Mapping the potential distribution of the invasive red shiner , Cyprinella lutrensis ( Teleostei : Cyprinidae ) across waterways of the conterminous United States

Predicting the future spread of non-native aquatic species continues to be a high priority for natural resource managers striving to maintain biodiversity and ecosystem function. Modeling the potential distributions of alien aquatic species through spatially explicit mapping is an increasingly important tool for risk assessment and prediction. Habitat modeling also facilitates the identification of key environmental variables influencing species distributions. We modeled the potential distribution of an aggressive invasive minnow, the red shiner (Cyprinella lutrensis), in waterways of the conterminous United States using maximum entropy (Maxent). We used inventory records from the USGS Nonindigenous Aquatic Species Database, native records for C. lutrensis from museum collections, and a geographic information system of 20 raster climatic and environmental variables to produce a map of potential red shiner habitat. Summer climatic variables were the most important environmental predictors of C. lutrensis distribution, which was consistent with the high temperature tolerance of this species. Results from this study provide insights into the locations and environmental conditions in the US that are susceptible to red shiner invasion.


Introduction
Alien invasive fishes have contributed to the global decline in stream fauna through predation, competition, and hybridization (Miller et al. 1989;Allendorf et al. 2001).The number and rate of new introductions have increased dramatically since the 1950s through the propagation and transportation of species around the world (Fuller et al. 1999).Predicting the future spread of non-native aquatic species continues to be a high priority for natural resource managers striving to maintain biodiversity and ecosystem function.Invasive species can have myriad impacts on native aquatic community structure including reducing native species diversity and abundance (Gordon 1998;Wilcove et al. 1998Wilcove et al. , 2000)), modifying hydrology (Ricciardi and MacIsaac 2000) and nutrient cycling (Simon and Townsend 2003;Strayer 2010), and altering food web dynamics (Baxter et al. 2004;van Reil et al. 2006).
Forecasting the potential geographic distribution of aggressive invaders is a crucial component of habitat conservation.Species distribution modeling (SDM) offers a method for generating spatially explicit information to prioritize invasive species management efforts by identifying habitats that are likely to be invaded.Recent advances in geographic information systems (GIS), data mining methods, and data availability have lead to the proliferation of a wide array of SDM approaches.
Maximum entropy (Maxent) is a high performing SDM method that uses species occurrence and environmental data for predicting potential species distributions (Phillips et al. 2006;Elith et al. 2006).Maxent is a machinelearning algorithm that compares presence locations to environmental variables at those locations and then across the study region to generate predictions of species distributions in un-sampled locations.It has performed well for mapping invasive species (Ward 2007;Stohlgren et al. 2010;Jarnevich and Reynolds 2011), although Maxent has not been used extensively for predicting the potential habitat of aquatic invaders (but see Kumar et al. 2008;Oliviera et al. 2010;Feria and Faulkes 2011;Poulos et al. 2012).
We used Maxent software to model the potential distribution of a widely distributed invasive minnow, the red shiner (Cyprinella lutrensis Baird and Girard, 1853).This species is a hardy cyprinid that thrives in a wide range of habitats (Marsh- Matthews and Matthews 2000).Red shiner invasion outside its native range has caused major shifts in fish assemblage structure through competition and extensive hybridization with native congeners (Hubbs and Strawn 1956;Heckman et al. 1987;Karp and Tyus 1990).We chose to model potential red shiner distribution because of the threat it poses to native fish diversity and ecosystem structure.
The red shiner is native to the Great Plains, American southwest, and Mexico in tributaries of the middle and lower Mississippi River basin, and Gulf of Mexico drainages westward to the Rio Grande, including several endorheic basins in Mexico (DFC 2010).Red shiners have been widely introduced outside their native range primarily through bait bucket (Hubbs and Lagler 1964;Jennings and Saiki 1990;Walters et al. 2008) and aquarium releases (Moore et al. 1976;Jenkins and Burkhead 1994).Initial introduction is often followed by the species' rapid population growth, dispersal, and aggressive colonization (Hubbs and Lagler 1958;Minckley and Deacon 1968;Minckley 1973).Where introduced, red shiner can dilute the gene pools of native Cyprinella via hybridization (Mettee et al. 1996;Fuller et al. 1999), and it has also affected the distribution and abundance of native fishes including spikedace (Meda fulgida Girard, 1856), woundfin (Plagopterus argentissimus Cope, 1874), and Virgin River chub (Gila seminude Cope and Yarrow, 1875) (Moyle 1976;Deacon 1988) through various mechanisms including larval predation and direct competition for habitat use.

Species biology
Red shiners are habitat generalists, primarily occurring in creeks and small rivers.They are tolerant of harsh environmental conditions and degraded habitats, including low or intermittent flows, excessive turbidity and sedimentation, and natural physiochemical extremes, which are conditions eschewed by many native minnows (Cross 1967;Sublette 1975;Matthews and Hill 1979;Baltz and Moyle 1993;Douglas et al. 1994).Red shiners are generally uncommon or absent from upland, clear water streams having moderate or high species richness (Matthews and Hill 1979;Matthews 1985;Yu and Peters 2002).This species can tolerate extreme thermal shock ranging from -21 to 10°C, as well dissolved oxygen as low as 1.6 ppm (Matthews and Hill 1977), and it has been observed in hot springs with temperatures as high as 39.5°C (Brues 1928).

Presence records
We compiled spatial red shiner occurrence data from within its native range using records from the Sam Noble Museum of Natural History in Oklahoma, the Kansas Aquatic Gap Database, the Fishes of Texas Project, the Texas Natural History Collection, the Illinois Natural History Survey Fish Collection, and the Tulane University Museum of Natural History (1000 native records total).Records from within the species' invaded range (133 non-native records) were taken from the Nonindigenous Aquatic Species (NAS) database (http://nas.er.usgs.gov)(Figure 1).We included both native and non-native records in our modeling effort because it encompassed the most comprehensive estimation of red shiner's ecological niche.Ibañez et al. (2009) highlighted the utility of this approach for modeling the potential distribution of alien invasive plants and Wolmarans et al. (2010) demonstrated that modeling invasive species distributions using records from a species' native and invaded range did not significantly affect model performance or result in overfitting.

Environmental data
We derived a set of 20 raster-based climatic, topographic, and spectral habitat predictor variables from a range of raster data sources (Table 1) in order to explore the environmental influences on species distribution patterns across the continental United States.We chose these candidate environmental predictors based on their relevance to riparian systems and their biological importance to red shiner.For example, both summer temperature and summer  We also resampled each grid to 1 km from its native spatial resolution, which ranged from 30 m to 1 km using ArcMap v9.3 (ESRI 2008).We chose this resolution based on our intent to model national-scale species distributions and because it was the coarsest native resolution of any single dataset.
The entire dataset of raster predictor variables was reduced through pairwise evaluation to reduce multi-collinearity among the predictors as suggested by Elith et al. (2010).We followed methods outlined by Stohlgren et al. (2010), which used the correlation matrix as a means of identifying highly correlated pairs of habitat predictors (r > 0.7).For correlated pairs, we removed the variable that captured less information.For example, if January minimum temperature and mean minimum temperature were highly correlated, we kept mean minimum temperature since it captured a longer record of winter temperature as a whole.

Species distribution modeling
Maxent uses a deterministic algorithm that finds the optimal probability distribution (potential distribution) of a species across a study area based on a set of environmental constraints.Maxent determines the best potential distribution by selecting the most uniform distribution subject to the constraint that each environmental variable in the modeled distribution matches its empirical average over the known distributional data (i.e.presence data).Maxent is sensitive to sampling biases in clustered or disparate datasets such as the one we use in this study.We addressed this issue by limiting the spatial extent from which Maxent could select background points to locations within 50 kilometers of an existing red shiner record (sensu Jarnevich and Reynolds 2001).
We ran each model by randomly dividing occurrence data into training and testing datasets (70% and 30% of localities, respectively).We produced maps for the conterminous US averaged across 25 model runs with random subsetting of the data at each iteration, including a map of suitable habitat and a map of the standard deviations of the 25 runs following Jarnevich and Reynolds (2011).We also displayed the model prediction results across US hydrologic units (USGS 8-digit HUC units) because of their value to managers for assessing red shiner invasion potential following DeVaney et al. (2009).The potential distribution of red shiner in each HUC unit was displayed using a suitability threshold to derive projected presence-absence distributions from the logistic outputs.The threshold indicating maximum training sensitivity plus specificity is considered a robust approach (Lui et al. 2005), so we used it to conduct the conversion into presence-absence predictions.
Maxent evaluates model performance by calculating the area under the curve (AUC) of the receiver operating characteristic plot.The AUC is a threshold-independent measure of model performance that ranges from 0 to 1. Values > 0.9 indicate high accuracy, values of 0.7-0.9indicate good accuracy, and values below 0.7 indicate low accuracy (Swets 1988).Average AUC values for the 25 runs were reported.

Model performance
The average testing AUC value across the 25 iterations of the Maxent model was 0.86 (SD = 0.009).August maximum temperature, summer precipitation, and summer heat were the most important predictors of potential red shiner habitat (Table 2).Baseflow, mean minimum temperature, and down slope elevation change were other minor environmental predictor variables.
Maxent predicted that red shiner was capable of spreading well outside its native range (Figure 2).In the West, red shiner could spread throughout all of the major California river systems, the Snake River in Idaho, the Gila, Snake and Santa Cruz rivers in Arizona, and in Nevada it could invade the Carson and Humboldt rivers, and tributaries of the Colorado River.Our model predicted that this species could colonize many rivers in the southeastern US including all of the major rivers in Mississippi and Alabama, the headwaters of the major river systems in Georgia (Altamaja, Oconee, Ocmulgee, and Flint rivers) and South Carolina (Savannah, Saluda, Brood, and Catawba Rivers), and the Peedee, Haw, and Tar rivers in North Carolina.The   model also predicted that red shiner could spread eastward from its native range into western Kentucky and Tennessee.The environmental response curves for red shiner were consistent with this species' biological tolerances and habitat preferences (Figure 3).Rivers with high August maximum temperatures, annual minimum temperatures, and summer precipitation had a high red shiner potential distribution.Locations with high baseflow were also more likely to be within the potential distribution of red shiner.

Discussion
The wide potential distribution of red shiner across the United States demonstrates its adaptation as a site generalist, which facilitates its success in newly invaded habitats.This species is the most thermotolerant minnow in North America (Brues 1928;Matthews and Hill 1979), and our maps suggest that red shiner has the potential to spread to other hot environments in the United States.The predicted habitat is consistent with the wide-ranging habitat associations of red shiner in its current native and invaded ranges (Marsh- Matthews and Matthews 2000).Sites with mean minimum temperatures above freezing, high mean maximum summer air temperatures, and a high summer heat index (August temperature/summer precipitation) are potential sites for invasion by red shiner.
The potential spread of this species both eastward and westward beyond its native and currently invaded ranges could threaten the stability of native US minnow populations with similar habitat requirements because of red shiner's ability to outcompete (Greger and Deacon 1988) and hybridize with natives (Burr and Page 1986;Larimore and Bayley 1996).Overlaps in the potential distribution of red shiner and native minnow species richness occur predominantly in the western United States, with the areas of highest minnow diversity and red shiner habitat suitability occurring in Arizona, New Mexico, and southern California (NatureServe 2004).This suggests that cyprinid congeners in these areas may be the most heavily impacted by red shiner spread.Walters et al. (2008) demonstrated that red shiner success can be facilitated through introgression in locations where congeners are present.Red shiner establishes first in locations with congeners, and then its subsequent expansion is driven primarily by hybrid minnows into new habitats.Limiting new introductions by prohibiting red shiner sales as aquarium and bait fish in these regions could lower the risk of red shiner introduction and hybridization with native species.
Red shiner expansion could also have largescale impacts on the abundance and distribution of other native fishes because of its negative influences on native larval fish survival (Ruppert et al. 1993;Douglas et al. 1994;Gido et al. 1999;Marsh-Matthews and Matthews 2000) and habitat use (Douglas et al. 1994).Red shiner occupies nursery habitats of young native fishes including the Red River pupfish (Cyprinodon rubrofluviatilis Fowler, 1916), Colorado pikeminnow (Ptychocheilus lucius Girard, 1856), spikedace (Meda fulgida Girard, 1856), and razorback sucker (Xyrauchen texanus Abbott, 1861), most of which are endangered.It is a well-known consumer of larval fish and it is an opportunistic drift feeder (Sublette 1975).The displacement of native cyprinids by red shiner is also well documented.For example, Douglas et al. (1994) demonstrated that biotic interactions between spikedace and red shiner involved interference competition for space, and that spikedace were displaced to less favorable habitats in the presence of this invader.Mooney and Cleland (2001) suggested that such niche displacement of natives by exotic fishes can have major evolutionary consequences on native populations and in some cases invasive competitiveness can lead to native fish extinction (Ricciardi and Rasmussen 1998;Ricciardi et al. 1998).
Potential distribution maps like the red shiner map produced in this study can be used by resource managers for rapid response and early mitigation of non-native fish invasion.Prohibiting the sale of red shiner as bait or aquarium fish in highly susceptible areas is one potentially effective method for reducing the risk of new introductions to sensitive habitats (Keller and Lodge 2007) as is limiting urban sprawl as a means of minimizing the availability of disturbed habitat since this species does well in such locations.While commerce remains a top policy priority for lawmakers (NISC 2008), restricting the movement of known aggressive invaders in highly vulnerable regions could protect uninvaded habitats from experiencing major shifts in ecosystem structure as a result of new species introductions.

Figure 1 .
Figure 1.Map of Red Shiner presence locations used in the Maxent model.The native distribution of Red Shiner (NatureServe 2010) is shown in dark gray.

Figure 2 .
Figure 2. Maxent results for red shiner including A) the habitat prediction map, and B) the standard deviation among the 25 model iterations using different subsets of point data to test for model sensitivity to presence locations, and C) the presence-absence habitat prediction map for each hydrologic unit (USGS 6-digit HUC) in the United States.The threshold for conversion to binary predictions was derived using the maximum sensitivity plus specificity criterion.The native distribution of C. lutrensis is shown in gray.

Figure 3 .
Figure 3. Environmental response curves for variables contributing greater than 5% to the red shiner model.Red lines indicate mean values for the 25 iterations of the Model.Blue shading indicates the range of environmental values of the 25 iterations of the model.

Table 1 .
Environmental Daymet: http://www.daymet.orgprecipitation were included in the analysis because this species is distributed across some of the hottest and driest parts of the United States.Grids were clipped to the extent of the USGS hydrography dataset to avoid modeling fish distributions outside riparian areas (USGS 2003).
predictor variables evaluated for inclusion in the red shiner species distribution models, their native spatial resolution, and data sources.Variables marked with asterisks were those that were included in the final Maxent model.

Table 2 .
Relative contribution of the environmental variables to the Maxent model.Percent Contribution reports the gain of the model by including a particular variable at each step of the Maxent algorithm The permutation importance reports the contribution for each variable to the final Maxent model which is determined by randomly permuting the values of that variable among the training points (both presence and background) and measuring the resulting decrease in training AUC.The percent contribution and permutation values are normalized to percentages.