A range-wide model of omnidirectional connectivity for the Mojave desert tortoise (Gopherus agassizii)

This 30-m raster layer depicts log-scale, range-wide, omnidirectional (i.e., 'coreless') connectivity for Mojave desert tortoise rescaled to 1-1000 to improve compression and reduce file size. It was derived from an empirically based model of landscape conductance to movement by desert tortoises and a circuit-theoretic approach to estimate connectivity (see Gray et al. 2019).

This data layer represents an empirically-based, statistical estimate of potential habitat connectivity for the Mojave desert tortoise across its range and at a 30-m pixel resolution. Results from the model reflect estimates of individual movement likelihood (measured in units of log-current flow or log-current density; note that this raster was rescaled linearly to range from 1-1000). These estimates do not reflect the location or locations of individuals or groups of tortoises. Every effort has been made to ensure the quality and completeness of the model results. Nevertheless, it may contain errors or omissions due to incomplete or incorrect source data or other factors. Input data and model results may be updated by the authors at any time.

This layer was created by Conservation Science Partners (CSP). Any applications or publications drawing on these data, in novel analyses, reports, peer-reviewed articles, theses, or other forms, should be undertaken in consultation with CSP. The source of the data should be properly referenced using the citation provided under Credits.

To model landscape conductance, we first estimated habitat quality for Mojave tortoise movement across its entire range. We obtained location data of 31 telemetered tortoises in the northeast Mojave desert from 2007-2010 (Drake et al., 2015; Ken Nussear, unpublished data) and used Brownian Bridge movement models (BBMMs) to estimate the probability of each individual moving through a given area between daily re-locations (Horne et al. 2007). BBMMs were estimated for each individual per year with at least 25 relocations in that year (resulting in n = 47 tortoise years and 1,554 total locations) using the BBMM package (Nielson et al. 2015) for R (R Development Core Team 2013). We identified environmental and landscape variables that are known to affect movement by desert tortoise and that showed sufficient variability within the area of BBMMs. Specifically, we focused on variables describing terrain, climate, vegetation, and the presence of minor roads. All variables were derived at a 30-m resolution across the range and were standardized and rescaled prior to relating them to BBMMs and fitting models to estimate habitat quality for movement.

We obtained range-wide (with the exception of NAIP-derived layers - see below), spatially-explicit data layers representing the identified environmental and landscape variables from public sources and from NatureServe (unpublished data produced for Defenders of Wildlife). Minor road features were selected from the 2016 TIGER/Line products, with the feature class code 'S1400' or 'S1500' (US Census Bureau 2016). Distance to the nearest minor road feature was then calculated at a 30-m resolution. Intermittent or ephemeral stream channels, also known as desert washes, are preferred habitat features and foraging corridors used by the Mojave desert tortoise. A model of desert washes was derived from 2015-2017 imagery from the National Agriculture Imagery Program (NAIP; 1m resolution; US Department of Agriculture Farm Service Agency, 2016). The Normalized Difference Vegetation Index (NDVI) was also derived from 2015-17 NAIP imagery and used to represent availability of vegetation, which tortoises rely on for thermoregulation and forage (Sadoti et al. 2017). NAIP imagery is not available over the Nevada Test and Training Range, and so washes and NDVI were not derived for this area in the north-central part of the tortoise's range, which accounts for approximately 13% of the range-wide area. For the remaining area, washes and NDVI were derived at 1-m resolution before resampling to 30-m resolution. To represent average climate, monthly climate normals for the 30-year period from 1981 to 2010 for 5,330 stations in the desert southwest were downloaded from the National Oceanic and Atmospheric Administration National Climatic Data Center and interpolated using a thin plate spline regression between the climate value, station locations, and elevation at 30m resolution (NatureServe unpublished data). Movement of tortoises is primarily limited to between April and October, encompassing the warmest quarter in the Mojave desert, and movement has been shown to be limited by the hottest temperatures in this period (Sadoti et al. 2017). We therefore included a variable of the maximum temperature in the warmest quarter, derived from the monthly climate normals.

Following methods of McClure et al. (2017), we used generalized linear mixed models and multimodel inference to model the relationship between each BBMM and the variables described above to estimate habitat quality for movement. We included both the linear and quadratic terms of all variables in the models, except NDVI, for which we only hypothesized a positive, linear relationship to tortoise movement within the local area where our relocation data were obtained. Habitat quality for tortoise movement was then mapped continuously across its range using the model-averaged regression coefficients. Where data was missing for washes and NDVI, we imputed these values with the mean variable value (equal to zero after standardizing and rescaling). This map ranged in values from 0-1 and was the baseline to represent landscape conductance for use in connectivity modeling.

We took two additional steps to account for the effects of terrain slope and land cover (i.e., barren ground, open water, and human disturbance) on habitat quality for tortoise movement. Since tortoises are known to largely avoid movement across steeper terrain, we imposed the following penalty to conductance for areas with steep slopes:

〖conductance〗^(1+(〖slope〗^2*40))

This follows a similar approach taken by Dickson et al. (2017) but assumes that conductance declines more rapidly with increasing slope and tortoises will almost completely avoid very high slopes. Lastly, we used a combination of three public datasets - the Global Human Settlement Built-Up Grid (GHSL; 38-m resolution; Pesaresi et al., 2015), the National Land Cover Database 2011 (NLCD; 30-m resolution; Homer et al., 2015), and TIGER/Line products (US Census Bureau 2016)- as well as heads-up digitized polygons of recent solar and golf course development, to assign lower values of habitat quality to human development, water, and barren ground. Built-up presence represented in the GHSL layer (through 2014), primary and secondary roads in the TIGER dataset (feature class code 'S1100' or 'S1200'), open water identified in the NLCD, and heads-up digitized development, were assigned the minimum, non-zero habitat quality value. Barren and planted or cultivated cover identified in the NLCD was assigned the 10th percentile of habitat quality value.

The global model of habitat quality for tortoise movement (i.e., conductance; including the linear and quadratic terms of all variables) had an AIC value 42 units lower (i.e., better) than a null model, suggesting that it approximated the data well (Anderson 2008). Intermediate distances from minor roads, intermediate values of annual average maximum temperature, and increasing density of desert washes were among the strongest predictors of movement habitat quality. There was also strong evidence for increased habitat quality for movement with increasing amounts of vegetation (i.e., values for NDVI). Using model-averaged regression coefficients of all variables, and the additional penalties for slope and land cover, the final conductance map was compiled in Google Earth Engine (Gorelick et al. 2016) and scaled from 1 - 100 before being exported at 30-m resolution.

We used the conductance layer to model and map omnidirectional connectivity, which predicts potential connectivity across a landscape without regard to the location of core population areas (Pelletier et al. 2014). To derive this model, we created two pairs of parallel nodes (one pixel wide) that extended across the east and west and north and south borders of the tortoise range, respectively. Using Circuitscape software (see McRae et al. 2008, implemented using the Julia numerical programming language), we calculated current flow at each pixel by 'injecting' 1-Amp of current into the node on one side and 'grounding' the node on the other (McRae et al. 2008). Once completed for each pair of nodes, the two models were combined to produce a 'wall-to-wall' map of potential current flow across the tortoises range. Resulting values represent the cumulative current flow (or current density) at each pixel and can be interpreted as relative values of potential connectivity.

Circuitscape outputs provide a current value for each pixel, representing the amount of flow going through that pixel. Care should be taken when interpreting current values. Higher values may imply importance of a pixel in maintaining connectivity, but should not necessarily be interpreted as corridors. A very wide ‘corridor’ of good habitat may have lower current values than a narrow corridor surrounded by poor habitat. This is because in the wide corridor, there are more potential pathways for current to travel at similar costs, resulting in the dispersal of current throughout the corridor.