Remotely sensed environmental measurements detect decoupled processes driving population dynamics at contrasting scales

Abstract The increasing availability of satellite imagery has supported a rapid expansion in forward‐looking studies seeking to track and predict how climate change will influence wild population dynamics. However, these data can also be used in retrospect to provide additional context for historical data in the absence of contemporaneous environmental measurements. We used 167 Landsat‐5 Thematic Mapper (TM) images spanning 13 years to identify environmental drivers of fitness and population size in a well‐characterized population of banner‐tailed kangaroo rats (Dipodomys spectabilis) in the southwestern United States. We found evidence of two decoupled processes that may be driving population dynamics in opposing directions over distinct time frames. Specifically, increasing mean surface temperature corresponded to increased individual fitness, where fitness is defined as the number of offspring produced by a single individual. This result contrasts with our findings for population size, where increasing surface temperature led to decreased numbers of active mounds. These relationships between surface temperature and (i) individual fitness and (ii) population size would not have been identified in the absence of remotely sensed data, indicating that such information can be used to test existing hypotheses and generate new ecological predictions regarding fitness at multiple spatial scales and degrees of sampling effort. To our knowledge, this study is the first to directly link remotely sensed environmental data to individual fitness in a nearly exhaustively sampled population, opening a new avenue for incorporating remote sensing data into eco‐evolutionary studies.


| INTRODUC TI ON
Understanding the environmental drivers of population stability and fluctuations is critical for effective natural resource management.
However, developing this understanding can require information about ecosystems and land cover at scales and sampling frequencies that are impractical to collect from field efforts alone. Beginning with the launch of the Landsat 1 satellite in July 1972, the National Aeronautics and Space Administration/U.S. Geological Survey Landsat Program has consistently provided medium spatial resolution satellite imagery of Earth's surface, with free and open access since 2008 (Wulder et al., 2022). Its data products have contributed to a rapid expansion of interdisciplinary research that relies on ecological knowledge and remote sensing data to describe a variety of patterns, including tracking loss of wetland habitat, detecting shifts in forest canopy composition, and monitoring shifts in phenological cycles (Vogelmann et al., 2016). Much of this work is forwardlooking, seeking to describe how natural systems evolve as climate change progresses and to construct relevant projections, but historical remote sensing data can also be used to add new dimensions to datasets lacking contemporaneous environmental measurements (e.g., Boult et al., 2018;Ndegwa Mundia & Murayama, 2009;Rossi & Leiner, 2022). Herein, we combine remote sensing and weather modeling data with previously collected demographic data to describe environmental factors influencing various components of population dynamics.
Our focal population of banner-tailed kangaroo rats (Dipodomys spectabilis) has been the subject of myriad studies, including investigations of mate choice patterns, genetic adaptation to arid environments, philopatry and dispersal, and many other eco-evolutionary dynamics (Busch et al., 2009;Jones et al., 1988;Marra et al., 2012;. These studies were largely based on detailed demographic and genetic sampling, including precise home mound locations for nearly all individuals in the population and a nearly complete pedigree linking parents and offspring (Waser & Hadfield, 2011;Willoughby et al., 2019). Analysis of this pedigree has previously shown that genetic variables, including degree of individual inbreeding or relatedness between mates, explain a portion of individual fitness, but individual birth year (i.e., non-genetic or environmental factors) accounted for a relatively larger proportion of variation in individual fitness (Willoughby et al., 2019).
To test which environmental characteristics contribute to these interannual differences in fitness, we used Landsat 5 Thematic Mapper (TM) images to obtain surface temperature data and three other descriptive indices via the Tasseled Cap Transformation (Tasseled Cap brightness, greenness, and wetness) (Kauth & Thomas, 1976). Tasseled Cap values can be used to describe variation in soil moisture content, ground cover type, and plant communities, with previous practical applications including assessing impacts of natural disasters, tracking shoreline changes, and charting the progress of desertification (Mostafiz & Chang, 2018;Shamsuzzoha & Ahamed, 2023;Zanchetta et al., 2016). We used these data alongside modeled precipitation and temperature data to summarize the environment of this population over 13 years. We analyzed these data in conjunction with demographic data at three different scales representing three distinct levels of field sampling effort-(i) individual microhabitat vs. individual fitness, (ii) population-scale macrohabitat vs. population fitness, and (iii) population-scale macrohabitat vs. population size-to test the suitability of remote sensing data for describing the effects that specific environmental variables can have on population dynamics at different scales. Because populations of banner-tailed kangaroo rats have been the subjects of numerous ecological and evolutionary studies over several decades, we were able to compare the patterns observed in our results against inferences drawn from prior field-based studies.
Previous studies of D. spectabilis and other heteromyid rodents have described positive relationships between the amount of habitat openness and survival or population size, perhaps because openness facilitates easier detection of or evasion maneuvers against predators or because higher quality food sources tend to grow in such habitats (Bowers et al., 1987;Germano et al., 2001;Waser & Ayers, 2003). We therefore expected to see a positive relationship between the Tasseled Cap brightness index and individual and population fitness, as brightness can indicate the ratio of open soil to plant cover (Crist & Cicone, 1984). We also expected to see a positive effect of precipitation and the Tasseled Cap wetness index-a measure sensitive to soil and vegetative moisture, but primarily characterizing soil moisture (Crist & Cicone, 1984)-on fitness, as increasing water availability may translate into increased food resources (Brown & Zeng, 1989;Munger et al., 1983). Subsequent increases in these resources may be captured by the Tasseled Cap greenness index, a measure shown to be correlated with leaf area index and vegetation biomass (Crist & Cicone, 1984). Specifically, we expected that higher greenness measures in the rainy seasons preceding breeding would lead to increased fitness, as previous studies have found lagged positive responses in rodent biomass or abundance to increased primary productivity (Ernest et al., 2000;Hernández et al., 2005;Previtali et al., 2009;Schooley et al., 2018). Finally, we anticipated that surface and air temperature measures would be negatively correlated with fitness, as increasing surface temperature corresponds to decreasing survival for D. spectabilis populations in the Chihuahuan Desert (Moses et al., 2012).
Although other studies have drawn important new ecological inferences by linking remotely sensed environmental measurements to approximations or correlates of fitness (where fitness is defined as the number of offspring produced by a single individual), such as apparent survival (Moses et al., 2012;Ward et al., 2018) or clutch size and fledging success (Regos et al., 2022;Riggio et al., 2023), ours is the first to use direct assessments of individual fitness as response variables. Specifically, the identification of parent-offspring pairs via genetic analysis allows for inclusion of adult individuals known to be alive but producing zero offspring within a year and for linking observations of specific individuals across years. Herein, we leverage this extensive demographic dataset to test our ecological predictions and, through these analyses, develop new ecological hypotheses regarding drivers of banner-tailed kangaroo rat population dynamics. Overall, we demonstrate that, in the absence of locally collected environmental data, remote sensing data can be used to draw novel inferences and generate new questions regarding fitness and population dynamics at multiple spatial scales and degrees of sampling effort.

| Study system
The study site is located in the Madrean Archipelago ecoregion, which comprises the Sky Islands-forested mountains interspersed among broad, flat desert scrub, and grasslands. The Chiricahua Mountains lie just to the north and west of the site, which is situated around a volcanic cinder cone surrounded by flatlands approximately 35 km southwest of Portal, AZ (31°36′27″ N, 109°15′48″ W) ( Figure 1a). Annual precipitation patterns typically include a summer rainy season from July to August (which supplies 50% of total annual precipitation) and a second, less intense winter rainy season from December to March (Adams & Comrie, 1997). The study area is primarily desert grassland, with rare to occasional half-shrubs and forbs present (Jones et al., 1988;Waser & Ayers, 2003).
Banner-tailed kangaroo rats rely on these plant communities for food, caching seeds in large mounds (1-3 m in diameter) constructed for food storage, reproduction and protection from predators and harsh environmental conditions (Edelman, 2011;Kay & Whitford, 1978). Each mound is typically occupied and defended by a single individual, with the exception of females and their dependent offspring (Schroder, 1979). When the offspring are between 2 and 7 months old, they disperse from their natal mounds to nearby vacant mounds to establish individual territories (Jones, 1984;. Exceptions to typical dispersal patterns may occur in years of high population densities, wherein individuals are more likely to remain in their natal mound to reproduce than to disperse to a new location (Jones et al., 1988;. Mating typically occurs between December and March with females producing 1-2 litters of 1-3 offspring per year (Jones, 1984). Individuals typically live up to 4 years, often producing offspring during the first mating season of their lives.
Twice annually, three traps were placed around each active mound on three consecutive nights, resulting in near-exhaustive population sampling (98% median capture probability for adults; 93% for juveniles; Skvarla et al., 2004). Each captured individual was uniquely marked with ear tags and sex and mound-specific capture location were recorded. It was also noted whether the individual was a juvenile (i.e., born in that year) or an adult. Ear tagging and subsequent recapture allowed individuals to be tracked across the landscape from year to year, and pinna biopsies were taken for genetic characterization. Biopsy samples were genotyped at nine polymorphic loci (Busch et al., 2009; and the resulting data were used alongside trapping records to construct a pedigree for the population (Waser & Hadfield, 2011;Willoughby et al., 2019).
Briefly, Waser and Hadfield (2011) used MasterBayes to build the pedigree, with parental assignment probabilities influenced by F I G U R E 1 (a) Map of the area surrounding the study site, which is situated in Arizona near the New Mexico and Mexico borders. The site is located just southeast of the Chiricahua Mountains. (b) Map of the study site with all mounds included in this study marked with points. The mounds are located on primarily flat areas surrounding a cinder cone.

| Environmental data curation and transformation
We downloaded all available Landsat 5 TM Collection 2 Level 2 images for our study site from 1989 to 2005. Our site was covered by both paths 34 and 35 in row 38 at 30-m spatial resolution. All images were processed and analyzed in R v4.0.3 (R Core Team, 2020). Because of the small size of our site relative to the footprint of a Landsat 5 scene, each image was cropped to a 2100 m × 2750 m extent using the raster package prior to further processing (Hijmans, 2022). We manually reviewed the cropped natural color image for each scene to verify absence of clouds or any other source of error.
For each of the surface reflectance bands, we applied the multiplicative scale factor (0.0000275) and additive offset (−0.2) specified in the Landsat 4-7 Collection 2 Level 2 Science Product Guide (U.S. Geological Survey, 2021, pp. 4-7). We also converted the surface temperature band to Kelvin (and later to degrees Celsius) using a multiplicative scale factor of 0.00341802 and an additive offset of 149. Using the spectralIndices function in the RStoolbox package, we calculated three Tasseled Cap indices for all images: Tasseled Cap brightness, greenness, and wetness (Crist, 1985;Leutner et al., 2019). To check for biased values with respect to path number, we plotted the mean value of each index per scene (i.e., timepoint) over time. Across all years examined, values calculated from path 34 were consistently higher than temporally adjacent values calculated from path 35, leading us to rely exclusively on path 35 scenes for downstream analyses. We also limited the dataset to scenes collected from 1993 to 2005 due to limited observations available in 1989-1992, leaving 167 scenes ( Figure A1; Table A1). All cell values across all years were z-transformed within each Tasseled Cap index.
After observing intra-and interannual patterns for these four variables, we calculated pairwise Pearson correlation coefficients using the cor function in R.
To link the remote sensing data to specific kangaroo rat mounds, we used GPS coordinates recorded for 188 mounds to assign them to corresponding cells in the raster. For 26 mounds, no GPS coordinates were available, but all mounds had been mapped during the original surveys using a custom coordinate system (units in meters) covering the study site (i.e., the position for each mound was recorded against a single reference point). Using the known coordinates for 188 mounds, we overlaid the meter-based locations for all mounds onto the raster and manually assigned the mounds lacking GPS coordinates to cells in the raster. In total, we assigned 214 mounds to raster cells (Figure 1b).
We also obtained precipitation totals and minimum, mean, and maximum temperatures from the Parameter-elevation Regressions on Independent Slopes Model (PRISM) at 4-km resolution (PRISM Climate Group, Oregon State University, 2009). The PRISM model incorporates a digital elevation model and other spatial datasets to calculate gridded estimates of multiple climatic parameters, while accounting for the effects of terrain on precipitation (Daly et al., 1997(Daly et al., , 2008. We used these estimated daily precipitation totals (mm) and minimum, mean, and maximum temperatures (°C) to calculate annual and seasonal means for the population-level analysis, as a single value for each PRISM variable was available for the entire population (see below).

| Individual fitness
Using the parent-offspring assignments generated by Waser and Hadfield (2011), we determined how many offspring each female produced in each year (n = 476 females) and, for females producing at least one offspring, how many of those offspring survived to age one (i.e., reproductive age; n = 282 females). We used the capture data to assign each female to a primary mound location within each year. For each female, we summarized remote sensing values by considering the cell containing her mound location and the eight adjacent cells.
Given that each cell is 30 m across, the maximum distance from the center of an individual's home range in the raster to the edge is 63 m.
Most banner-tailed kangaroo rats disperse <50 m over their lifetime (i.e., the distance between their natal and home mounds is <50 m), meaning that their raster-defined home range likely contains both their natal and reproductive environments (Skvarla et al., 2004). For each year, we calculated mean index and surface temperature values in three ways: (i) season-equalized 12-month (i.e., annual) average, wherein the average index values within each meteorological season were averaged to obtain a single annual value for each index; (ii) summary rainy season averages, calculated for July-August; and (iii) winter rainy season averages, calculated for December-March. We applied a 6-month lag to the environmental data, such that: means for July in year t − 1 through June in year t were used to predict the number of offspring produced in year t; means for July in year t through June in year t + 1 were used to predict the number of offspring produced in year t surviving to year t + 1 (Figure 2).
To check for relationships between individual microhabitat conditions and fitness, we conducted a series of Poisson and negative binomial regressions using the glm.nb function from the MASS package (Venables & Ripley, 2002) in R. The response variable was number of offspring produced with mean annual values for brightness, greenness, wetness, and surface temperature (K) as predictor variables. We used backward stepwise regression, manually removing one predictor variable at a time and examining model coefficients and AIC values until all predictors were significant (p < .05).
We compared the final Poisson and negative binomial regressions using a likelihood ratio test, checked the dispersion parameter for each model, and calculated generalized variance inflation factor (GVIF) values for final models with >1 predictor variable retained using the gvif function in the glmtoolbox R package to quantify the contribution of collinearity on uncertainty in each model (Hernando Vanegas et al., 2022). We repeated this process for the summer and winter rainy season means and for number of surviving offspring. To visualize the effects of predictor variables in models with multiple retained predictor variables, we used the effect_plot function in the jtools package in R, setting the non-focal predictor variable equal to its mean value.
Because some females were sampled in >1 year, we also constructed negative binomial linear mixed models with female identification number (ID) as a random variable for both number of offspring and number of offspring surviving. These models were built using the glmmTMB package in R, and we followed the same backward stepwise regression process as for the models that only included fixed effects (Brooks et al., 2017). For successful mixed models, the final models were compared against the corresponding models lacking random effects with likelihood ratio tests as implemented in the lrtest function in R.

| Population fitness
We used the capture data to determine the number of adult females alive in each year as well as the total number of offspring produced.
From these data, we calculated the average number of offspring produced per female and average number of offspring surviving to age 1 per female. To define the set of cells to be analyzed within each year, we began by identifying active mounds (i.e., mounds where a female was captured) within each year. We defined the total set of active cells as all cells containing an active mound plus the eight cells adjacent to each active cell. For each year, we calculated landscapelevel means for each remote sensing index and the PRISM variables as we did for the individual data (i.e., annually and for the summer and winter rainy seasons) and again applied a 6-month lag ( Figure 2).
We conducted a series of linear regressions to identify relationships between macrohabitat conditions and population-level fitness by testing each combination of a single environmental predictor variable and response variable separately. Two summer rainy season variables (wetness and brightness) were found to be significant predictors for average number of offspring surviving to age 1 (p < .05).
Because neither model met the homoskedasticity assumption, we permuted the y-values and calculated model coefficients 1000 times per model to generate permuted p-values.

| Population size
Again using capture data, we calculated the number of mounds with resident individuals within each year. We assumed that if a mound was occupied, an experienced surveyor of the site could reasonably identify occupied mounds as active based on signs left by residents (e.g., specific characteristic patterns left by banner-tailed kangaroo rat locomotion, recently excavated soil at mound entrances). We applied the same predictor variables and statistical approaches as for the population fitness data, using annual and summer and winter rainy season means from July in year t − 1 through June in year t to predict the number of active mounds in year t. After constructing the initial linear regressions, we were left with a single significant predictor variable (mean annual surface temperature; p < .05) and again permuted the y-values and calculated model coefficients 1000 times to generate p-values.
To confirm that number of active mounds is a reasonable proxy for population size, we constructed linear models to relate number of active mounds to number of adult females and census population size using the capture data. We also tested for relationships between both number of active mounds and number of adult females and fitness rates (number of offspring and number of surviving offspring per female) to determine whether fitness rates could be the result of density-dependent population processes. Finally, to account for the effect of population size in year t − 1 on population size in year F I G U R E 2 Schematic showing temporal alignments between the predictor and response variables tested in the study. For example: annual means used to predict the number of offspring produced in year t were calculated from environmental data collected from July, year t − 1 through June, year t, whereas winter rainy season means were calculated from data collected from December, year t − 1 through March, year t. Although not indicated in this figure, PRISM data were only used as predictor variables for population fitness and number of active mounds (i.e., not for measures of individual fitness). Summer and winter rainy season results are indicated by "S" and "W," respectively. t, we repeated our statistical approach using (i) absolute change in population size from year t − 1 to year t and (ii) proportional change in population size from year t − 1 to year t as response variables (i.e.,  Table A2 for all tested models). Summer rainy season mean brightness and winter rainy season mean wetness and surface temperature also positively affected number of offspring produced (Figure 5b). With respect to the number of offspring surviving to age 1, mean annual surface temperature and mean summer rainy season brightness were positive predictors (Table 1; Figure 5c; see Table A3 for all tested models). For both number of offspring and number of surviving offspring, greenness was not included in any of the final models. For the two individual fitness models with multiple predictor variables retained, GVIF values were close to 1 and therefore did not indicate an outsized contributed of collinearity to model uncertainty (maximum value was 1.05).

| RE SULTS
Including female ID as a random effect did not affect the results for number of offspring with respect to the identity or significance of retained predictor variables when compared to the negative binomial models that only included fixed effects (tested models presented in Table A4). Likelihood ratio tests comparing mixed-effects models to models excluding female ID as a random effect were non-significant. For number of offspring surviving, we could not construct reasonable models that included female ID due to convergence issues. These issues were likely due to the over-representation of individuals with only 1 year of observations (186 of 282 females were only observed in 1 year). For these reasons, we restrict further consideration and discussion to the results of the models that only include fixed effects for both number of offspring and number of offspring surviving.
There were only two statistically significant relationships linking environmental variables and population fitness: summer rainy season brightness ( Table 2) and wetness (Table 3), when averaged across the active landscape, were positively associated with average number of offspring surviving to age 1 per female ( Figure A4).
Despite summer rainy season wetness positively predicting average number of surviving offspring, total precipitation as modeled by PRISM was not correlated with fitness.
With respect to population size, only mean annual surface temperature was a significant predictor variable (Table 4; Figure 5d).
The direction of this relationship was negative, unlike the positive relationships described between surface temperature and individual fitness. In comparing number of active mounds against number of adult females and census population sizes, we found significant and strong statistical relationships ( Figure A5), suggesting that simply surveying the number of active mounds in an area occupied by banner-tailed kangaroo rats would produce a close estimate of population size. Neither number of adult females nor number of active mounds were significantly associated with fitness rates ( Figure A6), suggesting that fitness is not detectably influenced by population density. We also did not observe any significant relationships between environmental variables and absolute or proportional change in population size from year t − 1 to year t.
Although previous years' population sizes certainly influence contemporary population size, we were not able to capture these effects in our analyses.

| DISCUSS ION
For two variables-Tasseled Cap brightness and wetness-our results matched our expectations that were based on previously published relationships between kangaroo rat demographic measures and environmental conditions. In each individual-level model where brightness was retained as a significant predictor and at the population level, brightness positively affected fitness. This is consistent with previous studies that explicitly tested the relationship between habitat openness (i.e., plant density or shrub F I G U R E 3 Mean values of Tasseled Cap indices (a-c) and surface temperature (d) across days of the year. Means were calculated using all cells that were occupied in at least 1 year over the course of the study plus all cells directly adjacent to those occupied cells. Note that the xaxes are offset such that the axis begins with July 1 and ends with June 30. White lines connect dates from July 1 in each year through June 30 in the subsequent year. Vertices for shaded polygons encompass one standard deviation around each mean. The points to the right of the dashed line indicate annual and rainy season mean values within each year. cover) and kangaroo rat abundance (Bowers et al., 1987;Waser & Ayers, 2003). However, these studies primarily focused on the effect of shrub density on kangaroo rat populations, whereas the majority of plants at our study site are grasses. Therefore, brightness as measured in our study may be providing a summary of favorable conditions distinct from what was explicitly tested in previous studies of habitat openness and kangaroo rat abundance.
Mean winter and summer rainy season wetness values were also positively associated with individual and population fitness, respectively. This mirrors results of previous studies that have demonstrated a positive relationship between precipitation and rodent abundances in dry environments (Cárdenas et al., 2021), although mechanistic links between precipitation and rodent abundances are often complex (Ernest et al., 2000;Thibault et al., 2010;Thibault & Brown, 2008). Positive effects of precipitation on kangaroo rat survival or abundance could be mediated via decreased water stress on individuals or through increased availability of food resources that rely on rainy season precipitation to produce seeds. However, values for wetness and brightness were strongly correlated in our dataset (Pearson's r = .89; Figure 3b,c; Figure A2), making it difficult to definitively interpret changes in either index. This is likely due to the high ratio of bare soil: vegetation cover TA B L E 1 Summaries of the best negative binomial models describing individual fitness, with predictor variable values averaged over the time interval indicated. F I G U R E 4 Schematic summarizing the statistically significant relationships identified between environmental variables and fitness or population size. "S" and "W" indicate summer and winter rainy season results, respectively. The sign in each colored polygon indicates the direction of the relationship (i.e., the only negative relationship identified was between mean annual surface temperature and number of active mounds). Polygon color indicates environmental predictor variable with outline pattern indicating the scale at which variables were tested (i.e., individual and population fitness and population size).
at our site, with little variation across the cells being compared (Crist, 1985). The strength of this correlation does vary across the year (summer rainy season Pearson's r = .57; winter rainy season Pearson's r = .92), indicating that these two indices likely capture distinct soil characteristics, but we cannot explicitly define those characteristics without ground-truthed data.
Across all time intervals and scales, Tasseled Cap greenness was never retained as a significant predictor of fitness or population size. We expected greenness to increase during or immediately following the rainy seasons in each year, and this appears to have been the case for some years but not all (Figure 3). The uninformative nature of this particular index for our study site is likely related to semiarid shrub and grassland characteristics. In such ecosystems, spatial patterns of vegetative land cover are highly heterogeneous with respect to both plant community composition and density (Huenneke et al., 2001) and typically comprise dormant (i.e., non-photosynthetic) vegetative cover for large portions of the year (Browning et al., 2017;Okin, 2010;Yang & Guo, 2014). Regardless of season, areas with sparse vegetative cover may not reach greenness thresholds required for detection of vegetation in satellite data (Peng et al., 2021). In other words, the low density of green vegetation at our study site may not be sufficient to prompt an increase in Tasseled Cap greenness values on a per-cell basis, even when the plant community has reached maximum greenness. Further investigation would require either higher resolution data than is publicly available (e.g., Bankert et al., 2021;Browning et al., 2019) or groundtruthed data to calibrate conversions of spectral data to per-pixel vegetation fractions (Smith et al., 1990) on a temporal scale capable of capturing the often rapid changes in photosynthetic activity observed in desert plants (Reed et al., 1994). Without such information from the focal system, it may not possible to reliably ascertain  Tables 1-4 and Table A2. aspects of shrub or grassland phenology using multispectral data alone (Allnutt et al., 2002).

TA B L E 2
Whereas the Tasseled  winter temperatures (as was found for the number of offspring response variable), and warmer winters correspond to lower thermoregulatory costs for kangaroo rats (Edelman, 2011;Hinds & MacMillen, 1985). These reduced costs could help the kangaroo rats' seed caches to last longer, allowing females to produce greater numbers of litters in a single season. Whereas higher winter temperatures may correspond to greater numbers of offspring produced, higher summer temperatures may lead to higher rates of mortality. Although we did not detect a significant relationship between summer rainy season surface temperature and population size, hotter summers could perhaps decrease plant productivity, leading kangaroo rats to quickly exhaust their seed caches and spend more time gathering food at night, thereby also increasing their risk of predation. Additional environmental data (e.g., accurate measurements of plant community composition, abundance, and phenology) could provide greater context for interpreting the influence of surface temperature on population dynamics, but satellite-measured surface temperature alone may remain a critical and accessible measure of habitat suitability or population dynamics for many species as climate change progresses, including species of conservation concern and pest species (Albright Although four of our final models included surface temperature as a predictor variable, PRISM temperature estimates were never retained as significant predictors of fitness or population size, nor were PRISM precipitation estimates. One possible explanation is that PRISM estimates may not closely approximate the true values for our study site, which covers roughly 6% of a single PRISM grid cell. PRISM models account for elevation and topography, but precipitation is highly spatially variable in the Chihuahuan Desert, even over short distances (Petrie et al., 2014), making it difficult to assess the accuracy of PRISM estimates over very small areas. Additionally, large precipitation events can contribute the majority of annual rainfall in wet years (Petrie et al., 2014), and extreme weather events could influence kangaroo rat fitness or survival more strongly than the seasonal or annual averages (e.g., due to food resource spoilage (Valone et al., 1995)) we included in our analyses. Spatial variability in air temperatures may also impede detection of relationships between the modeled PRISM temperatures, but a more likely explanation is that surface temperature values are simply more representative of the environment kangaroo rats experience than air temperature estimates, further highlighting the utility of remotely sensed surface temperature measurements in this and similar habitats.
For all of the environmental variables we tested, we also checked whether these variables were predictive of either absolute or proportional change in population size from 1 year to the next. Population size in the preceding year certainly influences contemporary population size, but we did not detect any relationships between environmental variables and either measure of change in population size. It may be that surface temperaturethe only variable significantly associated with population size-is also correlated with some unmeasured aspect of the environment that limits population carrying capacity rather than rate of change in population size. We did find that number of active mounds is reliably predictive of population size, as has been previously described for this population over a different set of sampling years (Cross & Waser, 2000).

| CON CLUS IONS
Through our analysis of remote sensing and modeled climate data, we were able to identify potential ecological drivers of fitness and population size. Although most of our tested variables (i.e., the Tasseled Cap indices) will require pairing with groundtruthed data from the site to confirm, our results and conservative

ACK N OWLED G M ENTS
We thank Peter Waser for many constructive comments on this manuscript and for sharing expertise on this study system. This

TA B L E A 1 (Continued)
TA B L E A 2 Summary of all models tested using backward stepwise regression for each regression type (P = Poisson, NB = negative binomial) and time interval.

TA B L E A 3
Summary of all models tested using backward stepwise regression for each regression type (P = Poisson, NB = negative binomial) and time interval. Note: For all models in the table, the response variable was individual fitness as measured by number of offspring produced that survived to age 1. The final model selected for each time interval is bolded. In all cases, negative binomial was a better fit than Poisson with the same predictor variables, according to a likelihood ratio test for each final model.

TA B L E A 4
Summary of all negative binomial models tested using backward stepwise regression for each time interval with female ID included as a random effect.

F I G U R E A 1
Temporal distribution of retained Landsat 5 scenes (n = 167) across rainy seasons and meteorological seasons used for equalizing means within each year. Note that the y-axis indicates offspring year or the year in which a cohort of offspring was produced. Therefore, the environmental data used to predict the number of offspring produced in year t covers July 1 in year t − 1 through June 30 in year t (i.e., the first date included in this plot is July 1, 1992, and the last date included is June 30, 2005).

F I G U R E A 2
Correlations among remote sensing variables, with Pearson's r presented above the diagonals. (a) Points represent 10,000 cells sampled randomly across all time points (i.e., scenes) and all cells active in at least 1 year (n = 408 cells). (b) Points represent mean value of active cells within each time point (n = 167 scenes). Whereas strong correlation was noted between Tasseled Cap brightness and wetness, the relationships between greenness and these two indices conform to the classic "Tasseled Cap" shape.

F I G U R E A 3
Significant positive relationships between remote sensing measures and individual fitness (a-c, number of offspring; d, number of offspring surviving to age 1). For panels a and c, the effects of each predictor variable were calculated and are presented by setting the non-focal predictor variable in each negative binomial model equal to its mean value. For all panels, shaded polygons represent 95% confidence intervals. Statistical results for models are presented in Tables 1, A2, and A3.

F I G U R E A 4
Variables identified as significant predictors of population fitness, specifically the average number of offspring surviving to age 1 per female: mean summer rainy season (a) brightness and (b) wetness. Shaded polygons indicate 95% confidence intervals calculated from the unpermuted linear model. p-Values were calculated from 1000 permutations. Model results for brightness and wetness are presented in Tables 2 and 3, respectively.

F I G U R E A 6
No statistically significant relationships were found between (a,b) number of adult females or (c,d) number of active mounds and average number of offspring per female (a,c) or average number of offspring surviving to age 1 per female (b,d). These patterns suggest a lack of density-dependent influences on individual fitness for the years included in our study.