AIS Smart Prevention Tool 2.0

About the AIS Smart Prevention Tool 2.0

Wisconsin's 15,000 lakes and 45,000 miles of streams are among the state's most valuable natural resources. The abundance, diversity, and quality of Wisconsin’s aquatic resources provide the cornerstone of the state’s multi-billion travel and tourism industry, in addition to a wide range of ecosystem services, recreational opportunities, and aesthetic benefits. Unfortunately, there is an ever-expanding threat to our aquatic resources. Nuisance exotic species have already taken over the Great Lakes, causing major ecological and economic damage. Increasingly, they are spreading to inland lakes and streams by hitchhiking on recreational boats, and spreading through interconnected waterways, rivers, and canals. What does the arrival of these new nuisance species mean for our inland waters? And more importantly, what can we do to stop their spread and reduce their impacts?

In response to these questions, researchers at the Center for Limnology, University of Wisconsin – Madison have been conducting studies that are relevant to understanding and managing aquatic invasive species in Wisconsin. The central theme of this work can be summarized as ‘smart prevention.’ Because invasive species typically cannot be eradicated once they establish, preventing their spread is generally the best management option. But with approximately 15,000 lakes in Wisconsin, how and where should prevention efforts be focused? Our research has aimed at identifying the lakes that are most vulnerable to invasive species: those where invasive species are likely to be introduced, survive, and have negative impacts (Fig. 1). With such knowledge in hand, prevention, enforcement, and monitoring efforts can be directed to these vulnerable sites.

In addition to the vulnerability research described above, we have addressed a wide range of other questions relating to the spread, impact, and management of aquatic invasive species (see species accounts below). For example, we have found that impounded ecosystems are much more likely to be invaded than their non-impounded counterparts.

The goal of this website is to help make the results of recent research on invasive species spread, impact, and management conducted at the UW-Madison Center for Limnology available to lake managers, residents, and concerned citizens. Development of the original AIS Smart Prevention tool was led by Mona Papes with contributions from Jeff Maxted and hosted by the UW-Madison Applied Population Laboratory. Version 2.0 was launched in March of 2019 and includes the original models plus two additional species, spiny water flea and Eurasian watermilfoil. We hope that our efforts to communicate the result of this research to managers, stakeholders, and concerned citizens are helpful in some way in the ongoing battle against invasive species in Wisconsin, and elsewhere.

For support of our efforts to study and communicate the impacts of aquatic invasive species in our waters, we acknowledge the support of the Wisconsin Department of Natural Resources, the National Science Foundation (the North-Temperate Lakes Long-Term Ecological Research Site and the Bioeconomics of AIS project), and the Baldwin Wisconsin Idea Endowment.

The maps of lake suitability to invasive species presented here are based on statistical models and have not been tested in the field. Please familiarize yourself with the methods that were used to generate these predictions. By using this tool, you agree to use the information presented here only in addition to other information about invasive species, and NOT as maps of ACTUAL distribution of invasive species.

Zebra Mussel (Dreissena polymorpha)

Species Information

Zebra mussels were first discovered in the Great Lakes in 1988 and quickly spread throughout the commercial shipping waterways of the Mississippi River and Great Lakes drainages. Since then, recreational boaters and anglers have spread zebra mussel to many inland lakes in the upper Midwest, and most recently to the western U.S.

Though zebra mussels have spread to approximately 80 lakes in Wisconsin, in fact, many lakes are incapable of supporting this species. Which lakes provide suitable habitat? And which ones are incapable of supporting zebra mussels? We recently conducted an assessment of zebra mussel suitability for ~ 13,000 Wisconsin lakes. Lakes were classified as suitable, borderline, or unsuitable based on measured or estimated dissolved calcium concentrations (Papes et al. 2011). 14% of lakes were classified as suitable: suitable lakes were concentrated in the southern portion of the state, though some suitable lakes were found in northern Wisconsin as well.

Zebra mussel invasions are cause for serious concern. They have caused millions of dollars of economic losses, declines of native species, and have dramatically changed aquatic ecosystems. Zebra mussels efficiently remove algae from the water column, while simultaneously excreting large amounts of organic material and nutrients, which promotes bottom algae growth. Beyond this, it is unclear how many key aspects of lake ecosystems change following zebra mussel invasion. To address this question, researchers recently summarized and analyzed all of the available data on zebra mussel impacts (Higgins and Vander Zanden 2010). The study found that zebra mussels cause a 40-80% decline in the entire planktonic (open water) food chain and a 2-10x increase in the entire benthic (bottom) food chain. In other words, the effects are not confined to the bottom of the food chain, but rather they reverberate through the entire food web.

Model Information

Assessment of lake suitability to zebra mussel invasion based on calcium concentrations

Zebra mussel lake suitability analysis was carried out in 2009 by Monica Papes and Jake Vander Zanden, Center for Limnology, UW – Madison.

Zebra mussel lake suitability represent thresholds of calcium concentration reflecting previous studies of zebra mussel physiological requirements. Linear regression analysis was used to predict calcium concentration values from available conductivity measurements for over 11,500 lakes in Wisconsin and Upper Michigan Peninsula.

We reviewed the scientific literature addressing physiological constraints of zebra mussel and found seven studies that measured the effects of water calcium concentrations (mg/l, hereafter Ca) on zebra mussel development and growth (Jones and Ricciardi 2005, Allen and Ramcharan 2001, Hincks and Mackie 1997, Mellina and Rasmussen 1994, Ramcharan et al. 1992, Strayer 1991, Sprung 1987).

A 95% confidence interval (CI) was applied to the distribution of the calcium concentration thresholds identified in these studies to establish a lower limit for zebra mussel survival. The value obtained for this lower threshold was 10 mg/l (mean mg/l Ca – 95% CI). Similarly, we calculated an upper threshold of 21 mg/l (mean mg/l Ca + 95% CI).

These thresholds delimited three classes of lakes:

• not suitable to zebra mussel invasion (mg/l Ca < 10),
• borderline suitable (10<mg/l Ca < 21),
• suitable (mg/l Ca > 21).

Since we were able to gather Ca values from only about 1000 lakes in Wisconsin, we developed a linear regression function between Ca and conductivity based on two sources of data (EPA-Eastern Lake Survey, and CFL Landscape Position Project) that had both Ca and conductivity values.

The linear regression function obtained as follow:

lnCa = -3.11018 + 1.2034 x lnConductivity

The linear regression function was then used to predict mg/l Ca for about 11,000 lakes in Wisconsin and the Upper Peninsula of Michigan for which we had conductivity data. These data were gathered from a variety of past limnological studies, but mostly from the Surface Water Resources Inventory carried out in Wisconsin between 1960 and 1980 by the DNR and other state agencies.

The classification was done using ln-transformed thresholds:

• not suitable: lnCa< 2.302;
• borderline suitable: 2.302< lnCa< 3.044;
• suitable: lnCa>3.044).

Our final dataset thus included over 11,500 lakes with Ca values (measured or predicted) from Wisconsin and the Upper Peninsula of Michigan. We used it in Geographic Information Systems (GIS) to assign the lakes to three different suitability classes to zebra mussel invasion, as described above.

About 5500 lakes with surface area < 10 ha and 1300 lakes > 10 ha were omitted from this study due to lack of conductivity or Ca measurements.

Rainbow Smelt (Osmerus mordax)

Species Information

Rainbow smelt are native to coastal regions of eastern North America. They spread through the upper Great Lakes during the 1920s and 30s, and have since spread to inland waters throughout the region. Currently, they inhabit at least 24 inland Wisconsin lakes.

Rainbow smelt often become very abundant, and feed on the young of other fish species. As a result, they reduce native fishes such as lake herring, lake whitefish, yellow perch and walleye. Among their many documented effects, their invasion is responsible for the decline of walleye recruitment in Wisconsin lakes (Mercado-Silva et al. 2007).

Rainbow smelt tend to inhabit large, deep, unproductive lakes. Mercado-Silva et al. (2006) identified > 500 lakes in Wisconsin to be suitable for smelt invasion. Thus, there remains much potential for further spread of this invasive fish.

Model Information

Assessment of lake suitability to rainbow smelt invasion based on information from native range in Maine

Methods as described in Mercado-Silva, N., J.D. Olden, J.T. Maxted, T.R. Hrabik, M.J. Vander Zanden. 2006. Forecasting the spread of invasive rainbow smelt in the Laurentian Great Lakes region of North America. Conservation Biology 20: 1740-1749.

Our objectives were to develop a model for rainbow smelt presence and absence based on their distribution in their native range in coastal regions of Maine (U.S.A.) and to apply model predictions in other geographic areas to forecast future invasion potential.

Limnological data and rainbow smelt distribution data for Maine lakes were obtained from databases managed by the Public Educational Access to Environmental Information (PEARL) in Maine. We estimated that 354 Maine lakes are within the native range of rainbow smelt. Lake morphometry and physical-chemistry variables were obtained for 819 lakes from data sets in PEARL and from geographic information provided by Maine Department of Environmental Protection.

We compiled information for Wisconsin inland lakes from various sources. Maximum lake depth, area, and Secchi-disk readings were obtained from an extensive lake data set at the WDNR. pH information was obtained from the STORET database of the U.S. Environmental Protection Agency (STORET-EPA), and shoreline perimeter was calculated with a GIS (ESRI 2002) from data obtained from the WDNR. We obtained information for 5188 lakes in Wisconsin, but data availability was uneven among lakes (lake area = 5188 lakes, Secchi-disk readings = 3891, pH = 1190, maximum depth = 5142, shoreline perimeter = 1190).

We used classification trees (Breiman et al. 1984; Salford Systems 2002) to model rainbow smelt presence and absence within their native range in Maine. This methodology uses a recursive partitioning algorithm to repeatedly partition the data set according to the explanatory variables into a nested series of mutually exclusive groups, each as homogeneous as possible with respect to the presence or absence of rainbow smelt (see De’ath & Fabricius 2000).

For Wisconsin lakes, the model had a correct classification rate of 89%. The model identified 553 new lakes with the potential to be invaded by smelt. In Wisconsin and Ontario, <5% of lakes capable of supporting smelt currently do, indicating vast potential for future spread in both areas.

Round Goby (Neogobius melanostomus)

Species Information

Round goby is a small benthic fish native to the Black and Caspian Seas. It first appeared in the Great Lakes in 1990, and has since spread to all five Great Lakes. Round goby are currently spreading into coastal tributary streams in Wisconsin and elsewhere (Kornis and Vander Zanden 2010). In Wisconsin, round gobies were detected in 26 of 73 streams (36%), and were found more than 10 km upstream of Lake Michigan in nine separate watersheds. We forecasted the future spread of round goby in Wisconsin’s Lake Michigan basin. Up to the first stream barrier, 1369 km of stream are suitable (Fig. 1A). Within the watershed as a whole, 8878 km of stream were suitable (Fig. 1B). These results highlight the role of the Great Lakes as a springboard for invasive species to disperse into inland ecosystems (Kornis and Vander Zanden 2010). Current research is assessing impacts of round goby in stream ecosystems (Kornis 2013).

Model Information

Assessment of stream suitability to round goby invasion in Lake Michigan tributaries

Based on Kornis, M.S. and M.J. Vander Zanden. 2010. Forecasting the distribution of the invasive round goby (Neogobius melanostomus) in Wisconsin tributaries to Lake Michigan. Canadian Journal of Fisheries and Aquatic Sciences 67(3): 553-562.

We developed a model for round goby presence and absence based on their distribution in Wisconsin (U.S.A.) tributaries to Lake Michigan and applied model predictions to forecast future invasion potential. We were precluded from using native range data because round gobies are historically limited to large lakes and the lower reaches of large rivers.

Round goby distribution data were obtained from a field survey of all Wisconsin tributaries to Lake Michigan (130 sites on 73 streams). For a given stream, sampling began near the mouth and continued upstream at sites at least 1 km apart until no round gobies were detected. If round gobies were not present, we assumed they were absent upstream of that site as well. Two collection methods were used to sample the range of stream sizes in our study. Wade-able sites (<1m in depth) with high water clarity were sampled using a battery-powered backpack electrofishing unit with pulsed DC current. Such sites were electrofished for 20 minutes at a current of 3.5 to 4.0 amps, with focus on cobble or boulder substrate. Sites with low water clarity and/or depths greater than 1m were sampled using an overnight set of 10 wire-mesh Gee-style minnow traps baited with chicken liver. Probability of detection for both methods was assumed to be 1 in areas of high round goby abundance. Probability of detection at sites where round gobies were rare was 0.75 for electrofishing (determined by repeated sampling of sites where only 1 round goby was initially captured). Probability of detection using minnow traps was not calculated but assumed to be at least 0.75 because round gobies were more frequently detected by minnow traps than electrofishing and traps usually captured high numbers of round gobies if they were present.

We compiled habitat information for Wisconsin streams from various sources. Water temperature, conductivity, bankfull width, stream depth, and percent substrate composition were measured locally at each site by the field survey team. Landscape-scale variables (watershed area, percent land use, stream gradient, watershed slope, soil permeability, base-flow index, and mean July air temperature) and the coordinates of dams were obtained from a global information system (GIS) database provided by the US Geological Survey, Great Lakes Aquatic GAP Analysis Project. Landscape-scale variables were calculated on the scale of confluence-bound stream segments.

We used generalized linear mixed models (using the “logit” or logistic regression function) to model round goby presence/absence based on habitat factors in Wisconsin tributaries to Lake Michigan. Such models output the probability of round goby presence for a given site based on the predictor. Models were evaluated using the Bayesian information criterion (BIC) to select among models based on predictive ability. Best subsets analysis was used to extract the 20 best models (based on BIC score) from all possible models (i.e. all possible combinations of predictor variables). To test each of these models on independent data, each model was validated using k-fold cross validation as recommended for presence/absence models by Fielding and Bell (1997). Receiver operating characteristic (ROC) curves were used to determine the probability threshold distinguishing “predicted present” from “predicted absent”; thresholds were selected using the minimized difference threshold (MDT) criterion, which selects the probability value that minimizes the difference between sensitivity and specificity (Jiménez-Valverde and Lobo 2007). We used Cohen’s Kappa, which measures the proportion of agreement between actual and predicted data, as the primary statistic of model performance. This allowed us to evaluate each of the 20 models in our candidate set for their ability to perform on independent data (generated from the k-fold cross validation). This criterion was most appropriate given our goal of forecasting round goby presence/absence at sites yet to be invaded. The best model based on this criterion included watershed area, stream gradient, and watershed slope as predictor variables and was used to generate forecasted range.

For Wisconsin tributaries to Lake Michigan, the model had a correct classification rate of 80%. The model identified 1,369 km of stream downstream of impassable fish barriers (dams) as vulnerable to round goby invasion. Currently, round gobies only occupy 20% (280 km) of these streams, indicating a large potential for continued spread of round gobies in tributaries. The model also identified 7,482 km of streams upstream of impassable fish barriers as vulnerable to round goby invasion. To invade these streams, round gobies would need assistance from humans. This illustrates the vast potential for round gobies to invade inland streams and highlights the importance of educational campaigns and species movement regulations in preventing the spread of round gobies (and other invaders) upstream of dams.

Rusty Crayfish (Orconectes rusticus)

Species Information

Rusty crayfish are native to streams of the Ohio River basin, and have recently spread throughout much of North America. A recent study documented their rapid spread across the state, from 3% (pre-1980) to over 50% (present day) of crayfish occurrence (Olden et al. 2006). This study also found that rusty crayfish tend to displace native crayfish species.

Are there environmental factors that might limit the future distribution of rusty crayfish in Wisconsin? Rusty crayfish do not survive in waters with dissolved Ca2+ concentrations below 2-3 mg/l. Unfortunately, relatively few aquatic systems in Wisconsin fall below this level, indicating little potential to limit rusty crayfish distributions (Vander Zanden and Maxted 2008). Current research is using statistical models to better identify lakes and streams that are vulnerable to rusty crayfish invasion and impacts (Olden and Vander Zanden 2011).

Efforts to control or eradicate aquatic invasive species have achieved limited success, and much research is needed to develop better methods to control invasive species. There is also a need to understand the reversibility of ecosystem changes caused by aquatic invasive species. Researchers at the UW Center for Limnology have been conducting experimental removal of invasive rusty crayfish in Sparkling Lake (Vilas County) since 2001 (Hein et al. 2006; 2007). Removal efforts have produced sharp declines in rusty crayfish populations, and the decline has benefitted from increased bass and walleye abundances due to fishing regulation changes (Hein et al. 2006). Current research is assessing whether the rusty crayfish declines has resulted in the recovery of the ecosystem.

We still know little about the impacts of rusty crayfish on freshwater ecosystems. For example, can native species adapt to the presence of invasive species and find ways to coexist? Recent work on rusty crayfish-virile crayfish interactions indicates that native crayfish do adapt to the presence of invasive rusty crayfish, allowing them to potentially coexist (Hayes et al. 2009). Nevertheless, rusty crayfish have many impacts on aquatic ecosystems, including reducing aquatic invertebrates (McCarthy et al. 2006; Roth et al. 2006). Current work is investigating the implications of reduced invertebrate populations for fish, and the extent to which they disrupt energy flow in aquatic food chains (Nilsson et al. 2012).

Model Information

Forecasting the multi-staged invasion of rusty crayfish (Orconectes rusticus) in freshwater ecosystems

We used a multi-response artificial neural network (MANN) to model present-day (1985-2006) presence/absence of Orconectes rusticus as a function of environmental and human-related factors describing the likelihood of introduction and establishment. Our lake model included 9 variables describing human visitation, lake morphology and water chemistry, and our stream model included 11 variables describing channel characteristics, low flow regime, land use, superficial geology, and regional climate. We developed separate feed-forward MANNs trained by the backpropagation algorithm to model 292 lakes and 546 stream sites. Using the MANNs, we predicted the probability of occurrence for O. rusticus for 4,200 lakes and 35,855 stream segments (representing 90,360 river kms) across the entire state of Wisconsin.

O. rusticus occurrence was highly predictable according to the MANN. The cross-validated neural network correctly predicted O. rusticus in 93.2 % of the lakes and 94.3 % of the stream sites (cross-validated estimates), in addition to exhibiting both high sensitivity and specificity. Model performance was significantly greater than random based on Cohen’s Kappa for both habitat types.

Spiny Water Flea (Bythotrephes longimanus)

Species Information

The spiny water flea (SWF) is a free-swimming predatory crustacean, native to lakes of northern Europe and Asia. This species was discovered in Lake Huron in 1984. It quickly spread through the Great Lakes, and within a decade, had invaded hundreds of inland lakes in the Great Lakes region. SWF were discovered in Gile Flowage (Iron County) in 2003 by researchers at the UW Center for Limnology. Soon after, additional populations were discovered in Stormy Lake (Vilas County) in 2007 and the Madison chain of lakes (Mendota, Monona, Waubesa, Kegonsa) in 2009.

The current spread of SWF in Wisconsin is cause for concern. SWF are a voracious predator upon native zooplankton, and have caused dramatic declines in native zooplankton in Canadian lakes. The native zooplankton play two key roles in lakes: they are food for small fish and they consume phytoplankton and thus help control algal blooms. The long spine of SWF inhibits consumption by smaller fishes, highlighting their potential for disrupting the food web.

Their impact on Wisconsin’s lakes is not yet well understood. Researchers at the UW Center for Limnology are currently monitoring Stormy Lake and the Madison Lakes to understand their effects on water clarity, fishes, native zooplankton, and lake food webs as a whole. The Madison Lakes are among the most well-studied lakes in the world, and this unfortunate recent invasion does provide an opportunity to understand their impacts.

An urgent management challenge is to assess which lakes are vulnerable to spiny water flea invasion. One study found that SWF tend to inhabit large, deep, clear lakes in their native range. Application of those results to Wisconsin indicates that a large number of lakes are capable of supporting this invasive species. Researchers at the UW Center for Limnology are currently developing statistical models that can be used to identify which lakes are vulnerable to SWF (Papes et al. 2011).

Model Information

Assessment of lake suitability to spiny water flea based on modeled lake temperatures

Note: this project is a work in progress and the models presented here may be subject to changes as the modeling improves.

The suitability analysis for spiny water flea (SWF) was carried out in 2018 by Jake Walsh in collaboration with Jake Vander Zanden (UW-Madison), Jordan Read (USGS), and Gretchen Hansen (MNDNR/UMN-TC).

Background. Spiny water flea is a cool water zooplankter and is sensitive to warm summer surface temperatures at the southern edge of its range, which includes Wisconsin (Kerfoot et al. 2011, Walsh et al. 2016). We used a 10,774-lake dataset of daily modeled lake temperature profiles from three time periods (1981-2001, 2040-2060, and 2080-2100) in Minnesota, Michigan, and Wisconsin (Winslow et al. 2017) to estimate lake thermal suitability to SWF under contemporary and future projected lake thermal conditions. Water temperatures were modeled using seven different global climate models, six of which include projections under future climate change scenarios (see Winslow et al. 2017 for details). Notably, all six GCMs are modeled under the most extreme emissions scenarios (RCP 8.5).

MaxEnt model fitting. We modeled lake thermal suitability to SWF using a maximum entropy algorithm, MaxEnt, a commonly used technique for modeling species distributions (Phillips et al. 2006), developing our models using the package “maxnet” in R (Phillips 2017). The maxnet package fits environmental data to occurrence data using lasso regularization of generalized linear models to generate the Maxent probability distribution (Phillips 2017). Maxent is particularly useful for SDM modeling using presence only data (essential here given just 40 confirmed occurrences and 10,744 background lakes) as well as estimating non-linear effects of predictor variables (Phillips et al. 2006).

We selected three variables for model fitting: degree days above 20°C (SWF developmental rates plummet above 20°C; Lehman et al. 1997, Kim and Yan 2010), mean surface temperatures in October (°C; Bythotrephes is most abundant in fall in our study region (Walsh et al. 2018); and log10 maximum observed lake depth (m; since SWF is primarily pelagic, lake depth is the most likely morphological variable related to SWF habitat variability in lakes; e.g., Young and Yan 2008). These variables are weakly correlated: cor(gdd_wtr_20c, lake_depth) = -0.15, cor(gdd_wtr_20c, mean_surf_oct) = 0.51, cor(mean_surf_oct, lake_depth) = 0.63 (using contemporary conditions under the GFDL GCM), but correlation among variables shouldn’t matter as much when using Maxent relative to other niche modeling methods (Elith et al. 2011).

A key feature in MaxEnt modeling is the regularization multiplier, which essentially affects how tightly fitted the model is to the occurrence data. If a model is too tightly fitted to the data, it risks being ungeneralizable. Since we want to project our model into future lake temperature conditions, we increased the default regularization multiplier, in effect “under-fitting” the model. We used a regularization multiplier (“regmult”) of 2 (with the default being 1). Further, MaxEnt fits the models using multiple “feature classes” to allow for nonlinear effects of environmental variables. We allowed the models to use all feature classes (linear, quadratic, product, hinge, threshold), but note that with just 40 occurrences, the maxent algorithm won’t use the product feature classes. However, the final models used nearly identical combinations of linear and hinge features across all six GCMs (Fig. 1).

Fig. 1– Variable effects on raw maxent suitability output are plotted in colors that vary by GCM.

Variable effects. We found that increasing GDD > 20°C had a negative effect on SWF suitability while increasing lake depth had a positive effect on suitability (Fig. 1). Both of these findings are consistent with our current understanding of SWF ecology. The effect of mean surface temperatures in October on suitability was non-linear and U-shaped, possibly reflecting the positive effect of warming in a typically cool season when SWF is most abundant (Fig. 1). Since temperatures in October aren’t expected to warm past SWF thermal limits, any increases in temperature may increase developmental and population growth rates and may offset the negative effects of warmer summer surface temperatures. It is worth noting that this trend toward higher suitability at higher mean surface temperatures in October is driven by 11 of the 40 known occurrences in the dataset and not by a small number of outlying occurrences. These lakes are warm in both summer and fall, but tend to be deeper (e.g., > 9 m maximum observed depth).

Results in entire study regionWe defined suitability as any lake with a suitability score as high as the lowest score in a lake with confirmed SWF presence. Averaging suitability scores across all GCM-produced MaxEnt models, we found that roughly half of the lakes in our study region (5,204 of 10,774 lakes) would be suitable for SWF under contemporary conditions (Table 1). This fraction is expected to decline substantially with future warming (3,666 suitable lakes in 2090), though we note interesting non-linear patterns between 2050 and 2090 scenarios (3,165 lakes in 2050), likely due to the interaction between suitability and mean surface temperatures in October.

Table 1Model predictions (number of predicted presences where a presence is defined as a suitability score at least as high as the score of the least suitable known occurrence) across each GCM and time period. As a reminder: there are 10,774 total lakes in the dataset. Note: the mean is not the calculated mean of the columns, but predictions from suitability scores averaged across all GCMs.

Results for Wisconsin. For the purposes of this online tool, we’re primarily interested in SWF suitability Wisconsin. Of the 4,751 Wisconsin lakes in the dataset, 1,907 are predicted to be suitable under contemporary conditions (40%), 921 are predicted to be suitable under 2050 conditions (19%), and 1,229 are predicted to be suitable under 2090 conditions (26%) (Table 2).

Table 2 – Model predictions for Wisconsin (4,751 lakes in Wisconsin with SWF suitability data).

Metadata for understanding shapefile output

Model threshold for suitability: ≥-10.1526; (this is the lowest raw suitability score for a lake with a known SWF population)

Shapefile column names:

“lat” = latitude of the centroid of the lake shapefile

“long” = longitude of the centroid of the lake shapefile

“suitCont” = raw suitability score for the contemporary time period

“predCont” = binary suitability based on above threshold

“suit2050” = raw suitability score for the late century time period

“pred2050” = binary suitability based on above threshold

“suit2090” = raw suitability score for the late century time period

“pred2090” = binary suitability based on above threshold

“pa” = known SWF occurrences (1 = presence)

“RawSuit” = raw suitability score (redundant w/suitCont)

Eurasian Water Milfoil (Myriophyllum spicatum)

Species Information

M. spicatum is an invasive aquatic plant can be abundant in some waterbodies. The species is native to Europe, Asia and North Africa. The precise date of introduction to the United States is unknown, but likely occurred some time between 1880 and 1940 (Eiswerth et al. 2000). After the initial introduction, EWM spread throughout the continental US and Canada (EDDMapS 2017). It is now present in over 800 Wisconsin waterbodies, having steadily expanded its range northward after being introduced to southern Wisconsin during the 1960s (Trebitz et al. 1993). When abundant in certain waterbodies, EWM forms thick mats that prevent navigation, reduce property values and may affect native species (Boylen et al. 1999, Provencher et al. 2012). In 2011, over half of the total funding for Wisconsin’s Aquatic Invasive Species control grant program was allocated to M. spicatum, amounting to around $1-$2 million dollars annually, making it a high priority for an assessment of invasion vulnerability.

Model Information

Predicted occurrence of Eurasian Watermilfoil in Wisconsin based on physical, geographic and land use factors.

Note: Unlike the other species models displayed in this tool, the Eurasian watermilfoil model measures probability of occurrence, which incorporates dispersal-related risk of introduction in addition to environmental parameters influencing lake suitability. However, for the sake of consistent scheming with the other species in this tool, our map labels “present” / “absent” predictions as “suitable” or “unsuitable.”

The analysis of lake suitability and dispersal-related risk of introduction was conducted in 2017 by Alison Mikulyuk (WDNR) in collaboration with E. R. Kujawa (Lake Champlain Basin Program), C. M. Hein (WDNR), S. Van Egeren (WDNR), and M. J. Vander Zanden (UW-Madison).

Background: We sampled aquatic plant communities on 657 waterbodies from May to October, 2005 – 2016. EWM presence was observed using a rake sampler deployed from a boat at points arranged in a sampling grid that covered the surface of the waterbody (Mikulyuk et al. 2010). We assembled information on predictors with a hypothesized relationship to EWM occurrence for 9,285 Wisconsin lakes over 1 ha in surface area. Variables were related water quality and morphometry as well as dispersal, land use, geology and climate. The inclusion of dispersal and human use factors means that model results should not be interpreted solely as environmental suitability values, but should be taken to also reflect the likelihood of introduction.

Developing the model: We built species distribution models using logistic regression employed in a generalized linear modelling framework. The procedure employs a maximum likelihood optimization algorithm to estimate intercept and slope parameters for an equation that returns the probability that a given lake has been invaded. Model fitting was performed using Firth’s method of bias reduction in R using the function and package ‘brglm’ (Kosmidis 2017). We used a five-fold cross-validation procedure to evaluate model performance, developing the model on a randomly selected 80% of the data, and comparing predicted probabilities for the remaining 20% of data to observed presence/absence (Fielding and Bell 1997). We evaluated cross-validated model performance with a receiver operating characteristic analysis using the function ‘roc’ in package ‘pROC’ (Hosmer et al. 2013, Hussner et al. 2017). We also calculated overall model deviance, the percentage of deviance explained by the model (D2) and Tjur’s coefficient of discrimination, which can be interpreted much in the same way as an R2 value for linear regressions (Caffrey and Monahan 2006). All model performance statistics were averaged across cross-validations.

Results: Logistic regression models predicting M. spicatum presence performed well, with AUC = 0.81. The probability of M. spicatum occurrence was statistically significantly related to water quality, morphometry, climate, dispersal and geology. The probability of M. spicatum occurrence was positively related to road density, surface area, maximum air temperature of the warmest month and lake maximum depth. Occurrence probability was negatively related to %CaO in the watershed, annual temperature range and mean distance from all source populations. Mean cross-validated deviance was 515, Tjur’s coefficient of determination was 0.32, and the amount of deviance accounted for by the model was 22%. M. spicatum has broad environmental tolerances and can survive under a wide range of conditions, but it is not equally likely to be introduced to lakes across the state. Our occurrence model identifies locations where it is likely to be present, predicting that around a quarter of lakes in Wisconsin currently have or are likely to be invaded by M. spicatum.