Landscape-level models of potential habitat for Sonoran pronghorn

24 SPECIAL COVERAGE Landscape-level models of potential habitat for Sonoran pronghorn by Chantal S. O'Brien, Steven S. Rosenstock, John J. Hervert, Jill L. Bright, and Susan R. Boe Abstract Key Words A population of endangered Sonoran pronghorn (Antilocapra americana sonoriensis) exists in the United States, and 2 populations exist in Mexico. Because of the vulnerability of small, remnant populations of this subspecies to stochastic events, an important aspect of recovery planning is identifying suitable areas for establishment of new populations. To support translocation efforts, we developed landscapelevel Classification and Regression Tree (CART) and logistic regression models of potential Sonoran pronghorn habitat in southwestern Arizona through a 2-part modeling process. First, we used approximately half of Sonoran pronghorn locations (total n=3,219, collected from 1994 to 2002 from radiocollared animals in the United States) and unused points (total n=3,142, randomly generated in areas within the range of Sonoran pronghorn below 21% slope, but >1.6 km from pronghorn locations) to create habitat models from 5 explanatory variables (i.e., slope, aspect, biome, distance to wash, and soil category). We validated models with the second half of pronghorn and unused points. Both models determined whether areas would or would not be used by Sonoran pronghorn based upon values of explanatory variables at Sonoran pronghorn locations and unused points. The CART model correctly identified 63% of pronghorn locations and 65% of unused points. The logistic regression model correctly identified 57% of pronghorn locations and 62% of unused points. Second, we created a predictive Geographic Information System (GIS) map of Sonoran pronghorn habitat and applied it to the evaluation area. Both models identified >12,000 km 2 of potential habitat for Sonoran pronghorn on the evaluation area. Our models are a first step toward identifying potential translocation sites for Sonoran pronghorn. Potential translocation sites should be further evaluated with respect to habitat factors not included in our models, including barriers to pronghorn movements, water supplies, and forage resources. Antilocapra americana sonoriensis,arizona, CART, Classification and Regression Tree, habitat model, logistic regression, Sonoran pronghorn The Sonoran pronghorn (Antilocapra americana sonoriensis) is 1 of 5 subspecies of pronghorn, 4 of which occur in the United States. Three subspecies of pronghorn live in Arizona: Sonoran, Mexican (A. a. mexicana), and American (A. a. americana). Historically in Arizona, Sonoran pronghorn were reported south of Interstate 10 Address for Chantal S. O'Brien, Steven S. Rosenstock, and Susan R. Boe: Arizona Game and Fish Department, Research Branch - WMRS, 2221 W. Greenway Rd., Phoenix, AZ 85023-4312, USA e-mail for O'Brien: cobrien@azgfd.gov. Address for John J. Hervert and Jill L. Bright: Arizona Game and Fish Department, 9140 E. 28th St., Yuma, AZ 85363, USA. Wildlife Society Bulletin 2005, 33(1):24 34 Peer refereed

Sonoran pronghorn habitat models O'Brien et al. 25 (I-10) and west of Highway 19 (Figure 1; Wright and devos 1986). Current Sonoran pronghorn range in the United States is bordered by Highway 85 to the east, Cabeza Prieta Mountains to the west, Interstate 8 (I-8) to the north, and the United States Mexico International Border to the south (Figure 1). The total area of currently occupied range (defined as that used within the last 10 years) is 6,498 km 2 (Bright et al. 2001), and the current population numbers approximately 21 animals (95% CI 18 33; Bright and Hervert 2003). A sharp decline occurred in 2002, prior to which the already declining population consisted of 99 animals (95% CI 69 392; Bright et al. 2001). The taxonomy of pronghorn subspecies, including Sonoran, has been a source of debate (Paradiso and Nowak 1971, Cockrum 1981, United States Fish and Wildlife Service [USFWS] 1998, Malone et al. 2002). However, regardless of subspecies status, Sonoran pronghorn will be protected under the Isolated Vertebrate Population Policy in the Endangered Species Act (USFWS 1998). The USFWS listed the Sonoran pronghorn as endangered in 1967 (United States Department of the Interior, Office of the Secretary 1967). The Sonoran Pronghorn Recovery Plan was adopted by the USFWS in 1982 (USFWS 1982), revised in 1994 (USFWS 1994), and a Final Revised Sonoran Pronghorn Recovery Plan was issued in 1998 (USFWS 1998). Recovery criteria in the final plan identify the need to establish a second population of Sonoran pronghorn in the United States and for the current population to number >300 individuals, or for the current population to reach a size that will promote a stable population (USFWS 1998). Areas within historical range should be evaluated for translocation based on habitat use preferences learned from collared Sonoran pronghorn (USFWS 1998:41). Pronghorn habitat has been evaluated in Arizona (Ockenfels et al. 1996), but that evaluation ignored Sonoran pronghorn. One previous study modeled habitat use by Sonoran pronghorn in Organ Pipe Cactus National Monument (OPCNM; Marsh et al. 1999, Wallace et al. 2002). That model was developed for the small fraction of Sonoran pronghorn range within OPCNM, used specific Geographic Information System (GIS) vegetation coverages available only for that area, and had limited applicability elsewhere. Several studies have described habitat selection by Sonoran pronghorn on current range (devos 1990, 1998; Hughes 1991), but have not evaluated large, currently unoccupied areas as potential habitat for Sonoran pronghorn. Quantitative assessments of potentially suitable habitat could help guide future translocations and contribute to recovery of Sonoran pronghorn. Our objectives were to create landscape-level predictive models of habitat use by Sonoran pronghorn specific to the Sonoran Desertscrub biome (Turner and Brown 1994) in Arizona and to use those models to evaluate potential habitat in southwestern Arizona. We limited our models to the United States population of Sonoran For species such as Sonoran pronghorn, we believe that landscape-level models are an efficient and objective approach for preliminary evaluation of potential translocation sites. Once identified by the models, candidate sites can be further evaluated using groundbased assessments. pronghorn because of the limited information and locations available for Sonoran pronghorn and occupied habitats in Mexico and because Sonoran pronghorn in Mexico occur in different biomes than those occupied in the United States. Study area We created our models from existing locations of Sonoran pronghorn in the United States. The range of Sonoran pronghorn included portions of Barry M. Goldwater Air Force Range (BMGR), Cabeza Prieta National Wildlife Refuge (CPNWR; USFWS), and OPCNM (National Park Service). Average minimum temperature (Ajo station; Western Regional Climate Center 1973 2002) ranged from 7.5 o C (January) to 26.2 o C (July). Average maximum temperature ranged from 17.4 o C in January to 38.6 o C in July. Average annual precipitation was 19.9 cm/year (range=9.9 36.5). We evaluated the southwest corner of Arizona as potential habitat, including Yuma Proving Ground (YPG; United States Army), Kofa National Wildlife Refuge (KNWR; USFWS), BMGR, CPNWR, OPCNM lands west of Highway 85, and lands managed by Bureau of Land Management (Figure 1). The evaluation area included most of the historical range of Sonoran pronghorn in Arizona (Wright and devos 1986) and all of the current range and was bounded to the north by I-10, to the east by Highway 85, to the south by United States Mexico International Border, and to the west by the Colorado River (Figure 1). The current range and evaluation area were within the

26 Wildlife Society Bulletin 2005, 33(1):24 34 Figure 1. Map of historical and current range of Sonoran pronghorn and evaluation area in southwestern Arizona, USA. Historical range is from Wright and devos (1986), and current range is 100% MCP from radiocollared Sonoran pronghorn 1994 2002. The following public lands are identified: OPCNM = Organ Pipe Cactus National Monument, CPNWR = Cabeza Prieta National Wildlife Refuge, BMGR = Barry M. Goldwater Range, YPG = Yuma Proving Ground, and KNWR = Kofa National Wildlife Refuge. Sonoran Desert section of the Basin and Range physiographic province, an area characterized by block-faulted mountains and wide alluvial valleys (Chronic 1983). Biomes in the current range and the evaluation area were the Lower Colorado River Valley (LCRV) and Arizona Upland subdivisions of Sonoran Desertscrub (Turner and Brown 1994). The LCRV subdivision occurred through the wide, alluvial valleys. Dominant plant species in LCRV included creosote bush (Larrea tridentata), white bursage (Ambrosia dumosa), palo verde (Parkinsonia spp.), ironwood (Olneya tesota), honey mesquite (Prosopis glandulosa), smoketree (Psorothamnus spinosus), saguaro (Carnegiea gigantea), and ocotillo (Fouquieria splendens; Turner and Brown 1994). The Arizona Upland subdivision occurred in mountains and on upper bajadas. Greater precipitation in these areas supported plant growth that was more complex and dense than that found in the LCRV (Turner and Brown 1994). Common tree species in the Arizona Upland subdivision included palo verde, ironwood, and catclaw acacia (Acacia greggii). Shrub species included bursage (Ambrosia spp.) and creosote bush (Turner and Brown 1994). Numerous species of cacti were present, including buckhorn cholla (Opuntia acanthocarpa), chain fruit cholla (O. fulgida), teddy bear cholla (O. bigelovii), saguaro, and barrel cactus (Ferocactus spp.; Turner and Brown 1994). Methods Sonoran pronghorn locations Arizona Game and Fish Department (AGFD) captured and radiocollared a total of 35 pronghorn (28 F, 7 M) through separate efforts in 1994, 1997, 1998, 1999, and 2000 (Hervert et al. 2000, Bright and Hervert 2002) with a net-gun fired from a helicopter (Krausman et al. 1985). Collared pronghorn were located from fixedwing aircraft approximately weekly from 1994 2002, resulting in 3,718 locations. After a collared pronghorn was located, the observer attempted to get a visual sighting of the animal. If successful, the observer took a Global Positioning System (GPS) location with a handheld GPS unit (Garmin Corporation, Olathe, Kans.). The pilot avoided flying directly overhead to minimize disturbance of pronghorn, so GPS locations were offset by <0.4 km from actual locations. If a radiocollared pronghorn could not be located visually, the observer took a GPS location where they believed the animal was located based upon signal strength. Observers also collected GPS locations for uncollared animals encountered during radiotracking flights.

Sonoran pronghorn habitat models O'Brien et al. 27 Model development, validation, and application Due to the size of the current range and the evaluation area, we created a landscape-level model. Both modeling approaches we chose required locations used and unused by Sonoran pronghorn. Our modeling followed a 2-step approach: model creation and application. First, we delineated used and unused areas from pre-existing pronghorn locations, applied values of explanatory variables from statewide GIS coverages to used and unused points, ran logistic regression and Classification and Regression Tree (CART) analyses to model used and unused areas, and validated models with additional used and unused points. Second, we mapped results to predict potential habitat on the evaluation area. Delineation of used areas. Our data set consisted of 3,718 locations of Sonoran pronghorn, including locations where animals were captured and mortalities were recovered. We treated observations of multiple animals together (i.e., interacting) as a single location. We used all locations to create a 100% minimum convex polygon (MCP), which we considered to be Sonoran pronghornoccupied range in the United States. We considered all areas within occupied range as available to pronghorn, lacking mountains, deep-cut valleys, and other geographic features that provide poor escape terrain (Ockenfels et al. 1996). We used locations from 31 of 35 radiocollared animals (7 M, 24 F) for model development, removing 3 animals that died shortly after capture and were not relocated while alive and 1 whose radiocollar failed. To maintain consistency between data sets, we did not include the 10% of Sonoran pronghorn locations from areas with >20% slope. We included mortality and capture locations in the data set used to delineate Sonoran pronghorn occupied range and use areas but did not include them in model development because the ultimate location of an animal, radiocollar, or carcass could have been altered by pursuit during capture, attempted escape during predation, and subsequent scavenging, and thus may not have reflected habitat use. We used 3,219 pronghorn locations for model development, after removing capture and mortality locations. Delineation of unused areas. To account for offset of GPS-recorded positions, we used ARC/INFO (Experimental Systems Research Institute, Redlands, Calif.) to buffer all pronghorn locations. Wallace et al. (2002) buffered pronghorn locations by 5 km; however, we used a more conservative 1.6-km buffer to ensure that the used area included all actual locations but limited inclusion of areas potentially unused by pronghorn. We assumed that all areas with <20% slope and within the 100% MCP but outside buffered used points are unused by Sonoran pronghorn. We generated 3,142 random points within the unused areas for model development. Explanatory variable identification and assignment. We used available GIS coverages to determine values of 5 explanatory variables (i.e., aspect, biome, distance to wash, slope, and soil association) at pronghorn locations and unused points (Table 1). We considered aspect to be biologically meaningful because of rainshadow effects (Ingram 2000) and aspect-specific frost deposition that can influence vegetation growth (Turner and Brown 1994). We used biomes because they were the best available coverage of plant communities on the study area and reflected effects of dominant climatic variables (i.e., temperature and precipitation; Lowe and Brown 1994). We used distance to ephemeral washes as an explanatory variable because Sonoran pronghorn select areas close to washes (devos 1998) and commonly use thermal cover and forage found in washes, especially during dry summers (Phelps 1981, Wright and devos 1986, Hervert et al. 2000). We included slope as an explanatory variable because there might be differential use of slopes <20%. Sonoran pronghorn use bajadas and mountain slopes during autumn (Phelps 1981), and use of bajadas is greater than expected during all seasons (Hervert et al. 2000). We included soil association in our variable set because soil attributes that affect moisture retention and vegetation growth (McAuliffe 2000) could influence habitat use by pronghorn. The evaluation area contained soil associations absent on current pronghorn range; thus, we used the soil categorization approach of Wallace et al. (2002). We used chi-square goodness-of-fit tests (Neu et al. 1974, Byers et al. 1984) to compare observed versus expected use of 11 soil associations on current range in Table 1. List of GIS coverages used in development of potential habitat models for Sonoran pronghorn, southwestern Arizona, USA. GIS coverage Source, scale Derived attribute(s) 30-m Digital elevation map United States Geological Survey Slope (%) and aspect (degrees) 1:100,000 STATSGO state soil geographic USDA Natural Resources General soil association, identified database Conservation Service, 1:250,000 by map unit ID (muid) Ephemeral and perennial Arizona State Land Department, Distance to wash streams in Arizona 1:100,000 Biotic communities of the Arizona State Land Department Biome Southwest from Brown et al. 1980, 1:100,000

28 Wildlife Society Bulletin 2005, 33(1):24 34 areas with slope <20%. We compared proportions of pronghorn observations among soil associations with percent of total area of each soil association within the Sonoran pronghorn population 100% MCP (calculated with ARC/INFO). We calculated simultaneous 95% Bonferroni confidence intervals to infer selection (i.e., use > availability) or avoidance (i.e., use<availability) when we found a statistically significant difference (α= 0.05) between expected and observed use of individual soil associations. Based upon results of the chi-square goodness-of-fit tests and Bonferroni confidence intervals, we placed the 11 soil associations into preferred, avoided, or neutral categories. We classified soils on the evaluation area as neutral if they were not among the original 11 encountered on current range. We replaced soil association with soil category as a variable for model development. Modeling. We used Microsoft Excel (Microsoft Inc., Redmond, Wash.) to assign random binary numbers to pronghorn locations and unused points. We used these random numbers to split pronghorn locations and unused points into approximately equal groups, creating learning and test sets for model development. We used the Statistica 6.1 (StatSoft Inc., Tulsa, Okla.) implementation of CART (Breiman et al. 1984) and logistic regression analyses. We used CART because it is relatively free of assumptions, has been used successfully in habitat modeling (Andersen et al. 2000, De ath and Fabricius 2000, Debeljak et al. 2001, McGrath et al. 2003), and produces decision trees that are readily applied in a management context. The CART models consist of a decision tree with binary (i.e., yes no) splits based upon specific values of predictor variables. Decision pathways originate from a starting node that contains all observations and end at multiple terminal nodes containing unique subsets of observations. Terminal nodes are assigned a final outcome based on group membership of the majority of observations (i.e., either used or unused ). Our analysis used the Gini goodness-of-fit measure, estimated prior probabilities of group membership from proportions in the learning data set, and specified equal misclassification costs for used and unused predictions. We used P- value=0.05 for selection of the variables used to create binary splits. We pruned candidate trees on misclassification costs, specified splitting end if terminal nodes contain <15 observations (0.5% of learning sample), and selected the final tree using the 1-SE rule (Breiman et al. 1984). We used logistic regression as an adjunct to CART because it is widely used in habitat models (O Connor 2002) and produces a continuous probability rather than a binary outcome. We created logistic regression models by the best subsets method (Hosmer and Lemeshow 1989) and chose the model with the lowest Akaike s Information Criterion (AIC; Burnham and Anderson 2002). We calculated AIC differences ( i ) and weights (w i ) to determine similarity between the top 5 models (Burnham and Anderson 2002). We used outputs from the CART and logistic regression models to create GIS maps of habitat suitability (mapping unit=8,100 m 2 ). The CART model map was delineated into habitat and nonhabitat. The logistic regression model map divided probabilities of habitat use by Sonoran pronghorn into classes of: 0.00 0.25, 0.26 0.50, 0.51 0.75, and 0.76 1.00. Results Sonoran pronghorn did not use soil associations in proportion to availability (χ 2 =218.52, P<0.0001). One 10 soil association was preferred, 4 were avoided, and 6 were used as expected (i.e., neutral; Table 2). We grouped soils into 3 soil categories based upon the 3 levels of preference and used these categories in subsequent modeling. The final CART model had 10 splits and 11 terminal nodes, 7 of which described areas used by pronghorn (Figure 2). The CART model used all 5 explanatory variables. Distance to wash was the most important explanatory variable for used areas, followed by aspect, soil category, % slope, and biome. The CART model had an overall classification accuracy of 64%, with classification accuracy of 63% and 65% for used and unused points, respectively. Cross-validation with test data yielded overall correct classification of 62%, with classi- Table 2. Bonferroni confidence intervals for use versus availability of soil associations by Sonoran pronghorn (n = 3,219 locations) within the population 100% MCP in southwestern Arizona, USA 1994 2002. Soil % Use association available % use 95% CI category a AZ008 1.82 0.37 0.000667 0.006789 2 AZ009 34.65 33.30 0.309348 0.356697 3 AZ010 27.25 22.71 0.206044 0.248134 2 AZ016 20.76 29.23 0.269479 0.315174 1 AZ017 3.03 1.78 0.010324 0.023227 2 AZ021 4.79 5.28 0.041577 0.064046 3 AZ022 0.12 0.28 0.000144 0.005448 3 AZ023 0.25 0.37 0.000667 0.006789 3 AZ028 4.15 3.23 0.023426 0.041190 2 AZ030 2.21 2.87 0.020210 0.036950 3 AZ050 0.94 0.68 0.002696 0.010973 3 a 1 = use > availability, 2 = use < availability, and 3 = neutral use.

Sonoran pronghorn habitat models O'Brien et al. 29 Figure 2. The CART model for Sonoran pronghorn habitat use in southwestern Arizona, USA. Decision rules at splits apply to the left branch, while the opposite rule applies to the right branch. Numbers (1 and 0) inside nodes indicate majority classification of each node: 1 = pronghorn locations and 0 = unused locations. fication accuracy of 61% for used locations and 63% for unused points. The CART model identified 13,370 km 2 of potentially suitable habitat, with 2,609 km 2 on current range (Figure 3). The CART model identified 10 areas of potential habitat (La Posa Plain, Castle Dome Plain, King Valley, Palomas Plain, Hyder Valley, Yuma Desert, Lechuguilla Desert, Tule Desert, Childs Valley, and La Abra Plain). The top 5 models from best subsets logistic regression each contained 3 5 variables (Table 3). All of the top 5 models were good approximations to the data ( I <4; Table 4). We chose a model with 4 variables and 1 interaction term. Distance to wash, soil category 2, and interaction of distance to wash and aspect negatively influenced use by pronghorn. Aspect and soil category 1 positively influenced use by pronghorn. The low Akaike weight of the chosen model (Table 4) suggests that it may not be the best model if additional replicate samples were available. Using 0.50 as the cutpoint between used and unused points, the logistic regression equation correctly classified 59% of total locations, with classification accuracy of 57% and 62% for used and unused points, respectively. Cross-validation produced overall classification accuracy of 60%, with classification accuracy of 58% and 62% for used and unused points, respectively. The logistic regression model identified 12,489 km 2 of potentially suitable habitat with 2,621 km 2 on current range (Figure 4). Current range contained 101 km 2 with 0 0.25 probability of use by pronghorn, 3,612 km 2 with 0.26 0.50 probability of use by pronghorn, 2,605 km 2 with 0.51 0.75 probability of use by pronghorn, and 16 km 2 with 0.76 1.00 probability of use by pronghorn. The evaluation area contained 516 km 2 with 0 0.25 probability of use by pronghorn, 11,766 km 2 with 0.26 0.50 probability of use by pronghorn, 12,460 km 2 with 0.51 0.75 probability of use Table 3. Top 5 models from best subsets logistic regression predicting Sonoran pronghorn presence and absence on current range in southwestern Arizona, USA. Model 1 is the model we chose for evaluation of potential habitat. Distance Aspect*Distance Model Intercept Aspect Biome a to wash Soil 1 b Soil 2 c Slope to wash Slope*Aspect 1 0.097393 0.002440 0.000046 0.411638 0.447351 0.000002 2 0.231048 0.002594 0.179887 0.000055 0.476258 0.480657 0.000002 3 0.005422 0.001580 0.000204 0.428344 0.450081 0.013385 0.000068 4 0.124174 0.002639 0.000109 0.000002 5 0.159564 0.001679 0.189534 0.000215 0.492132 0.479856 0.023924 0.000081 a Included if biome = LCRV. b Included if soil category = 1. c Included if soil category = 2.

30 Wildlife Society Bulletin 2005, 33(1):24 34 Figure 3. Map of potential Sonoran pronghorn habitat in southwestern Arizona, USA, identified by CART model. Current range is delineated by the 100% MCP. by pronghorn, and 27 km 2 with 0.76 1.00 probability of use by pronghorn. The model selected La Posa Plain, Castle Dome Plain, Palomas Plain, Hyder Valley, Table 4. AIC differences ( i ) and weights (w i ) for the top 5 models from best subsets logistic regression predicting Sonoran pronghorn use of current range in southwestern Arizona, USA. Model 1 is the model we chose for evaluation of potential habitat. Model i w i 1 0.0 0.48 2 1.9 0.18 3 1.9 0.18 4 3.5 0.08 5 3.9 0.07 Lechuguilla Desert, Childs Valley, King Valley, and La Abra Plain as potential habitat. Discussion Habitat model The models we created do not imply causal inference between the explanatory variables and Sonoran pronghorn habitat use. It is possible that some explanatory variables we used were strongly correlated with other, unmeasured variables that actually affected pronghorn habitat use. The logistic regression required fewer variables than the CART model and had the advantage of providing probability of pronghorn use, allowing managers to decide the probability level of use that corresponds to adequate habitat. However, the CART model performed better than the logistic regression at identifying used and unused points and provided a simple dichotomous key that made it intuitive to use. The CART model performed better than the logistic regression model, identifying used and unused areas with higher accuracy. The logistic regression model classified as habitat a slightly larger area on current range, which would explain why more errors of commission with the unused points occurred. However, we would expect that a more inclusive model would likewise have an increased correct classification rate (CCR) for used locations. Despite the tendency of the logistic regression model to include a larger area of current range as habitat, it had a lower CCR for used locations than the CART model. The CART model was more conservative on current range, identifying less habitat than the logistic regression model, but was more successful at identifying used and unused areas on current range and identified more habitat overall in the evaluation area. The CART and logistic regression models

Sonoran pronghorn habitat models O'Brien et al. 31 Figure 4. Map of potential Sonoran pronghorn habitat in southwestern Arizona, USA, identified by logistic regression model. Probability of pronghorn use in areas with slope <20% is broken into 4 categories: 0 0.25, 0.26 0.45, 0.46 0.75, and 0.76 1.00. Current range is delineated by the 100% MCP. The areas that fell within the probability range of 0.76 1.00 do not appear on the map due to the size of the mapping unit (8,100 m 2 ). identified La Posa Plain, King Valley, Castle Dome Plain, Palomas Plain, Hyder Valley, Lechuguilla Desert, Childs Valley, and La Abra Plain as potential habitat. In addition, the CART model also classified Yuma and Tule Deserts as potential habitat. Model improvement If Sonoran pronghorn are translocated, habitat-use patterns in newly occupied areas could be used to validate and further refine our initial models. Our modeling effort concentrated on year-round habitat use by Sonoran pronghorn. Previous studies have identified annual and seasonal differences in habitat use by pronghorn (Phelps 1981, Hughes 1991, Hervert et al. 2000, Wallace et al. 2002). Models of seasonal habitat use perform better on OPCNM than year-round models (Wallace et al. 2002). Future range-wide modeling efforts could be enhanced by focusing on particular seasons (e.g., dry vs. wet), critical periods of pronghorn life history (e.g., fawning), or habitat use at the individual animal level. Variability in pronghorn habitat use not explained by our models may reflect the relatively coarse resolution of available GIS coverages, potential misidentification of used areas as unused areas, and influences of habitat variables not included in our models. Forage availability and vegetation association are important predictors of habitat use by Sonoran (Wright and devos 1986, Hughes 1991, devos 1998) and other subspecies of pronghorn (Irwin and Cook 1985, Ockenfels et al. 1994). Future models could be improved by incorporation of finer-scale vegetation characteristics, some of which may be obtained from an intensive vegetation mapping project underway on current range (J. Malusa, University of Arizona, personal communication). Management implications Managers of pronghorn have created several guides for rating potential translocation areas for other subspecies of pronghorn (Hoover et al. 1959, Yoakum 1980, International Union for Conservation of Nature and

32 Wildlife Society Bulletin 2005, 33(1):24 34 Natural Resources 1987). However, these approaches rely on subjective ground-based assessments and do not directly relate habitat use on current range to potential translocation sites. For species such as Sonoran pronghorn, we believe that landscape-level models are an efficient and objective approach for preliminary evaluation of potential translocation sites. Once identified by the models, candidate sites can be further evaluated using ground-based assessments. Our models are an initial step toward identifying potential sites for Sonoran pronghorn translocation. Several additional steps for evaluating translocation sites include public input and reviewing predator presence and density, fencing, and presence of preferred forage and water (USFWS 1998). Other suggestions for selecting translocation sites include determining whether land managers will support pronghorn recovery and ensuring that factors that led to initial reduction in population size on current range are not present in the translocation area (O Gara and Yoakum 1990). Our models classify habitat or probability of use by Sonoran pronghorn but do not address quality of potential habitat. We identified several other influences that could affect habitat quality, but we were unable to include these in our models. These include levels of disturbance and development; barriers presented by roads, canals, highways, and railroads; availability of water; and forage availability and quality. Highways and other transportation corridors located on or adjacent to current range may prevent access to historically used areas, but have not been considered a serious limiting factor for the current population because there are large, continuous tracts of land available without roads (Hervert et al. 2000). Sonoran pronghorn avoid roads, selecting areas >5 km away (devos 1998). Paved roads and fences are important barriers to movement of pronghorn elsewhere in Arizona (Ockenfels et al. 1994, Bright 1999). Highways and other paved roads may prevent pronghorn movement outside the area we evaluated as potential habitat. Interstates 10 and 8 (Figure 1) are large, multi-lane, busy highways that have additional barriers associated with them, including fences, canals, and railroad tracks. Highway 95 bisects 2 areas of potentially suitable habitat identified by our models (La Posa and Castle Dome Plains, Figure 1). Water use by Sonoran pronghorn has been questioned (Monson 1968), but recent studies have documented extensive use of freestanding water during summer months (Hervert et al. 1995, 2000; USFWS 1998). We did not include distance to water in our model because of uncertainties concerning availability of specific sources of water to pronghorn on current range. There are wildlife water developments and natural waterholes on current range, but some are not accessible to or used by Sonoran pronghorn because they are fenced, surrounded by dense vegetation (e.g., earthen tanks; Cutler 1996, Hervert et al. 2000), located outside of areas used during summer (Cutler et al. 1996), or in mountainous terrain (Carr 1981). Pronghorn also obtain water from ephemeral pools that occur on disturbed areas on BMGR (Hervert et al. 2000). Sonoran pronghorn make heavy use of chain fruit cholla, which appears to be an important food and water source during drought (Phelps 1981, Hervert et al. 2000). Chain fruit cholla and year-round water sources accessible to pronghorn are rare in the evaluation area. Thus, availability of water will be an important consideration in future evaluation of potential translocation sites. Acknowledgments. R. Ockenfels shared his knowledge of pronghorn literature. Members of the Sonoran Pronghorn Recovery Team and personnel from the Bureau of Land Management, Yuma Proving Ground (YPG), Kofa National Wildlife Refuge, Cabeza Prieta National Wildlife Refuge, and Organ Pipe Cactus National Wildlife Refuge offered insights and suggestions that helped refine our modeling effort. Constructive reviews of the manuscript were provided by 2 anonymous reviewers. The United States Army, YPG Conservation Program funded our model-based assessment of the evaluation area. Pronghorn locations and other data used in our models were obtained from efforts funded by the United States Marine Corps, United States Fish and Wildlife Service, United States Air Force, National Park Service, and Arizona Game and Fish Department. Literature cited ANDERSEN, M. C., J. M. WATTS, J. E. FREILICH, S. R. YOOL, G. I. WAKEFIELD, J. F. MCCAULEY, AND P. B. FAHNESTOCK. 2000. Regression-tree modeling of desert tortoise habitat in the central Mojave Desert. Ecological Applications 10: 890 900. BREIMAN, L., J. H. FRIEDMAN, R. A. OLSHEN, AND C. J. STONE. 1984. Classification and regression trees. Chapman and Hall, New York, New York, USA. BRIGHT, J. L. 1999. Home ranges, habitat use and distribution around water sources of pronghorn in northern Arizona. Thesis, Northern Arizona University, Flagstaff, USA. BRIGHT, J. L., AND J. J. HERVERT. 2002. Sonoran pronghorn population monitoring progress report: 1999 2001. Arizona Game and Fish Department, Nongame and Endangered Wildlife Program technical report 209, Phoenix, USA. BRIGHT, J. L., AND J. J. HERVERT. 2003. Sonoran pronghorn 2002 aerial survey summary. Arizona Game and Fish Department, Nongame and Endangered Wildlife Program technical report 236, Phoenix, USA. BRIGHT, J. L., J. J. HERVERT, AND M. T. BROWN. 2001. Sonoran pronghorn

Sonoran pronghorn habitat models O'Brien et al. 33 2000 aerial survey summary. Arizona Game and Fish Department, Nongame and Endangered Wildlife Program technical report 180, Phoenix, USA. BROWN, D. E., C. H. LOWE, AND C. P. PASE. 1980. A digitized systematic classification for ecosystems with an illustrated summary of the natural vegetation of North America. United States Department of Agriculture Forest Service General Technical Report RM-73. Rocky Mountain Forest and Range Experiment Station, Fort Collins, Colorado, USA. BURNHAM, K. P., AND D. R. ANDERSON. 2002. Model selection and multimodel inference: a practical information-theoretic approach. Second edition. Springer-Verlag, New York, New York, USA. BYERS, C. R., R. K. STEINHORST, AND P. R. KRAUSMAN. 1984. Clarification of a technique for analysis of utilization-availability data. Journal of Wildlife Management 48: 1050 1053. CARR, J. N. 1981. Habitat of the Sonoran pronghorn. Pages 11 19 in The Sonoran pronghorn. Special report number 10. Arizona Game and Fish Department, Phoenix, USA. CHRONIC, H. 1983. Roadside geology of Arizona. Mountain Press Publishing, Missoula, Montana, USA. COCKRUM, E. L. 1981. Taxonomy of the Sonoran pronghorn. Pages 2 10 in The Sonoran pronghorn. Special report number 10. Arizona Game and Fish Department, Phoenix, USA. CUTLER, P. L. 1996. Wildlife use of two artificial water developments on the Cabeza Prieta National Wildlife Refuge, southwestern Arizona. Thesis, University of Arizona, Tucson, USA. CUTLER, T. L., M. L. MORRISON, AND D. J. GRIFFIN. 1996. Final report: wildlife use of Jose Juan and Red Tail Tank, Cabeza Prieta National Wildlife Refuge, southwestern Arizona. United States Department of Defense Contract N68711 93 LT 3026. University of Arizona, Tucson, USA. DE ATH, G., AND K. E. FABRICIUS. 2000. Classification and regression trees: a powerful yet simple technique for ecological data analysis. Ecology 81:3178 3192. DEBELJAK, M., S. DZEROSKI, K. JERINA, A. KOBLER, AND M. ADAMIC. 2001. Habitat suitability modelling for red deer (Cervus elaphus L.) in south-central Slovenia with classification trees. Ecological Modelling 138: 321 330. DEVOS, J. C., JR. 1990. Selected aspects of Sonoran pronghorn research in Arizona and Mexico. Pages 46 52 in P. R. Krausman and N. S. Smith, editors. Managing wildlife in the Southwest: proceedings of the symposium 16 18 October, 1990. Arizona Chapter of the Wildlife Society, Phoenix, USA. DEVOS, J. C., JR. 1998. Habitat selection patterns by Sonoran pronghorn in southwest Arizona. Pages 34 37 in T. J. Tibbitts and G. J. Maender, editors. First conference on research and resource management in southern Arizona national park areas: extended abstracts. May 5 7, 1998. Organ Pipe Cactus National Monument and Cooperative Park Studies Unit, University of Arizona, Tucson, USA. HERVERT, J. J., J. L. BRIGHT, M. T. BROWN, L. A. PIEST, AND R. S. HENRY. 2000. Sonoran pronghorn population monitoring:1994 1998. Arizona Game and Fish Department, Nongame and Endangered Wildlife Program technical report 162, Phoenix, USA. HERVERT, J. M., B. HENRY, M. BROWN, D. W. BELITSKY, AND M. E. KREIGHBAUM. 1995. Sonoran pronghorn population monitoring: progress report. Arizona Game and Fish Department, Nongame and Endangered Wildlife Program technical report 98, Phoenix, USA. HOOVER, R. L., C. E. TILL, AND S. OGILVIE. 1959. The antelope of Colorado: a research and management study. Colorado Department of Fish and Game technical bulletin 4, Fort Collins, USA. HOSMER, D. W., AND S. LEMESHOW. 1989. Applied logistic regression. John Wiley & Sons, New York, New York, USA. HUGHES, K. S. 1991. Sonoran pronghorn use of habitat in southwest Arizona. Thesis, University of Arizona, Tucson, USA. INGRAM, M. 2000. Desert storms. Pages 41 50 in S. J. Phillips and P. Wentworth Comus, editors. A natural history of the Sonoran Desert. Arizona-Sonora Desert Museum, Tucson, USA. INTERNATIONAL UNION FOR CONSERVATION OF NATURE AND NATURAL RESOURCES. 1987. The IUCN position statement on translocation of living organisms: introductions, reintroductions, and restocking. International Union for Conservation of Nature and Natural Resources, Species Survival Commission, Gland, Switzerland. IRWIN, L. L., AND J. G. COOK. 1985. Determining appropriate variables for a habitat suitability model for pronghorns. Wildlife Society Bulletin 13: 434 440. KRAUSMAN, P. R., J. J. HERVERT, AND L. L. ORDWAY. 1985. Capturing deer and mountain sheep with a net-gun. Wildlife Society Bulletin 13: 71 73. LOWE, C. H., AND D. E. BROWN. 1994. Introduction. Pages 8 16 in D. E. Brown, editor. Biotic communities: southwestern United States and northwestern Mexico. University of Utah Press, Salt Lake City, USA. MALONE, C. L., J. C. DEVOS, JR., J. R. HEFFELFINGER, AND O. E. RHODES, JR. 2002. Genetic distinction of the Sonoran pronghorn antelope. Pronghorn Antelope Workshop Proceedings 20: 15 (abstract). MARSH, S. E., C. WALLACE, AND J. WALKER. 1999. Evaluation of satellite remote sensing methods for regional inventory and mapping of desert resources. Final Report. Arizona Remote Sensing Center, Office of Arid Lands Studies, University of Arizona, Tucson, USA. MCAULIFFE, J. R. 2000. Desert soils. Pages 87 104 in S. J. Phillips and P. Wentworth Comus, editors. A natural history of the Sonoran Desert. Arizona-Sonora Desert Museum, Tucson, USA. MCGRATH, M. T., S. DESTEPHANO, R. A. RIGGS, L. L. IRWIN, AND G. J. ROLOFF. 2003. Spatially explicit influences on northern goshawk nesting habitat in the interior Pacific Northwest. Wildlife Monographs 154. MONSON, G. 1968. The desert pronghorn. Desert Bighorn Council Transactions 12: 63 69. NEU, C. W., C. R. BYERS, AND J. M. PEEK. 1974. A technique for analysis of utilization-availability data. Journal of Wildlife Management 38: 541 545. OCKENFELS, R. A., A. ALEXANDER, C. L. DOROTHY TICER, AND W. K. CARREL. 1994. Home ranges, movement patterns, and habitat selection of pronghorn in central Arizona. Arizona Game and Fish Department Research Branch Technical Report 13, Phoenix, USA. OCKENFELS, R. A., C. L. TICER, A. ALEXANDER, AND J. A. WENNERLUND. 1996. A landscape-level pronghorn habitat evaluation model for Arizona. Arizona Game and Fish Department Research Branch Technical Report 19, Phoenix, USA. O CONNOR, R. J. 2002. The conceptual basis of species distribution modeling: time for a paradigm shift? Pages 25 33 in J. M. Scott, P. J. Heglund, M. L. Morrison, J. B. Haufler, M. G. Raphael, W. A. Wall, and F. B. Samson, editors. Predicting species occurrences: issues of accuracy and scale. Island Press, Washington, D.C., USA. O GARA, B. W., AND J. D. YOAKUM. 1990. Additional capture methods and habitat suitability criteria for pronghorn translocations. Pronghorn Antelope Workshop Proceedings 14: 51 62. PARADISO, J. L., AND R. M. NOWAK. 1971. Taxonomic status of the Sonoran pronghorn. Journal of Mammalogy 52: 855 858. PHELPS, J. S. 1981. Biological observation on the Sonoran pronghorn. Pages 28 33 in The Sonoran pronghorn. Arizona Game and Fish Department Special Report Number 10, Phoenix, USA. TURNER, R. M., AND D. E. BROWN. 1994. Sonoran desertscrub. Pages 181 221 in D. E. Brown, editor. Biotic communities: southwestern United States and northwestern Mexico. University of Utah Press, Salt Lake City, USA. UNITED STATES DEPARTMENT OF THE INTERIOR, OFFICE OF THE SECRETARY. 1967. Federal Register 32. 11 March 1967; page 4001. UNITED STATES FISH AND WILDLIFE SERVICE. 1982. Sonoran pronghorn recovery plan. United States Fish and Wildlife Service, Albuquerque, New Mexico, USA. UNITED STATES FISH AND WILDLIFE SERVICE. 1994. Sonoran pronghorn recovery plan revision. United States Fish and Wildlife Service,

34 Wildlife Society Bulletin 2005, 33(1):24 34 Albuquerque, New Mexico, USA. UNITED STATES FISH AND WILDLIFE SERVICE. 1998. Final revised Sonoran pronghorn recovery plan. United States Fish and Wildlife Service, Albuquerque, New Mexico, USA. WALLACE, C. S. A., J. J. WALKER, AND S. E. MARSH. 2002. Characterizing landscapes accessed by Sonoran pronghorn antelope using remote sensing data. Final Project Report, Cooperative Agreement 17 00 2 0068. Arizona Remote Sensing Center, Office of Arid Lands Studies, University of Arizona, Tucson, USA. WESTERN REGIONAL CLIMATE CENTER. 1973 2002. Arizona climate summaries. Available online at http://www.wrcc.dri.edu/summary/ climsmaz.html (accessed 9 October 2004). WRIGHT, R. L., AND J. C. DEVOS, JR. 1986. Final report on Sonoran pronghorn status in Arizona. Arizona Game and Fish Department, Phoenix, USA. YOAKUM, J. D. 1980. Habitat management guides for the American pronghorn antelope. United States Department of the Interior, Bureau of Land Management technical note 347. Chantal (Chasa) O Brien (center) has been a research biologist with the Arizona Game and Fish Department since 2002. She received her M.S. in wildlife ecology from the University of Arizona and B.S. in wildlife and fish conservation biology from the University of California, Davis. Her current projects focus on desert ungulates in southern Arizona. Steven (Steve) Rosenstock (left) has been a research biologist with the Arizona Game and Fish Department since 1992 and also serves as adjunct professor in the School of Forestry, Northern Arizona University. His current research efforts focus on wildlife ecology and management in the Sonoran Desert of southwestern Arizona. John J. Hervert (not shown) is a wildlife program manager for the Arizona Game and Fish Department in Yuma. He received his B.S. and M.S. from the University of Arizona. His research focus is ungulate ecology in desert environments (25 years) and population estimation. Jill L. Bright (not shown) is the Sonoran pronghorn projects coordinator for the Yuma Region, Arizona Game and Fish Department. She received a B.S. in wildlife management from Utah State University and an M.S. from Northern Arizona University. Currently she specializes in Sonoran pronghorn management, and has worked on other endangered species such as blackfooted ferrets and spotted owls. Susan (Sue) Boe (right) has been a GIS analyst with the Arizona Game and Fish Department since 1992 after receiving her MS in biology from the University of Minnesota, Duluth. She supports numerous biologists in the Research Branch with spatial analysis, modeling, and various GIS needs. Special Section Associate Editor: Krausman