Deserts of the southwestern United States have many competing human uses that can, at local scales, compromise and potentially extirpate populations of the federally threatened Agassiz's desert tortoise, Gopherus agassizii, (Lovich and Bainbridge 1999; Lovich et al. 2011; Averill-Murray et al. 2012). At a range-wide scale, concerns regarding the effects of climate change on the sustainability of tortoise populations have also emerged (Henen et al. 1998; Duda et al. 1999; Longshore et al. 2003; Barrows 2011; McCoy et al. 2011; Averill-Murray et al. 2013; Lovich et al. 2014). Although as ectotherms, reptiles have the capacity to behaviorally and physiologically adapt to temporary, multiple-year temperature and drought fluctuations, (e.g., López-Alcaide et al. 2014), G. agassizii mortality increases with extended drought conditions (Woodbury and Hardy 1948; Turner et al. 1984; Peterson 1994; Berry et al. 2002; Longshore et al. 2003; Zylstra et al. 2013; Lovich et al. 2014). A resurvey of a historic plot has already documented substantial climate-implicated tortoise population declines where previously dense populations occurred (Lovich et al. 2014).

Dispersal, usually via latitudinal or elevation shifts, is one expected response by species to a changing climate (Parmesan and Yohe 2003; Hof et al. 2011; Reif and Flousek 2012). However, while a gradual climate shift may allow species generations to track and disperse to preferred climates, rapid shifts in climate can create challenges for less-mobile species, such as reptiles, putting them at a greater risk of local extinction (Root and Schneider 2002; Hannah et al. 2005). Although the current rapid rate of climate change has been contrasted against what was once thought to be more-gradual rates during the Pleistocene, that perception of more-gradual past climate shifts has been challenged (Alley et al. 2003; Hof et al. 2011; Annan and Hargreaves 2013). Habitat suitability for tortoises, as well as for other indigenous species of the southwestern deserts, is changing spatially and temporally (e.g., Lovich et al. 2014). Recent and currently occupied tortoise habitat will not necessarily be suitable in coming decades. As the desert becomes fragmented by land uses incompatible with tortoise habitat, there is an increasing challenge to determine how and where to mitigate those impacts.

The broad distribution of tortoises in the Mojave Desert today may indicate how they survived past climate shifts, at least some of which were rapid (Henen et al. 1998; Hof et al. 2011). For a species with low dispersal capacities, one mechanism to survive rapid climate change is to occupy habitats where both existing and the future climate conditions overlap that species climate tolerance. Such regions are climate refugia, and individuals occupying refugia would not need to disperse to sustain populations if temperatures increase as predicted (Barrows and Fisher 2014; Hannah et al. 2014; Keppel et al. 2015). If refugia were dispersed across the landscape, once climate stabilizes tortoises could, over time, populate nearby newly suitable habitats as long as critical areas of connectivity are maintained (Dissanayake et al. 2012; Averill-Murray et al. 2013). Tortoise climate refugia, especially those large enough to sustain populations over extended periods, thus have great conservation importance and should be protected from land uses incompatible with persistence of those populations.

Here we present results of a study aimed at identifying potential desert tortoise climate refugia within training areas, and adjacent study areas, of the US Marine Corps Air Ground Combat Center (MCAGCC) at Twentynine Palms, California; one purpose of which is to help MCAGCC manage its lands for G. agassizii. This region experiences many uses from military training to recreation (e.g., off-road vehicle use), mining, livestock grazing and utility corridors, has growing urban centers, and is part of intense studies for renewable energy projects. It is also a region of North America that may experience some of the most-severe departures from current conditions as a result of climate change (Seager et al. 2007; Diffenbaugh et al. 2008; Kerr 2008). While tortoise habitat has been modeled using similar methods (Heaton et al. 2008; Nussear et al. 2009; Barrows 2011), our approach is novel in that we 1) explicitly identify climate refugia, which are high-value conservation targets, 2) model tortoise habitat at a fine scale that more readily identifies those refugia and focuses management decisions, and 3) specifically focus on lands where tortoise protection must be balanced with other land uses. Using this approach, we provide a framework for more-effective conservation of and mitigation for this threatened species.

METHODS

Study Area

We confined our study area to the MCAGCC, expansion study areas (Department of Navy 2012), and a 10-km buffer extending from those boundaries (Fig. 1). This area was divided into 381,690 150 × 150 m (2.25 ha) cells, encompassing an area of 858,802.5 ha.

View Figure

• Download as Powerpoint
• Download Figure

Habitat Suitability Modeling

We used the Mahalanobis distance statistic (D²) (Clark et al. 1993; Rotenberry et al. 2002, 2006; Browning et al. 2005) to model the distribution of suitable habitat for G. agassizii. The Mahalanobis statistic is one of several habitat suitability modeling tools available and consistently ranks well in its performance compared to other methods (Griffin et al. 2010; Liang et al. 2014). Unique among the techniques, the Mahalanobis statistic may be refined by partitioning it into separate components (Dunn and Duncan 2000; Rotenberry et al. 2002, 2006). This partitioning is based on a principal components analysis of the selected model variables in the calibration data set. Each of the partitions are additive, orthogonal variable combinations that explain increasingly more variance until the final partition, the full model, captures the full range of variance exhibited in the calibration data. The first partition is typically the most restrictive partition; the last partition is often the most inclusive.

For any location within the study area, Mahalanobis models yield an index of habitat similarity (HSI) to the multivariate mean of the habitat characteristics of a calibration data set derived from target species locations (Rotenberry et al. 2002, 2006). This statistic requires only species-presence data for the dependent variable. Because only positive occurrence data are required, any valid record can be used regardless of survey method, as long as there is sufficient precision to the record's location. This also avoids the dubious assumption of correctly identifying unoccupied habitat (Knick and Rotenberry 1998; Rotenberry et al. 2002; Browning et al. 2005). Recent tortoise location data for our analysis had two main sources, 2752 observations from 2008 and 2009 surveys adjacent to MCAGCC and 920 observations from 2010 and 2011 surveys within western training areas of MCAGCC; all observations were recorded with Global Positioning System (GPS) coordinates in the field with a horizontal precision of < 10 m (A.E. Karl, unpubl. reports). Tortoise presence in a cell represents the dependent variable.

Independent environmental variables were calculated for all cells; each cell was then scored based on presence of a tortoise observation and used to construct the calibration data set from which we created a desert tortoise niche model based on Mahalanobis distances. Once a model was created, Mahalanobis distances were used to calculate an HSI for every cell on the map. Following Rotenberry et al. (2006), HSIs were rescaled to range from 0 to 1, with 0 being the most dissimilar and 1 being the most similar to the multivariate mean of known occupied habitat characteristics based on the calibration data set. ArcGIS 9.3.1 software (ESRI 2009) was used to illustrate the model of habitat similarity, with an HSI assigned for each cell within the mapped study area.

While we had 3672 recent ‘current conditions' tortoise observations (within the past 7 yr) to use for constructing habitat suitability models, tortoise surveys did not evenly cover our entire study area. To reduce clustering, we randomly selected 500 occupied cells across the surveyed area for the calibration set so as to be more-evenly distributed within those survey areas; the more-spatially unbiased and unclustered the observations the better the resulting model is at representing the distribution of suitable habitat (Phillips et al. 2009). We minimized biases or over-weighting by using occupied cells only once regardless of how many observations (live tortoise, fresh scat, tracks, or an active burrow) may have occurred in that cell. Also, to prevent model over-fitting we maintained a ratio of more than 10 occupied cells per independent variable (Osborne and Costello 2004).

Model Selection

Identifying the model that best fit the distribution of occupied cells was a multistep process. First, the median HSI values for each model partition were inspected to identify those that had the highest median values. From those partitions with the highest values we then examined the mapped representations of the selected model partition to determine which partition best encompassed the observation points without including areas known to be unoccupied (i.e., lava flows, playas, and steep mountains). From these steps we selected our best current conditions model. As a final step, we compared the distribution of the remaining 3172 occupied cells that we did not include in constructing the models to the coverage of the modeled habitat. While there is no established threshold HSI for selecting models, our model had the highest HSI (partition 7, HSI = 0.964), explained 90% of the total variance in the calibration cells, and encompassed 94% of the 500 calibration cells and 92% of those remaining cells that were not used in constructing the model. For each model iteration, the hectares of suitable habitat for HSI values ≥ 0.6 were calculated and the distribution of that suitable habitat was mapped. The selection of HSI values of ≥ 0.6 for model comparisons, while somewhat arbitrary, corresponded to visually close fits between location data and the niche model distributions, and on the 0 to 1 scale indicated moderate to high habitat suitability. This HSI criterion successfully modeled G. agassizii habitat in nearby Joshua Tree National Park (Barrows 2011) as well as that of several other species (Barrows and Fisher 2014).

Model Variables

We selected the same independent variables and successfully modeled the niche of desert tortoises nearby (Barrows 2011). Those variables were limited to and derived from existing publicly available and limited-access geographic information system (GIS) layers or were derived from a digital elevation model (DEM) of the study area. We constructed partitioned Mahalanobis D² models with different suites of abiotic variables derived from GIS layers readily available in 2008 (Table 1); soils (Natural Resources Conservation Service 2010); ruggedness (Sappington et al. 2007; US Geological Survey 2009); and climate (PRISM Climate Group 2006). Given the large number of observations, we were able to include all independent abiotic variables describing climate and topography employed in constructing the niche model (Table 1). Sixty-seven variables were calculated; from those, 11 were identified that provided meaningful metrics which influenced the distribution of desert tortoises (Table 1).

Table 1. Variables used in the development of the niche model for desert tortoises at the Marine Corps Air Ground Combat Center (MCAGCC).

View Fullscreen

Modeling Climate Sensitivity

Downscaled Global Circulation Models (GCM) have, as with any model, a measure of uncertainty. When applied to a complex topographic unit such as our study area, that uncertainty and the potential for error may be amplified. Therefore, we opted to use a temperature gradient approach, not making any assumptions as to what position along that gradient will match actual levels of climate change (Barrows and Murphy-Mariscal 2012). Using our best-performing niche model based on current environmental conditions, we then fit that model to new, map-point data sets with new values for mean maximum July temperature increased by 1°C and then by 3°C. A +3°C shift approximates a maximum expected change from current conditions over this century (Intergovernmental Panel on Climate Change [IPCC] 2013). We used maximum July temperatures because July represents a period of activity (Henen 1997; Henen et al. 1998) and physiological strain (Nagy and Medica 1986; Peterson 1996a; Christopher et al. 1999) for G. agassizii.

Based on the available GIS climate layers, current mean July maximum temperatures within our region have a strong negative correlation with precipitation (r = −0.9524, p < 0.0001) (PRISM Climate Group 2006). Assuming the current relationship between temperature and precipitation remains the same, (i.e., as summer temperatures increase, precipitation decreases), the inclusion of that temperature variable alone serves as a proxy for shifts in mean precipitation under increasing temperatures. Decreasing winter rainfall as temperatures increase is consistent with recent modeled climate change predictions for this region (Gao et al. 2012). Alternatively, incorporating downscaled GCM precipitation projections across the complex topography of our study area would add unknown, but potentially large, errors to projected rainfall amounts. Modeling a range of possible shifts in precipitation would also result in decoupling the current correlation between the two climate variables and would create temperature–precipitation combinations that do not currently occur within that landscape. Even if the new modeled climate was within the physiological tolerances of the tortoises, the model would not match current combinations of climate variables, even at higher elevations, and would indicate that little or no future suitable habitat will occur. This potential false negative for identifying suitable habitat represents a source of error if precipitation is added as separate model variable. We opted for a conservative approach of retaining mean maximum temperature as a proxy for a future precipitation condition.

RESULTS

The Habitat Suitability Model (HSM) created a multivariate synthesis of the independent variables for the current conditions model (Fig. 2). This synthesis identified 283,900 ha (33% of our study area) as currently suitable tortoise habitat, with high summer temperature averaging 38.3°C, SD = 1.7°C and elevation averaging 764 m, SD = 205 m. Tortoise habitat was less rugged, less steep, had less clay and silt, and was cooler and sandier than the mean for areas not modeled as suitable for tortoises. Areas not modeled as currently suitable tortoise habitat comprised 574,900 ha and had a mean elevation of 614 m, SD = 346 m and a mean high summer temperature of 39.5°C, SD = 2.7°C. Suitable tortoise habitat was concentrated in the western portion of the study area and excluded the lowest, flattest areas, which are primarily in the eastern portion of the study area, and the rugged, rocky, and higher mountain slopes.

View Figure

• Download as Powerpoint
• Download Figure

Simulating a modest temperature increase of 1°C resulted in a westward shift in suitable habitat and a decrease in area to 214,665 ha, a decrease of 24% from the current conditions HSM. The average elevation of suitable habitat in this model was 836 m. A potential, roughly end-of-the-century increase of 3°C in summer temperature shifted the regions of suitable tortoise habitat west and the mean elevation was 950 m; thus 186 m higher than the modeled current habitat. The area of the 3°C-shifted habitat shrunk by 55% to 127,650 ha (< 15% of our study area) (Fig. 3). While this represents a substantial reduction in suitable habitat, much of that climate-shifted area remained in large, contiguous polygons and overlapped the modeled area for currently suitable tortoise habitat. The overlap between the current and +3°C climate-shifted modeled habitat was 115,790 ha; comprising 91% of the +3°C climate shifted modeled habitat (Fig. 3).

View Figure

• Download as Powerpoint
• Download Figure

DISCUSSION

By identifying climate refugia, there is an opportunity to focus protection of tortoises within areas where it will be most effective. This includes efforts to reduce other multiple use-related stressors that may exist within those refugia as well as identifying areas that may be suitable for population augmentation via tortoise translocation (US Fish and Wildlife Service [USFWS] 2011). By first modeling the current distribution of suitable tortoise habitat, and then increasing the mean maximum summer temperature variable in the HSM by 1°C and then 3°C, we were able to model where suitable habitat for desert tortoise is likely to shift under projected levels of climate change, and so identify climate refugia. An additional objective of our analyses is to create a framework for focusing future tortoise survey efforts that will improve our understanding of the distribution of suitable habitat and the effects of climate on the sustainability of tortoise populations.

A +3°C shift in mean maximum summer temperatures approximates a maximum expected change from current conditions over this century (IPCC 2013). The modeled reduction in suitable habitat with a +3°C shift is substantial; nevertheless, large areas within our study area will likely remain suitable. A substantial portion (41%) of the current modeled habitat would still be suitable under the +3°C scenario. For tortoises, their likely response to climate change will be differential recruitment and survivorship, favoring areas within their current distribution that are, and continue to be, within a preferred climate envelope. Lower recruitment and higher mortality will occur in areas that become unsuitable due to increased drought intensity and duration. High tortoise mortality rates are already being observed at lower elevation sites in this region that previously were known to have productive, high-density tortoise populations (Lovich et al. 2014). Climate refugia for tortoises experiencing a +3°C shift in mean maximum July temperatures do not require the tenuous assumption that the tortoise will move and locate “newly suitable” habitat as temperatures increase and precipitation decreases. Areas where these refugia occur in relatively large, connected polygons have heightened conservation importance, and they should be managed to minimized threats of future development or other detrimental land uses.

Three other HSMs have been developed for desert tortoises prior to our current model, all with the objectives to support conservation and focus mitigation efforts for this threatened species. Nussear et al. (2009) constructed an HSM for G. agassizii covering the full extent of the species distribution. Over such a large region their cell size for constructing the model was necessarily large (1 km²), and the low resolution limited the ability to make management decisions (Barrows and Murphy-Mariscal 2012). The Nussear et al. (2009) model was constructed using many of the same abiotic variables we employed; however, they did not conduct climate change simulations. Heaton et al. (2008) modeled areas appropriate for translocating tortoises from areas of anthropogenic habitat loss or degradation associated with military operations at Fort Irwin. Their cell size was larger (2.59 km²) than that used by Nussear et al. (2009). The Heaton et al. (2008) model incorporated variables that defined many of the potential conflicting land uses that would render lands unsuitable for translocations but did not consider the effects of climate change. Our results improved upon those earlier efforts by using finer-scale cells (150 m²) and then explicitly identifying climate refugia, those regions where tortoises should find both suitable habitat today and under projected climate shifts. By not requiring tortoises to locate or be translocated to these refugia in order to survive otherwise persistent warmer and drier future conditions in nonrefugia areas, these refugia have high conservation value. The finer-scale analysis can allow land managers (e.g., agencies) clearer distinction between lands of current and future high and lesser tortoise habitat suitability, allowing them to make informed decisions as they balance both tortoise protection and multiple-use mandates. When translocation is required to take tortoises out of harm's way, managers will be able to move them to refugia where they will find suitable habitat currently and in the future.

Our model used independent variables similar to Nussear et al. (2009) and the same independent variables as Barrows (2011) used for Joshua Tree National Park (JTNP), an area a few kilometers south of MCAGCC. Both models identified the same independent variables as important in forming the niche model, reinforcing that these variables are likely critical to the biology of G. agassizii. Although we do not know whether these variables relate to tortoise mobility (e.g., ruggedness and slope), friability of soils for digging burrows (e.g., soil composition), or the biology of the tortoises or their food plants (e.g., temperatures), their commonality suggest they should be evaluated in detailed, field-based studies. Combined with such field-based studies, ours and other models may help us better understand what environmental elements are critical to protect for this threatened species.

Rainfall events are critical drivers of temporal shifts in habitat quality for tortoises (Barrows 2011). Winter rain stimulates the germination of annual plants, a crucial food source for Agassiz's desert tortoises (Turner et al. 1984; Henen 1997; Jennings 2002; Jennings and Berry 2015), and summer rains enable tortoises to drink, forage, and avoid negative energy and protein balances (Nagy and Medica 1986; Peterson 1996a, 1996b; Henen 1997). While acknowledging the importance of precipitation, incorporating downscaled GCM precipitation projections across the complex topography of our study area would add unknown, but potentially large, errors to projected rainfall amounts. Additionally, any shift away from the current close correlation between maximum summer temperatures and rainfall would generate a model that cannot match any current combination of climate conditions within the modeled area and would indicate that little or no suitable habitat will occur, even if the new modeled climate is within the physiological tolerances of the tortoises. This potential false negative for identifying suitable habitat represents an unavoidable source of error if precipitation is added as separate model variable. We opted for a conservative approach of retaining mean maximum temperature as a proxy for a future precipitation condition consistent with a prediction of reduced winter rains as temperatures increase (Gao et al. 2012).

Habitat suitability modeling is typically restricted to variables describing mean climate conditions, which is a shortfall when used in deserts that are defined, in part, by extreme variability and unpredictability (Louw and Seely 1982). The desert's discrete, stochastic events such as isolated summer downpours, floods, or severe extended droughts directly impact the localized and temporal dynamics of habitat quality and tortoise recruitment and survivorship (Barrows 2011). Three extended droughts have been recorded in the Mojave Desert, the first from 1910 through the early 1930s, another from the mid-1940s through the early 1970s (Redmond 2009), and a third from the late 1980s to the present. Due to the unpredictability of droughts, we were unable to incorporate drought events in our models because available climate data include only long-term (e.g., decades) means. Habitat suitability models represent testable hypotheses and should never be accepted without on-the-ground validation and assessments of current tortoise habitat as well as the impacts of long-term droughts. While desert tortoises possess behavioral and physiological adaptations to periods of weather stress, such as droughts (Nagy and Medica 1986; Peterson 1996a; 1996b; Henen 1997; Henen et al. 1998; Duda et al. 1999; McCoy et al. 2011), they endure a 40% loss of body mass, a 60% loss of body water (volume, % of body mass), and large variations in plasma osmolality (Peterson 1996a). Additionally, juveniles and adults experience increased mortality in droughts (Woodbury and Hardy 1948; Turner et al. 1984; Peterson 1994; Berry et al. 2002; Longshore et al. 2003; Zylstra et al. 2013; Lovich et al. 2014).

We also did not assess the impacts of multiple human uses on habitat suitability for tortoises. Those stressors have been identified as reducing habitat suitability (Averill-Murray et al. 2012); however, each has a unique impact that has not been fully quantified, mapped, and incorporated into the GIS layers necessary for incorporation in HSMs. We view the identification of climate refugia as the first step in a conservation strategy for desert tortoises. To the extent that those uses occur within sites identified as climate refugia, they could, depending on the extent, characteristics, and intensity of that activity, reduce the ability of those refugia to sustain tortoise populations, even if climatic conditions remained suitable.

No model is a perfect representation of a species habitat. Nevertheless, models provide heuristic guidelines to focus management actions where they will have the greatest benefit for tortoise conservation. These models should not be viewed as endpoints but as an iterative process by which the models are improved with new observations and insights to critical habitat features that enable populations to persist over time. Until validated with field studies, habitat suitability models represent hypotheses as to the current and future distributions of appropriate tortoise habitat. These hypotheses should foster additional research identifying whether tortoise densities and demographic structure are more secure and whether tortoises can adapt to shifting climates more effectively within than outside modeled refugia. With an understanding of how climate and topography relates to habitat suitability for tortoises, the effects of additional stressors from a multitude of human uses of the desert can be better partitioned and understood.

Another potential use of these results is to create a tiered scale for tortoise habitat compensation. Areas not modeled as either current or future habitat would be at the bottom of that scale, lands not currently suitable but likely suitable in the future and lands modeled as current but not future habitat as being mid-scale, and climate refugia considered as the highest level for the long-term sustainability of tortoise populations because those areas would not require tortoise movement to find suitable habitat. This tiered approach could also be used to locate multiple and competing uses of the desert landscape and limit their encroachment to the highest level conservation areas.

Multiple potential conflicting land uses, diseases, invasive plants, and climate change present challenges for the conservation of Agassiz's desert tortoises. Mitigating habitat losses due to land use shifts to renewable energy, agriculture, urban development, off-road vehicle recreation, and military training generally include protecting lands elsewhere, and in some cases translocating tortoises to those or other more secure habitats. Climate change complicates this mitigation approach. Protecting existing tortoise habitat, which will become unsuitable as temperatures rise and droughts increase in intensity and duration, does not mitigate for anthropogenic habitat losses. Our framework for the identification of climate refugia, regions where tortoises currently occur and where suitable habitat will likely persist despite climate shifts, provides an important step toward providing effective mitigation to ensure that desert tortoise populations can persist into the future.