USFWS and state natural resource agencies. We omitted records prior to 1990 (these areas may no longer be suitable), records lacking spatial coordinates, occurrences with poor estimated viability (if available via Biotics databases), and transient roost and mist-net records that occurred during times when the species is known to be migrating. In cases where multiple potential points were associated with a single colony or a single individual, we selected those that were most informative by removing records <1 km from others. In these cases, we preferentially selected maternity roosts over other types of roosts, roosts over feeding and

Region name Region description Area (km2)

Indiana Indiana 94,321

Ohio Ohio, except the islands in Lake Erie, plus the northern panhandle of WV (Brooke, Hancock, Marshall, and Ohio counties), plus a 3 km buffer.

113,293

Kentucky-Tennessee

Tennessee and Kentucky, plus a 3 km buffer, plus far western Virginia (Buchanan, Dickenson, Lee, Russell, Scott, Washington, Wise, Bristol, and Norton counties), because this area seemed to share bat populations with Kentucky.

230,154

West Virginia

West Virginia, except the northern panhandle (Brooke, Hancock, Marshall, and Ohio counties), which were included with Ohio. We also included all Virginia and Maryland counties bordering West Virginia, except Buchanan and Loudoun.

80,674

Pennsylvania Pennsylvania, plus a 10 km buffer to include locations in northern Carroll County, Maryland (a shared population). 132,874

New York-New Jersey New York, plus most of northern New Jersey (Bergen, Essex, Hunterdon, Mercer, Middlesex, Morris, Passaic, Somerset, Sussex, Union, and Warren counties), plus a 3 km buffer.

137,725

Table 1: Regions where we modeled M. sodalis.

57


commuting locations, and confirmed locations over those estimated via triangulation. When multiple bats were tracked to an area, we selected females over males, adults over juveniles, multiple year observations over a single year, multiple night observations over a single night, more recent locations, and points in the center of clusters.

After mapping these data, we identified gaps in the distribution where Indiana bats may occur but had not been verified. There were too many gaps to sample during one season using traditional mist-netting techniques. Therefore, we deployed clusters of AnaBat detectors at the best available commuting and foraging habitat within these gaps. We stored calls on portable flash cards and regularly uploaded them to the servers at Environmental Solutions and Innovations (ESI) in Cincinnati, Ohio. All calls were analyzed using a pair of statistical filters. The first filter removed files containing lots of noise caused by wind, insects or other non-bat sources (Britzke and Murray 2000). The second filter, developed by the Kentucky Field Office of the USFWS as a regulatory aid, identified probable Indiana bats at sites where netting efforts did not detect the species. The tool is based on a discriminant function model which has been widely used as an aid to the identification of free-ranging bats (Francl et al. 2011, Britzke et al. 2002, 2005; 2011). The filter was designed by USFWS to be highly conservative and thus prone to Type II error. In a practical sense, however, one must be aware that Type I errors do occur when many calls of confusing species are recorded at a single site.

Environmental Variables

We calculated 18 variables within each study area (Table 2), using Model Builder for ArcGIS (version 9.3.1; ESRI, 2009). For surrounding landscape features, we based a 1 km radius on the average foraging range of 11 individuals tracked by Sparks et al. (2005), and a 3 km radius on their averaged maximum linear distance from roost. Values from this study are similar to those observed in other telemetry studies (Sparks et al. 2004, Menzel et al., 2005), but were collected from an area where suitable habitat occurs next to extensive urban development. Because of computer limitations, we had to resample land cover variables to a resolution of 90 m for 3 km focal statistics.

Maxent Modeling

Maximum entropy modeling (performed using the software Maxent; Phillips et al. 2004; 2006) is a machine learning technique that can be used to predict the geographic distribution of animal or plant species or other entities of interest. Maxent compares a set of samples from a distribution over a defined space (such as recorded locations of a particular species), to a set of features, such as relevant environmental variables over that same space. Maxent estimates the spatial distribution of the species (i.e., predicts suitable habitat) by matching the relationships between recorded occurrences and underlying variables. We used version 3.3.2 of the program by Dudik et al. (2010). The software weighted each point equally.

Maximum entropy modeling does not require reliable species absence data, which are rarely available (Elith et al. 2006, Phillips et al. 2006, Baldwin 2009). Maxent can produce solutions with a small set of observations, although the larger the number of observations, the more accurate the model output is likely to be. For example, Pearson et al. (2007) reported significant predictive ability with sample sizes as low as five, but cautioned that uncertainty might be high in such cases. Comparing 12 different models of 46 species distributions, Wisz et al. (2008) found that model accuracy decreased and variability increased for all algorithms, including Maxent, as sample size decreased. No algorithm predicted consistently well with a small sample size (n = 10), although Maxent performed better than alternative techniques. The benefit of additional locations in Maxent appears to plateau around 50 (Hernandez et al., 2006). Thus we set an apriori goal of obtaining at least 50 samples per geographic region.

We used 10-fold cross-validation and examined the mean and standard deviation of the replicate runs. Although the use of 10 replications was somewhat arbitrary, it was based on standard practice, as well as time limitations. We used Maxent’s default settings, except we used all possible variable relationships (linear, quadratic, product, threshold, and hinge). We set output format to logistic.

58


Table 1: Regions where we modeled M. sodalis.

Variable name Variable description Source data

PCT_FORWET_1K Percent forested wetlands within 1 km 2001 National Land Cover Data (NLCD), with a 30 m cell size

PCT_FORSTR_1K Percent area within 1 km of blocks of deciduous forest and forested wetlands containing unchannelized streams or rivers

NLCD and 1:24,000 National Hydrography Dataset (NHD); Ftype = STREAM/RIVER or ARTIFICIAL PATH

PCT_DEVEL_1K Percent development within 1 km NLCD

PCT_EDGES_1K Length of forest edges with fields or open water within 1 km NLCD

STR_DFOR_M_1K Length of unchannelized streams or rivers in deciduous forest within 1 km NLCD and NHD

PCT_FORWET_3K Percent forested wetlands within 3 km NLCD

PCT_FORSTR_3K Percent area within 3 km of blocks of deciduous forest and forested wetlands containing unchannelized streams or rivers

NLCD and NHD

PCT_DEVEL_3K Percent development within 3 km NLCD

PCT_EDGES_3K Length of forest edges with fields or open water within 3 km NLCD

STR_DFOR_M_3K Length of unchannelized streams or rivers in deciduous forest within 3 km NLCD and NHD

DIST_MAJ_RDS Distance from major roads ESRI’s StreetMap USA (except Indiana: IN DOT roads with Average Annual Daily Traffic >10,000)

DIST_HIBERNAC

Distance to known winter hibernacula with at least 100 M. sodalis individuals recorded in surveys between 2000-2008. We used 100 as a cutoff based on the Kentucky Bat Working Group.

USFWS + state heritage programs

JUNSOL_FMN_1K Mean June insolation within 1 km Calculated by TCF from 30m U.S. Geological Service Digital Elevation Models (DEMs) and latitude

JN_SOLAR June insolation at grid cell Calculated by TCF from 30m DEMs and latitude

TMIN_06 June min. temperature Climate Source, ~400m resolution, 1971-2000 average

TMAX_06 June max. temperature Climate Source, ~400m resolution, 1971-2000 average

TMEAN_06 June mean temperature Climate Source, ~400m resolution, 1971-2000 average

PRECIP06 June mean precipitation Climate Source, ~400m resolution, 1971-2000 average

59


FunConn Modeling

The FunConn modeling toolbox for ArcGIS (Theobald et al., 2006) provides a spatial context for wildlife species, using graph theory approaches to define functional habitat patches and landscape connectivity. FunConn uses a habitat quality layer, along with the distance that the animal moves in the landscape while foraging (i.e., its home range), to delineate habitat patches that would be used by the organism. These patches contain core habitat as well as some more marginal habitat that is traversed while seeking food, water, shelter, and other needs.

We used the Maxent average model output, converted to an integer grid between 0 and 100, as the initial habitat quality layer for FunConn. We then modified it by giving major roads and developed land a habitat value of 0 (i.e., the lowest possible suitability). We set the minimum patch size as the average foraging range (335 ha) of 11 Indiana bats tracked by Sparks et al. (2005); and the foraging radius to the average maximum linear distance that these bats traveled from their roost (3020 m). Because this study took place at an urban/rural interface where extensive management was required to protect a colony, it likely also represents a minimal amount of suitable habitat that is needed to support a colony (Sparks et al. 2009).

However, it is noteworthy that overall nocturnal behavior at this site has proved typical of that observed by other researchers throughout the range (Sparks et al. 2004, Menzel et al. 2005, Watrous et al. 2006, Butchkoski and Hassinger 2002). We examined various resource quality thresholds, generally selecting the average model output that captured at least 90% of occurrences while capturing the least amount of area (i.e., attempting to minimize Type I and II errors).

RESULTS

Locational Data

We collected 642 nights of AnaBat data in eight states. This yielded 241,466 files that survived the noise filtering process. The Indiana bat filter identified probable Indiana bats at 68 sites. These included 21 occurrences in the Ohio region, 4 in the Kentucky-Tennessee region, 11 in the West Virginia region, 27 in the Pennsylvania region, and 5 in the New York-New Jersey region. We combined this with existing locational data collected with more traditional field-based techniques to obtain a grand total of 629 sites that were incorporated into models (Table 3).

Table 3: Number of M. sodalis locations used for Maxent models, by region.

Region name USFWS and state locations

AnaBat survey locations

Total locations

Indiana 204 Did not sample 204

Ohio 64 21 85

Kentucky-Tennessee 100 4 104

West Virginia 15 11 26

Pennsylvania 23 27 50

New York-New Jersey 155 5 160

60


Maxent Models

Figure 2 shows Maxent modeled habitat suitability for Myotis sodalis in Indiana, and Figure 3 shows modeled suitability throughout the study area. Output was discontinuous between the separate regions. Table 4 lists the training and test area under the receiver operating characteristic curve (AUC) for the replicate runs for each region. Training AUC’s varied between 0.909 and 0.974, and test AUC’s between 0.788 and 0.930. Table 4 also lists the variables contributing 5% or greater to the Maxent models, in descending order of relative contribution, and the relationship between the variables and increased probability of M. sodalis presence. Variable contributions are heuristic estimates, may be correlative rather than causative, and should be interpreted with caution. Values in Table 4 were averages over the ten replicate runs.

In the western regions (Indiana, Ohio, and Kentucky/Tennessee), riparian forest (e.g., PCT_FORSTR_1K) appeared important. Forest-field edges were important in Indiana, Ohio, and New York-New Jersey. Distance to hibernacula was important in all the regions except Indiana, although the relationship was not always intuitive, with alternating maxima and minima. Myotis sodalis was generally less likely to be found near developed land. Solar warming (insolation) was an important variable in West Virginia, Pennsylvania, and New York-New Jersey. Finally, precipitation and maximum temperature were important in West Virginia and Pennsylvania.

We did not have enough M. sodalis locations in West Virginia for cross-validation; thus this model should be considered preliminary only. We removed distance from hibernacula as a variable in this model after observing strong overfitting with its use. The remaining top variable, PCT_FORSTR_3K, conflicted with models in other states, exhibiting declining probability of M. sodalis presence as nearby area of deciduous forest and forested wetlands containing unchannelized streams or rivers increased.

Figure 2: Maxent modeled habitat suitability for Myotis sodalis in Indiana.

Figure 3: Maxent modeled habitat suitability for Myotis sodalis throughout the study area.

61


Region Mean training AUC

Mean test AUC

Variables contributing >5% to the mean Maxent model

Percent contribution

Values corresponding with increased probability of M. sodalis presence

Indiana 0.946 0.878 PCT_FORSTR_1K 39.5 >22%

STR_DFOR_M_1K 16.3 >1.6 km, then continued increase

PCT_EDGES_3K 8.6 peak around 35%

PCT_FORSTR_3K 6.0 >10%, then continued increase

Ohio 0.918 0.788 STR_DFOR_M_1K 13.9 continued increase, but anomalous drop at 6 km

DIST_HIBERNAC 13.8 peak around 100 km

PCT_FORSTR_1K 12.6 >20%, then continued increase

PCT_FORSTR_3K 12.1 >15%, further increase >60%

STR_DFOR_M_3K 9.4 Continual increase, with anomalous peak ~0.5 km

PCT_EDGES_3K 5.9 >5%, further increase >55%

PCT_EDGES_1K 5.5 Values did not produce much gain in the model

PCT_DEVEL_3K 5.2 <70%

Kentucky-Tennessee 0.949 0.886 DIST_HIBERNAC 26.0 <25 km, but also with a peak

around 110 km

PCT_DEVEL_1K 13.1 Rapid decline between 0-5%, and slower decline after that

PCT_FORSTR_1K 12.9 Continual increase, especially >25%

PCT_FORWET_1K 8.4 Peak at 50-55%

STR_DFOR_M_3K 5.9 Increase >1800 m, but with a trough around 3000-5000 m

DIST_MAJ_RDS 5.3 <6 km

PCT_FORWET_3K 5.3 Peak around 30%

Table 4: Maxent AUC values by region, and relative contributions of the environmental variables.

62


West Virginia 0.909 No Data PCT_FORSTR_3K 23.4

Variable response conflicted with models in other regions declining as values increased

PCT_DEVEL_1K 13.3 Logistic decline with increasing nearby development

JN_SOLAR 12.5 <180,000

STR_DFOR_M_3K 7.6 >1 km

PRECIP06 6.3 Peak around 95 mm

TMAX_06 6.0 >26 C

PCT_FORSTR_1K 5.5 Variable response conflicted with models in other regions, declining as values increased

Pennsylvania 0.974 0.895 PRECIP06 19.0 Peak around 105 mm

DIST_HIBERNAC 15.4 Peaks around 90-150 km and at ends (0 km and 200 km)

PCT_FORSTR_3K 11.6 Peak at 5%, then a decline

PCT_DEVEL_3K 7.9 Peak at 0%, followed by rapid decline

TMAX_06 6.5 Peaks at lowest end (21.6 C) and at 27.7 C, with lowest probability around 22.4-24.0 C

STR_DFOR_M_1K 5.7 Continual increase as stream length increases, with a dip ~500-4200 m

JUNSOL_FMN_1K 5.6 Peak around 178000-182000

New York-New Jersey 0.960 0.930 DIST_HIBERNAC 36.4 <60 km, especially <40 km

PCT_EDGES_3K 11.2 >23%

PCT_FORWET_3K 8.6 >9%

JUNSOL_FMN_1K 8.3 peak around 177,000

PCT_DEVEL_3K 6.9 peak around 10%

PCT_FORWET_1K 6.0 gradual increase after 5%

Table 4 continued

63


FunConn Models

Figure 4 shows FunConn modeled habitat patches for Myotis sodalis throughout the study area. Table 5 lists the FunConn resource quality threshold, number of habitat patches, the percent of land occupied by these patches,

and the percent of recorded M. sodalis locations within these patches, for each region. Modeled habitat patches captured 80.8-98.0% of recorded M. sodalis locations, depending on the region. Pennsylvania had the least percentage of suitable habitat (14.7%); Indiana, Ohio, and New York-New Jersey had the highest (35.0-38.7%).

Figure 4: Modeled habitat patches for Myotis sodalis.

Table 5: FunConn resource quality threshold, number of habitat patches, the percent of land occupied by these patches, and the percent of recorded M. sodalis locations within these patches, for each region.

Region Resource

quality threshold

Number of patches

Percent of study area contained

Percent of M. sodalis locations captured

Indiana 17.7 277 36.8 97.1

Ohio 30.0 496 38.7 86.5

Kentucky-Tennessee 23.6 624 22.2 92.3

West Virginia 44.5 172 31.0 80.8

Pennsylvania 33.7 148 14.7 98.0

New York-New Jersey 4.8 230 35.0 96.3

64


DISCUSSION

Predicted Summer Habitat for M. Sodalis

The summer habitat models for M. sodalis varied by region. Model output was discontinuous from one region to another (e.g., between Pennsylvania and its neighbors). This probably reflects a combination of differences in which locational data was available as well as real regional differences.

In Indiana and Ohio, which are primarily agricultural states, riparian forest and adjacent edges with fields appeared to provide the best habitat. This was also the case in the Gulf Coastal Plain of Kentucky and Tennessee, where the landscape is also dominated by larger, relatively flat fields. It is also noteworthy that both areas are known to contain multiple maternity colonies (USFWS 2007). In more rugged portions of Kentucky and Tennessee, modeled habitat and summer locations alike tended to be near hibernacula. Historically, this area has been known to support few maternity colonies, but males are not unusual (USFWS 2007, Brack et al. 2002). We had few locations in Tennessee, which is on the periphery of the species range; indicating this state may have influenced the model in Kentucky. Indiana bats in Tennessee likely winter in Kentucky hibernacula. We did not have enough locations in West Virginia to confidently draw conclusions in that region, but our results may indicate that the bats are limited by open space that is often used for foraging (Sparks et al. 2004).

In the fragmented northwest corner of Pennsylvania, modeled habitat resembled that in Indiana and Ohio, identifying riparian forest and adjacent edges with fields. In the forested center of the state, modeled habitat and summer locations tended to be near hibernacula. In the southeast corner of the state, modeled habitat concentrated largely in rural agricultural valleys and Piedmont, reflecting the concentration of observations in the fragmented Gettysburg area. Finally, the model selected much of the northeast corner, which is a rural mix of farm fields, forest, and lakes that ESI sampled intensively. As a caveat, qualitative analyses identified hundreds of M. lucifugus (the closely related little brown bat) and few M. sodalis calls at some of the northeast Pennsylvania sites, making call filter misidentifications likely. Concurrent with our

study, large-scale sampling efforts using mist-nets failed to detect Indiana bats in this region (Francl et al. 2011), but did detect Indiana bats in adjacent portions of New Jersey. At those sites where bat detectors were independently deployed, Indiana bats were somewhat regularly identified. Both this section and the northwest corner were glaciated, creating a less rugged topography than the center of the state. A noteworthy exclusion was extreme southwestern Pennsylvania. This region has been home to a stable colony of Indiana bats since at least 2007, and several new colonies have been detected since this study was completed (V. Brack Unpublished) .

The Maxent and FunConn models identified about a third of the New York-New Jersey region as suitable summer habitat, mostly within 60 km of hibernacula, which is consistent with the hypothesis that this area has only recently been occupied by the species (Watrous et al. 2006). The Adirondack and Catskill mountains were not selected as viable habitat and may be too cold and wet for M. sodalis, as put forth by Watrous et al. (2006). Predicted habitat somewhat resembled the potential range depicted by New York Natural Heritage Program (2009), with the greatest difference being predicted habitat in western New York. The western addition was based on our detection of Indiana bats in this region where they were not previously recorded. However, much like north-central Pennsylvania, many calls of little brown bats were identified during the qualitative analysis.

Individual Variable Contributions

Some of the variable relationships were consistent with the literature cited in the introduction. Riparian forest, a key predictor in agricultural areas, is more likely to contain older trees than upland forest because it is less suitable for farming (Weber and Boss 2009), and may have been retained for nutrient reduction, erosion control, fuel wood, or flood control purposes. In agricultural landscapes, most remaining contiguous forest is usually along streams. Duchamp and Swihart (2008) found that within the highly fragmented landscape of the upper Wabash River basin, Indiana, Myotis spp. were found in less fragmented forest. They wrote that bats like Myotis spp. adapted to flying in forest rely on this habitat for survival. Furthermore, floods

65


may kill trees, creating snags ideal for bat roosts. Finally, water in streams, rivers, ponds, and wetlands provides a source of aquatic insects that are commonly contained in the diet (Sparks et al. 2004).

Forest edges with fields and water, another predictive variable, are utilized by bats for foraging. National Land Cover Data could not accurately discriminate between agricultural fields and meadows (Weber and Allen 2010), but foraging is more concentrated above and around foliage surfaces than over the fields themselves (LaVal et al. 1977). The probability of M. sodalis occurrence declined in the heavily forested Pennsylvania and West Virginia regions, as the nearby percentage of deciduous forest and forested wetlands containing unchannelized streams or rivers increased. This may have nothing to do with streams; rather, M. sodalis may prefer an intermediate level of fragmentation, with not only adequate roosting resources but also adequate forest edges for foraging. Sparks et al. (2004) predicted that this species would preferentially use open habitats in heavily forested areas.

Distance to hibernacula was also a strong predictor of M. sodalis presence. In New York, for example, all hibernacula with at least 100 bats in pre-White Nose Syndrome surveys had clusters of summer locations within about 30-60 km, except the cave in Albany County, which was listed as an ecological trap. Other regions had mid-range peaks. The biological reason for this, if there was one, remains unclear. Myotis sodalis was generally less likely to be found in urban areas. This is consistent with previous studies (Menzel et al. 2005; Sparks et al. 2005) and may be the result of limited food resources, light pollution, or even competitive interactions with big brown bats.

However, many variable responses were either counterintuitive (e.g., having multiple peaks or responses contrary to expectations) or fitted around values that correlated with particular geographic areas. For example, the Pennsylvania habitat model avoided the rainiest portions of the state, which were also cold and on higher elevations, but the cold, northeast corner of the state contained modeled habitat. The relationship with temperature had multiple, non-intuitive peaks in Pennsylvania, yet this variable contributed to the probability distribution in Maxent.

In summary, our models were not mechanistic, and variable responses may have been correlative rather than causative. While Maxent is a useful tool for predicting spatial distributions, the contributions of individual variables to Maxent models must be interpreted with caution. Variables were interrelated, and combined in a variety of ways to create the models. In this study, they did not always correspond to plausible ecological mechanisms.

According to Elith (2011), Maxent “is unlikely to be improved” by variable pre-selection, and intercorrelation is not the issue it is with procedures like stepwise regression. After conducting this study, we reconsidered this and now (for subsequent studies) cull predictive variables that are more than 80% correlated, particularly where the initial list of variables is large compared to the number of sample points (Weber, unpublished data). While the distribution model may not be affected by correlations between variables, interpretability can be.

Model Applications

We used modeled M. sodalis summer habitat to help identify suitable mitigation opportunities and suggest locations for future survey work. Because this species is endangered, we purposely preferred to minimize errors of omission over errors of commission. Therefore, we strongly recommended field verification within the modeled habitat patches before making specific conservation decisions. Suitable conservation sites should include primary and secondary roosting trees, foraging locations, drinking water, and other habitat needs. Suitable roost trees cannot be identified remotely, only inferred; and identification of such microhabitat features requires ground level surveys. For mitigation of projected NiSource impacts, regional biologists preferred conservation sites containing confirmed M. sodalis hibernacula or roosts, at least 40 suitable roost trees/ha (either live or dead, as long as they provide shelter with sun exposure), at least 60% canopy closure, streams or other water bodies, suitable travel corridors, recruitment of preferred roost tree species, adjacency to existing protected land, large property sizes, and low likelihoods of on-site or nearby human impacts such as large-scale clearcutting or development (The Conservation Fund, 2010).

66


Finally, White Nose Syndrome continues to decimate bat populations across the study area. For many of the areas, this research effort may represent the last opportunity to obtain distributional data prior to (or even continuous with) the emergence of wide-spread WNS fatality. At the time of this writing, we intend to update the M. sodalis summer habitat model for Ohio and Indiana, and to expand it to part of the species’ western range.

ACKNOWLEDGMENTS

The authors thank Virgil Brack, David Jeffcott, Adam Fannin, Cory Murphy, and Jodi Farrell Sparks of Environmental Solutions & Innovations, Inc. (ESI) for their help collecting data in the field. John Tempone, Daniel Cox, Casey Swecker, and Justin Boyles, also of ESI, helped analyze the AnaBat data. Jazmin Varela, Strategic Conservation Planning Information Manager at The Conservation Fund, and Michael Schwartz, Senior Environmental Associate at The Conservation Fund’s Freshwater Institute, provided GIS data, as did USFWS and several state agencies. Michael Dougherty, GIS Programmer Analyst at West Virginia Division of Natural Resources, provided advice regarding Maxent. Will Allen, Director of Strategic Conservation at The Conservation Fund, provided overall project management.

This research was conducted as part of a project, titled “Development of a Multi-Species Habitat Conservation Plan for NiSource Natural Gas Transmission Facilities in Cooperation with NiSource Gas Transmission and Storage,” undertaken by The Conservation Fund and the U.S. Fish and Wildlife Service, with inputs from the natural gas company NiSource and the states of Delaware, Indiana, Kentucky, Louisiana, Maryland, Mississippi, New Jersey, New York, Ohio, Pennsylvania, Tennessee, Virginia, and West Virginia. This overall project aimed to create a multi-species, multi-state habitat conservation plan covering future construction, operation, and maintenance of NiSource’s pipelines and ancillary facilities. Funding for this project was provided by the U.S. Fish and Wildlife Service from the Cooperative Endangered Species Conservation Fund, via the State of Indiana.

LITERATURE CITED

Baldwin, R. A. 2009. Use of maximum entropy modeling in wildlife research. Entropy 11, 854-866.

Bat Conservation International. 2001. Bats in eastern woodlands. Bat Conservation International, Inc., Austin, TX.

Brack, V., Jr., C. W. Stihler, R. J. Reynolds, C. M. Butchkoski, and C. S. Hobson. 2002. Effects of climate and elevation on distribution and abundance in the Mideastern United States. in The Indiana Bat: Biology and Management of an Endangered Species (A. Kurta and J. Kennedy, eds.) Bat Conservation International, Austin, Texas

Britzke, E. R. 2005. Acoustic surveys of bats in the eastern United States. in Indiana Bat and Coal Mining, A Technical Interactive Forum (K.C. Vories and A. Harrington, eds.) U.S. Department of Interior, Office of Surface Mining. Alton, Illinois.

Britzke, E. R., and K. L. Murray. 2000. A quantitative method for selection of identifiable search-phase calls using the Anabat system. Bat Research News 41, 33–36.

Britzke, E. R., K. L. Murray, J. S. Heywood, L. W. Robbins. 2002. Acoustic identification. in The Indiana Bat: Biology and Management of an Endangered Species (A. Kurta and J. Kennedy, eds.) Bat Conservation International, Austin, Texas

Britzke, E. R., J. E. Duchamp, K. L. Murray, R. K. Swihart, and L. W. Robbins. 2011. Acoustic identification of bats in the eastern United States: A comparison of parametric and nonparametric methods. Journal of Wildlife Management 73, 660-667.

Butchkoski, C. M., and J. D. Hassinger. 2002. Ecology of a maternity colony roosting in a building. . in The Indiana Bat: Biology and Management of an Endangered Species (A. Kurta and J. Kennedy, eds.) Bat Conservation International, Austin, Texas

Carter, T. C. 2004. Summer Habitat Assessment. in Indiana Bat and Coal Mining: A Technical Interactive Forum (K.C. Vories and A. Harrington, eds.). U.S. Department of the Interior, Office of Surface Mining. Alton, Illinois.

Clark, F., et al. 2006. A GAP analysis of Indiana: December 2006 final report. U. S. Fish & Wild. Serv., Bloomington, IN.

67


Conservation Fund, The. 2010. Determining mitigation needs for NiSource natural gastransmission facilities ‐ implementation of the multi‐species habitat conservation plan (MSHCP): mitigation site report - Indiana Bat. The Conservation Fund, Chapel Hill, NC.

Duchamp, J. E. and R. K. Swihart. 2008. Shifts in bat community structure related to evolved traits and features of human-altered landscapes. Landscape Ecology 23, 849-860.

Dudik, M., S. Phillips, and R. Schapire. 2010. Maxent, Version 3.3.2. http://www.cs.princeton.edu/~schapire/maxent/ .

Elith, J., Graham, C. H., Anderson, R. P., Dudík, M., Ferrier, S., Guisan, A., Hijmans, R. J., Huettmann, F., Leathwick, J. R., Lehmann, A., Li, J., Lohmann, L. G., Loiselle, B. A., Manion, G., Moritz, C., Nakamura, M., Nakazawa, Y., Overton, J. McC., Peterson, A. T., Phillips, S. J., Richardson, K. S., Scachetti-Pereira, R., Schapire, R. E., Soberón, J., Williams, S., Wisz, M. S. and Zimmermann, N. E. 2006. Novel methods improve prediction of species’ distributions from occurrence data. Ecography 29, 129-151.

Elith, J., S. J. Phillips, T. Hastie, M. Dudík, Y. E. Chee, and C. J. Yates. 2011. A statistical explanation of MaxEnt for ecologists. Diversity and Distributions 17, 43–57.

ESRI. 2009. ArcGIS 9.3.1. ESRI, Redlands, CA.

Farmer, A. B. S. Cade, and D. F. Staufer. 2002. Evaluation of a habitat suitability model. in The Indiana Bat: Biology and Management of an Endangered Species (A. Kurta and J. Kennedy, eds.) Bat Conservation International, Austin, Texas

Fenneman, N. M. 1938. Physiography of the eastern United States. McGraw-Hill, New York, New York. 714 pp.

Francl, K E, C. Bland, J. S. Lucas, and V. Brack, Jr. 2011. Comparison of survey techniques for documenting summer bat communities in Pennsylvania and New Jersey. Journal of the Pennsylvania Academy of Science 85, 52-56.

Gardner, J. E., and and E. A. Cook. 2002. Seasonal and geographic distribution and quantification of potential summer habitat. in The Indiana Bat: Biology and Management of an Endangered Species (A. Kurta and J. Kennedy, eds.) Bat Conservation International, Austin, Texas

Hernandez, P. A., C. H. Graham, L. L. Master , and D. L. Albert. 2006. The effect of sample size and species characteristics on performance of different species distribution modeling methods. Ecography 29, 773-785.

LaVal, R. K., R. L. Clawson, M. L. LaVal, and W. Caire. 1977. Foraging behavior and nocturnal activity patterns of Missouri bats, with emphasis on the endangered species Myotis grisescens and Myotis sodalis. Journal of Mammalogy 58, 592-599.

Loeb, S. C. and E. A. Winters. 2013. Indiana bat summer maternity distribution: effects of current and future climates. Ecology and Evolution 3, 103–114.

Menzel, J. M., W. M. Ford, M. A. Menzel, T. C. Carter, J. E. Gardner, J. D. Gardner, and J. E. Hofmann. 2005. Summer habitat use and home-range analysis of the endangered Indiana bat. Journal of Wildlife Management 69, 430-436.

New York Natural Heritage Program. 2009. Online Conservation Guide for Myotis sodalis. Available from: http://www.acris.nynhp.org/guide.php?id=7405. Accessed May 27, 2010.

Pearson, R. G., C. J. Raxworthy, M. Nakamura, and A. T. Peterson. 2007. Predicting species distributions from small numbers of occurrence records: a test case using cryptic geckos in Madagascar. Journal of Biogeography 34, 102-117.

Phillips, S. J., M. Dudík, and R. E. Schapire. 2004. A maximum entropy approach to species distribution modeling. Proceedings of the Twenty-First International Conference on Machine Learning, 655-662.

Phillips, S. J., R. P. Anderson, and R. E. Schapire. 2006. Maximum entropy modeling of species geographic distributions. Ecological Modelling 190, 231-259.

Rittenhouse, C D, W. D. Dijak, F. R. Thompson III, and J.J. Millspaugh. 2007. Development of landscape-level habitat suitability models for ten wildlife species in the central hardwoods region. General Technical Report NRS-4. U.S. Department of Agriculture, Forest Service, Northern Research Station, Newtown Square, Pennsylvania. 47 pp

Rommé, R.C, K. Tyrell, V. Brack, Jr. 1995. Literature summary and habitat suitability index model: components of summer habitat for the Indiana bat, Myotis sodalis. 3D/Environmental, Cincinnati, OH.

68


Sparks, D. W., C. M. Ritzi, J. E. Duchamp, and J. O. Whitaker, Jr. 2005. Foraging habitat of the Indiana bat (Myotis sodalis) at an urban–rural interface. Journal of Mammalogy, 86(4).

Sparks, D. W., J. O. Whitaker, Jr., and C. M. Ritzi. 2004. Foraging ecology of the endangered Indiana bat. in Indiana Bat and Coal Mining: A Technical Interactive Forum (K.C. Vories and A. Harrington, eds.). U.S. Department of the Interior, Office of Surface Mining. Alton, Illinois.

Thogmartin, W. E., R. A. King, P. C. Mckann, J. A. Szymanski, and L. Pruitt. 2012. Population-level impact of white-nose syndrome on the endangered Indiana bat. Journal of Mammalogy 93, 1086–1098.

Theobald, D.M., J.B. Norman, M.R. Sherburne. 2006. FunConn v1 User’s Manual: ArcGIS tools

for Functional Connectivity Modeling. Natural Resource Ecology Lab, Colorado State University, Fort Collins, CO.

U.S. Fish and Wildlife Service (USFWS). 2007. Indiana Bat Draft Recovery Plan: First Revision. U.S. Fish and Wildlife Service, Fort Snelling, MN.

Watrous, K. S., T. M. Donovan, R. M. Mickey, S. R. Darling, A. Hicks, and S. L. Von Oettingen. 2006. Predicting minimum habitat characteristics for the Indiana bat in the Champlain Valley. Journal of Wildlife Management 70, 1228-1237.

Weber, T. C., and W. L. Allen. 2010. Beyond on-site mitigation: An integrated, multi-scale approach to environmental mitigation and stewardship for transportation projects. Landscape and Urban Planning 96, 240-256.

Weber, T. C. and D. E. Boss. 2009. Use of LiDAR and supplemental data to estimate forest maturity in Charles County, Maryland, USA. Forest Ecology and Management 258, 2068-2075.

Wilson VanVoorhis, C. R., and B. L. Morgan. 2007. Understanding power and rules of thumb for determining sample sizes. Tutorials in Quantitative Methods for Psychology 3, 43-50.

Wisz, M. S., R. J. Hijmans, J. Li, A. T. Peterson, C. H. Graham, A. Guisan, and NCEAS Predicting Species Distributions Working Group. 2008. Effects of sample size on the performance of species distribution models. Diversity and Distributions 14, 763-773.

summer habitat identification of an endangered bat ... · 55 weber, sparks / journal of...

Documents