We studied movements of adult tortoises in seven different locations throughout the Mojave Desert of Nevada and California (Figure 1) between 2011 and 2021. Sites primarily consisted of Mojave desert scrub dominated by creosote-bursage (Larrea tridentata-Ambrosia dumosa) associations, and other important shrub components (e.g., Yucca spp., Psorothamnus spp., Cylindropuntia spp.; Turner, 1994). Sites varied topographically from rugged mountain terrain in McCullough Pass, NV (max slope 38°) to flat sandy terrain at the Nipton site.

Collectively the sites were impacted by an array of disturbances including roads, railroads, solar development, OHV use, and recent wildfire (Table 1). Most sites had dirt roads used by recreational OHVs as well as power line maintenance traffic. The Silver State, NV and ISEGS South, CA sites had tortoise exclusion fencing surrounding utility-scale solar plants that were installed prior to data collection (2014 and 2011; respectively). The Nipton, CA site had a low-traffic volume paved road with no posted speed limit that served as access for an active railroad maintenance site that also intersected the study site. The China Lake, CA site had two unfenced, fairly high-traffic highways (speed limit = 88.5 km/h) that resulted in several tortoise mortalities in the years prior to and during our study (J. Hendrix, Naval Air Weapons Center—China Lake—pers. comm.). Portions of the surrounding public lands were designated for unrestricted “open” travel of OHVs; two areas were designated as seasonal grazing allotments for sheep in years of good forage production (Spangler Hills and Cantil Common; grazed 2019 and 2020 during our study period). The eastern portion of the site was within the off-limits boundaries of the China Lake Naval Weapons Station, which does not allow public access. The remaining areas in the China Lake site and all other sites were located on land managed by the United States Bureau of Land Management (BLM). The Hidden Valley, NV site experienced lightning caused wildfires during 2005 which led to large areas being burned (Drake et al., 2015). Site characteristics are summarized (Table 1) and maps of sites are provided (Figure 1).

Tortoises were found during surveys of one-km² plots at Nipton, McCullough Pass, Sheep Mountain, ISEGS South, and Silver State as part of another study (Mitchell et al., 2021). Surveys were also conducted in the area within 1 km of Randsburg Wash Road (China Lake) and on 400 m² vegetation study plots in Hidden Valley (Drake et al., 2015).

All tortoises involved in the study were outfitted with Very High Frequency (VHF; Model RI-2B; Holohil Systems Ltd.) radio-transmitters. Tortoises at all sites except Hidden Valley were equipped with GPS loggers (i-gotU GT-120; Mobile Action Technology). Tortoises at these sites were physically relocated once per month using VHF telemetry equipment. During these visits the GPS loggers were swapped for battery recharge and data downloading. Tortoises at the Hidden Valley site were outfitted with custom-made Advanced Telemetry Systems GPS/VHF combination loggers. Tortoises at the Hidden Valley site were relocated at approximately one-week intervals, and GPS loggers were replaced approximately every 3 months for data downloading and charging. All tortoises were greater than 200 mm midline carapace length; tortoises smaller than this size could not carry both the radio and logger without excessively modifying the vertical profile of the tortoise. Most GPS loggers were programmed to collect a point every hour, but some of the Hidden Valley and McCullough Pass data were collected every half hour during initial deployment trials and later filtered to provide an hourly dataset. The number of tortoises monitored, and GPS locations recorded per site are provided in Table 1. Thirteen of the twenty tortoises at the Silver State site were translocated from the footprint of the nearby utility-scale solar installation in 2014, and one of the nine tortoises at the ISEGS South site was translocated from the nearby utility-scale solar installation in 2011. Tortoises were translocated to areas just outside the footprint of the utility-scale solar installations. All but one of the tortoises that interacted with the fences were translocated. Therefore, we were not able to separate movement behavior differences due to the fences and translocation; translocation is known to alter movement behavior in tortoises (Nussear et al., 2012). All tortoises were handled in accordance with a USFWS Permit (permit TE-030659-10), Nevada Department of Wildlife Scientific Collection Permit 317351, University of Nevada, Reno Animal Care and Use Committee protocol (IACUP 00671), and a Memorandum of Understanding with the California Department of Fish and Wildlife (all to T. Esque).

We chose environmental predictors that we hypothesized were important to tortoise movement behavior and used them as covariates in our model construction. We used the TIGER road datasets for dirt and paved roads in our study areas and modified them to better match satellite imagery (United States Census Bureau, 2021). We log-transformed distance to linear features (roads, railroad, and fences) to reflect the spatial decay around the impact of these features as we expected the effect of these covariates to be localized near the disturbance. We used the National Land Cover Dataset (NLCD) shrub coverage layer to determine if differences in shrub cover may act as a cue for tortoises to recognize disturbed areas (Homer and Fry, 2012; Rigge et al., 2021). To represent areas burned in wildfire, we digitized a polygon that corresponded with the burned areas seen on Google satellite imagery as available layers from the BLM did not accurately reflect the wildfire boundaries visible on satellite imagery and from data collected while tracking tortoises (Drake et al., 2015). Additionally, we considered slope as it has been shown to alter tortoise movement (Hromada et al., 2020); slope was derived from the USGS Digital Elevation Model using the terrain function in package raster (U.S. Geological Survey, 2017; Hijmans et al., 2022).

Open OHV areas on BLM lands typically have both established routes designated by the BLM and user-created routes that are often missing from BLM road and trail inventories. BLM designated open routes are intended for travel by street legal vehicles, while user-created routes are tracks created by OHVs driving outside of established routes. We found that established routes in the Spangler Hills OHV Area at the China Lake site were best represented by data sourced from Owlshead GPS, a service intended for OHV recreationists (Friends of Jawbone, 2022). We also used a polygon from the same source to represent the entire open OHV area. However, we found that although the entire area is designated for off-trail OHV use, certain areas within the boundaries used by our tortoises saw little to no off-trail use, likely due to their greater distance from major staging areas, or rougher terrain. These included the areas north of Randsburg Wash Road and areas that were rocky hills with large boulders. We created another polygon that excluded these areas and tested both as categorical predictors in our modeling efforts. Sheep grazing occurred during years of higher productivity across the area contained within the entire OHV polygon; thus, the two polygons represent a combination of grazing pressure and relatively high OHV use versus mostly just grazing pressure with relatively low OHV use.

Hidden Markov models (HMMs) are a popular approach for analyzing temporally regular animal location data for inference on movement behavior (McClintock et al., 2012; McClintock and Michelot, 2018). HMMs are a form of state-space models that assume that observed data arise from a number of “hidden” (i.e., latent) states. In the case of telemetry data these are movement behavior states that can represent activities such as exploration, foraging or rest (Patterson et al., 2009). The state changes can be identified from differences in the distribution of the step lengths and turning angles of the relocation dataset (Morales et al., 2004; McClintock et al., 2012). Many HMMs fit to animal movement data can only identify two biologically meaningful states; one that corresponds to a moving state and one that corresponds to a resting or encamped state (Beyer et al., 2013; McClintock et al., 2014). An extension of the HMM framework (known as a generalized HMM) allows for inference of changes in behavioral states and movement parameters using environmental and individual level covariates (McClintock et al., 2017; Carter et al., 2020).

We fit generalized discrete-time hidden Markov movement models (HMMs) to tortoise GPS location data using the package momentuHMM (version 1.5.4) in program R (version 4.0.4; McClintock and Michelot, 2018; R Core Team, 2018). Hidden Markov movement models require bursts of continuous GPS data; however, tortoise behavior (i.e., burrowing and cave dwelling) often results in missed fixes. As we are primarily interested in above-ground activity in this study, we were not concerned about missed fixes during burrow use. We partitioned our GPS telemetry data first to contain temporally continuous segments that had no more than 2 h of missing data and greater than 36 consecutive hours. To filter out inaccurate fixes from the GPS loggers, we used the uere.fit function from package ctmm to estimate error for the GPS loggers based on locations taken with stationary loggers placed at study sites (Calabrese et al., 2016). We placed GPS loggers within unoccupied tortoise burrows to determine how this positioning would alter fix accuracy. For the custom made ATS GPS loggers (used in the Hidden Valley site), we fit an error model using the HDOP (Horizontal Dilution of Precision) data column as a predictor of horizontal error to the stationary logger data. For the i-gotU GPS loggers, we fit an error model without using the EHPE (Expected Horizontal Position Error) column as much of the data were retrieved from the loggers without the column. We filtered extreme outlier points for the i-gotU GPS loggers by removing points from both the stationary logger and tortoise telemetry datasets that had greater than 25 m of elevation error recorded altitude—elevation from a digital elevation model (Laver et al., 2015), and then fit an error model on the stationary logger data. We then used the outlie function in ctmm to detect and remove telemetry points that had a movement rate greater than that expected for a tortoise (300 m/h; Nussear and Esque, unpublished data) after accounting for locational error.

As HMMs assume exact temporal consistency, GPS data had to be imputed by fitting a continuous-time correlated random walk to the data using the crawlWrap function from the momentuHMM package as a wrapper around the crwlMLE function from package crwl (version 2.2.1; Johnson and London, 2018). We fit 2-state HMM models with the two states representing a “moving” state and an “encamped” state. The “moving” state was described by longer hourly step lengths and low turning angles while the “encamped” state was described by short hourly step lengths and high turning angles. We fit a two-state model as we anticipated that our data would not allow for differentiation between different tortoise activities when movement lengths were below 10 m (e.g., foraging vs. thermoregulating) and different movement purposes (e.g., nesting vs. socializing; Pohle et al., 2017). We used a gamma distribution for the step length and a von Mises distribution for turning angles as well as a zero-mass parameter to account for time steps with no movement.

We first fit models to determine if environmental covariates altered the probability of state transition between the moving and stationary states. Tortoises are largely diurnal, and their activity is largely determined by environmental conditions such as temperature and food availability which change throughout the year (Woodbury and Hardy, 1948; Ruby et al., 1994b; Rautenstrauch et al., 2002; Ennen et al., 2012). We fit models incorporating hour and Julian date and their interaction to find the best combination of the two variables to explain transition probability of tortoises changing from encamped to moving states. We used the cosinor function from momentuHMM on the hour variable to allow state transition probability to reflect the typical diurnal activity patterns of desert tortoise (Zimmerman et al., 1994). We also tested models that looked for relationships between sex and transition probability, as male tortoises are often more active than females (Agha et al., 2015a). We then used the top-ranked ranked model (based on AIC), representing circadian and circannual movement patterns, to test additional environmental predictors that we hypothesized to alter the transition probabilities of movement states.

We fit generalized HMMs to the best performing state transition model that related lengths of hourly time steps to environmental covariates and the sex covariate. We restricted our modeling of step related covariates to only consider effects of environmental covariates on steps made while in the moving state, as the mean step length in the encamped state was below that of the positional error (~ 10 m) of our datasets. We fit models relating both the mean and standard deviation of step length and the mean and concentration of turning angles to the environmental covariates. We compared models with additive combinations of step covariates using AIC to find the most supported combination of covariates. For the Hidden Valley site, we examined the potential for an interaction term between the categorical burn predictor and distance to dirt roads. We fit and compared models for each site separately as not all anthropogenic features were present at each site, and we wanted to determine if covariates had similar effects among sites. We also attempted to fit random intercepts for each tortoise as a covariate in both the transition and step models to allow for the potential of individual-level differences in parameters.

The top-ranked behavioral state probability models were similar across our seven study sites, although the best models for some sites did not include all covariates tested, as presented in Table 2. AIC tables for models run at each site are given in the Supplementary Materials. All top models performed better than null models with no covariates. We found that an interaction between the continuous day-of-year variable with the cyclical hour variable best accounted for state, where tortoises were more likely to be moving in the earlier days of the active season (spring) and more likely to be moving during the daylight hours (Figure 3). The best model for every site included sex as a predictor, with males more likely to be in the moving state and more likely to remain in the moving state once moving (Figure 3).

State transition probabilities were affected by distance to dirt roads at every site except Hidden Valley, though the AIC difference was low between the best model at that site and the model including the dirt road covariate (Δ2.4 AIC). In all sites with distance to dirt road in the transition model, tortoises were more likely to be in the moving state when close to a dirt road, except for China Lake and Silver State which showed the opposite relationship (Figure 4). Distance to paved road was included in the top transition model for the Nipton site, but was not in the top model for China Lake, although the difference in AIC was relatively low between the best model and the model that included the paved road covariate for this site (Δ1.1 AIC). Tortoises at the Nipton site were more likely to be moving when closer to the paved road. The distance to railroad covariate was not included in the top model for the Nipton site; likely because tortoises were seldom located close to the railroad. Distance to fence was included in the top model for both sites where fences were present and indicated that tortoises were more likely to be moving when close to the fence (Figure 4).

The categorical covariate representing the OHV use area was in the top model for China Lake; tortoises were more likely to be moving in areas with relatively low OHV use and grazing than in areas without these disturbances. Distance inside the burn scar was in the top transition model for the Hidden Valley site, suggesting tortoises were more likely to be moving in areas closer to the edge of the burned area (Figure 5). Slope was included only in the best transition model for the Sheep Mountain site, suggesting that tortoises are more likely to be in the moving state in areas of higher slope. The only site that contained the shrub covariate in the top model was Nipton, which suggests that tortoises were more likely to be moving in areas of higher shrub cover. For all sites, a random intercept model for each individual was found to improve the AIC as there was considerable variation amongst individuals (Supplemental Figure 1). However, we chose not to include this in subsequent model fits as it added considerably to computational times and did not provide additional information on the covariates of interest.

All top-ranked step length models performed better than null models; selection tables of all models fit examined are available in the Supplemental Materials. Dirt roads were included in the top model for every site where present. Hourly step lengths were both longer and more variable when close to dirt roads at every site except Silver State (Figure 4). The covariate for distance to paved road was in the top step parameter model for both the Nipton and China Lake sites, though they had the opposite effects. When moving closer to the paved road in Nipton, tortoises had longer hourly steps. In contrast, tortoises at China Lake had shorter hourly steps when closer to the paved road (Figure 4). Distance to fence was included as a covariate in the step length models for both the Silver State and ISEGS South sites, which were the only sites with fences. Hourly steps were longer when tortoises were moving near the fences (Figure 4). The OHV covariate representing the relatively heavy use area was included in the step model for China Lake, indicating that tortoises took longer hourly steps when moving in the high OHV use area (Figure 5). The fire scar covariate was included in the best model for the Hidden Valley Site with an interaction with the road covariate that indicated that tortoises took longer hourly steps near roads in burned areas but shorter hourly steps near roads in unburned areas (Figure 4). All the sites that had varied terrain (McCullough Pass, China Lake, Sheep Mountain) included slope in the top step model, indicating that tortoises took shorter steps in areas of higher slope (Figure 5). The shrub covariate was in the best model for McCullough pass and Nipton and indicated that hourly steps were shorter in areas that had higher shrub cover for McCullough pass, but hourly steps were longer in areas of high shrub cover for Nipton (Figure 5).

We found that tortoise movement behavior is influenced by natural features such as slope and shrub cover. Tortoise movement is limited by high slope (Hromada et al., 2020), and we found that tortoises made shorter hourly movements in higher slope areas at all three sites with varied terrain. This finding could be due to the higher cost of movement through rougher terrain, or potentially higher availability of resources in rugged areas. Sloped areas in tortoise habitat often feature rock outcrops that can provide areas where water pools during rain, shelter sites, forage cover, and areas of higher thermal inertia that may buffer extreme temperatures (Nowicki et al., 2019). The only site where we recovered a relationship between movement state and slope is also the only site on which tortoises moved between creosote flats and hills. Some studies have suggested that Mojave desert tortoise populations in areas with varied terrain may be more resilient to declines than those in flat areas (Allison and McLuckie, 2018; Berry et al., 2020c). Future research should focus on the causes and potential importance of these areas in preserving the species. We found that shrub cover was an important predictor of tortoise movement at only two of the seven sites in our study (Nipton and McCullough Pass). At Nipton, we found that tortoises were more likely to move, and take longer hourly steps, in areas of higher shrub cover, while tortoises at McCullough Pass were more likely to take shorter hourly steps in areas of higher shrub cover. The Nipton area is flat, sandy, and has relatively low shrub cover; areas of higher shrub cover generally occur along washes where tortoises prefer to move, though foraging opportunities may be lower (Jennings and Berry, 2015; Gray et al., 2019; Hromada et al., 2020). In contrast, McCullough pass is rugged and rocky with much higher overall shrub cover; tortoises may move more slowly in areas of higher shrub cover here due to higher availability of desirable resources (e.g., shade, forage underneath shrubs), and higher slopes often associated with that cover. While previous studies have shown that tortoises prefer moving in areas of higher perennial vegetation cover (Sadoti et al., 2017; Hromada et al., 2020), our results illustrate the importance of site-specific context in how tortoises are responding to environmental conditions.

Although our analysis leveraged a large GPS dataset, we were limited in scope to only making inference on tortoise surface activity, as our GPS units could not record data while the tortoises were underground. Thus, our inference is on the surface activity of tortoises as data on time spent inactive in burrows was not available for this analysis. Although tortoises spend most of their time in burrows (Rautenstrauch et al., 2002), we were mostly concerned with how tortoises are moving. However, understanding where tortoises use burrows in relation to disturbances is also important in describing how disturbances alter tortoise space use. The use of an additional data loggers, such as a light-sensitive loggers, would enable researchers to capture the times that a tortoise was inactive and in a burrow. This could allow additional inference on surface activity and on how different disturbances and environmental settings may influence this. However, we have demonstrated the utility of these methods to describe the movement behavior of a reptile and how it is influenced by disturbed habitat from different sources. Though outside the scope of this study, the addition of site and year specific weather covariates that may also influence tortoise movement (Duda et al., 1999) would allow for a more comprehensive model of tortoise movement behavior.

Individual-level variation in behavior can be an important aspect to consider in movement ecology (Merrick and Koprowski, 2017; Spiegel et al., 2017). Tortoises have distinct personalities that can play an important role in determining movement patterns and responses to stressors (Currylow et al., 2017; Germano et al., 2017; Le Balle et al., 2021). Animal personality potentially determines how an individual responds to an environmental disturbance or features (Michelangeli et al., 2022), and it is likely that tortoises with different personalities have different propensities to cross roads or pace fences. If so, tortoises that are more likely to cross a road are more susceptible to mortality. As a result, the distribution of the variation in personality syndromes in a population may be altered, which could have emergent effects on population function (Spiegel et al., 2017). Although our results indicated that there was considerable individual-level variability around movement state probability, these models took considerable computational time and models with random slope terms typically failed to converge. Further exploration of individual-level movement variation and personality is warranted to better understand how personality in this species may influence spatial distribution of tortoises and the implications for genetic connectivity.

Our results indicate that tortoise movement behavior is altered by both local and landscape scale disturbances. Both the initial, and the revised recovery plan for the Mojave desert tortoise recommends non-essential or redundant routes to be closed (U. S. Fish and Wildlife Service, 1994, 2011), and our results suggest that this action could aid in restoring historic connectivity of tortoises populations. Although fences alter tortoise movement behavior, they continue to be important tools in restricting tortoise access to dangerous areas; further work could better describe how mitigation (e.g., shade structures) may alter how tortoises interact with fences. OHV activity and grazing continue to be prevalent uses of public lands in the southwestern US; our results suggest that these uses may be degrading tortoise habitat. Further study of how these disturbances affect tortoise populations is warranted to better understand how areas designated for these purposes influence landscape scale connectivity and viability of tortoise populations.