Forecasting the combined effects of anticipated climate change and agricultural conservation practices on fish recruitment dynamics in Lake Erie

This is an open access article under the terms of the Creat ive Commo ns Attri bution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited. © 2020 The Authors. Freshwater Biology published by John Wiley & Sons Ltd 1Department of Evolution, Ecology, and Organismal Biology, Aquatic Ecology Laboratory, The Ohio State University, Columbus, OH, U.S.A. 2Department of Food, Agriculture, and Biological Engineering, The Ohio State University, Columbus, OH, U.S.A. 3Department of Natural Sciences, Tusculum University, Tusculum, TN, U.S.A. 4Blackland Research and Extension Center, Texas A&M University, College Station, TX, U.S.A. 5United States Department of Agriculture, Agricultural Research Center, Temple, TX, U.S.A. 6United States Geological Survey, Upper Midwest Water Science Center, Middleton, WI, U.S.A. 7The Nature Conservancy, Michigan Field Office, Lansing, MI, U.S.A. 8United States Department of Agriculture, Resource Assessment Division, Natural Resources Conservation Service, Temple, TX, U.S.A.

Highly productive coastal and large-lake ecosystems that receive substantial river-transported runoff from agricultural catchments are typically negatively affected by the combined impacts of climate change and NPS pollution (Buchheister, Bonzek, Gartland, & Latour, 2013). Because these ecosystems (e.g. Chesapeake Bay: Breitburg, 2002;Kemp et al., 2005;Buchheister et al., 2013;northern Gulf of Mexico: de Mutsert, Steenbeek, Lewis, Buszowski, & Cowan, 2016; Lake Erie: Ludsin et al., 2001) often support valuable commercial and recreational fisheries, which can be adversely affected by eutrophication, regulatory agencies and policy-makers have typically sought to improve habitat conditions (e.g. water quality) by altering farming practices in the catchment (Hagy, Boynton, Keefe, & Wood, 2004;Keitzer et al., 2016;Ohio EPA, 2013;Scavia et al., 2014;Wilson et al., 2019). Agricultural conservation practices (ACPs), which are farming practices that reduce runoff, prevent erosion, curb excessive nutrient loading, and mitigate ecosystem degradation, offer one potential means to reduce the negative effects of NPS pollution on fish production while, ideally, not compromising agricultural production (USDA Keitzer et al., 2016;NRCS, 2011). The benefits of such practices are well-documented and have been shown to limit in-stream nutrient and sediment loading, improve water quality, promote diverse stream-fish assemblages, and even reduce nutrient inputs into downstream recipient water bodies (Bosch, Allan, Selegean, & Scavia, 2013;Keitzer et al., 2016;Richards, Baker, & Crumrine, 2009). Even so, how ACP implementation amidst a changing climate would affect fish production in large ecosystems that receive substantial upstream nutrient and sediment loads remains a conspicuous information gap. Such information could help managers design more resilient and adaptive management strategies (Hansen et al., 2017;Lynch et al., 2016;Paukert et al., 2016).
Climate change and nutrient loading can directly and indirectly drive fishery dynamics by affecting the recruitment of individuals to the fishable population (Farmer et al., 2015;Hansen et al., 2017). | 1489 DIPPOLD et aL. and direction of response to increasing temperature is linked to a species' thermal guild, with cool-and cold-water species showing negative responses to warming and warm-water species showing positive ones (Comte et al., 2013;Hansen et al., 2017;Lynch et al., 2016). Similarly, a species' tolerance, or lack thereof, of eutrophic, nutrient-rich waters can affect recruitment positively or negatively through top-down or bottom-up processes (Briland, 2018;Ludsin et al., 2001). Thus, understanding the response of fish recruitment, and ultimately fishery production, to climate change and the implementation of ACPs is a key component to future management and conservation (Pritt, Roseman, & O'Brien, 2014).
Towards this end, we explored how anticipated climate change and the implementation of realistic ACP scenarios might alter the recruitment dynamics of three ecologically and economically important fish populations (walleye Sander vitreus, yellow perch Perca flavescens, and white perch Morone americana) in the highly productive and dynamic west basin of Lake Erie (Figure 1). The recruitment dynamics of these species have been shown to respond to climate-related factors (e.g. winter and spring temperatures: Busch, Scholl, & Hartman, 1975;Clady, 1976;Farmer et al., 2015;Hokanson, 1977;Johnson & Evans, 1990) and factors associated with NPS inputs from rivers (Carreon-Martinez, Wellband, Johnson, Ludsin, & Heath, 2014;Jones, Shuter, Zhao, & Stockwell, 2006;Ludsin et al., 2011;Mion, Stein, & Marschall, 1998;Reichert et al., 2010). We focused on western Lake Erie for several reasons.
We forecasted walleye, yellow perch, and white perch age-0 (juvenile) abundance, which has been shown to be a strong predictor of future recruitment to the fishery at age-2 (Farmer et al., 2015;WTG, 2017), under different combinations of anticipated  climate change and varying levels of ACP implementation in the WLEB catchment. Our specific research goals were to: (1) quantify the independent and combined effects of climate change and ACP implementation on walleye, yellow perch, and white perch recruitment; (2) explore whether climate change and ACP implementation might alter western Lake Erie's fish community by differentially affecting native cool-water fish species (walleye and yellow perch) versus nonnative warm-water ones (white perch); and (3) provide insights to Lake Erie management agencies regarding the potential future of their fisheries. We hypothesised that, in general, climate warming would negatively affect walleye and yellow perch recruitment, owing to their seeming dependence on long winters for successful reproductive output and strict thermal requirements during the spring (Busch et al., 1975;Farmer et al., 2015;Hokanson, 1977). However, depending on their magnitude, reductions in total phosphorus (TP) inputs from ACP implementation could offset or exacerbate anticipated climate-driven declines in yellow perch recruitment. For example, reduced NPS TP loading could offset warming-induced recruitment declines by alleviating the adverse effects of eutrophic conditions (e.g. bottom hypoxia; Caddy, 1993;Roberts et al., 2009;Scavia et al., 2014). By contrast, reduced NPS inputs of phosphorus (and associated sediments) could exacerbate declines in recruitment by reducing the extent and productivity of turbid Maumee River plumes during the spring, which appear to offer protection to larval yellow perch from predators such as invasive white F I G U R E 1 Map of the western Lake Erie basin catchment (MI, IN, OH, U.S.A.) and the west basin of Lake Erie (U.S.A.-Canada). Trawling stations are denoted by solid black circles, weather stations by black triangles, and the Maumee River gauge station used to validate the Soil and Water Assessment Tool model by a gold star perch (Carreon-Martinez et al., 2014;Ludsin et al., 2011) and may also enhance foraging through bottom-up effects (Barbiero, Balcer, Rockwell, & Tuchman, 2009;Grimes & Finucane, 1991). In this way, the relationship between yellow perch recruitment and TP inputs from the Maumee River could be expected to be dome-shaped (unimodal), a notion supported by previous studies that have quantified the relationship between ecosystem productivity metrics and fish(eries) yield (e.g. Caddy, 1993Caddy, , 2000Oglesby, Leach, & Forney, 1987). By contrast, we hypothesised that white perch recruitment would be positively affected by climate warming and negatively affected by increased ACP implementation because of this species' higher thermal optimum compared to walleye and yellow perch (Johnson & Evans, 1990) and its tolerance of eutrophic conditions (Briland, 2018).

| West basin of Lake Erie
Lake Erie is the smallest of the Laurentian Great Lakes (by volume), but is the most biologically productive, supporting numerous commercial and recreational fisheries (Bunnell et al., 2013;Ludsin et al., 2014). The lake has three distinct basins (west, central, and east), each having unique chemical and physical properties. The focal area of this study, the west basin (Figure 1), is the warmest, shallowest, and most biologically productive of the three (Bolsenga & Herdendorf, 1993;Leach & Nepszy, 1976;Ludsin et al., 2001). Owing to these properties, the west basin has historically provided spawning and nursery habitat for a diversity of fishes including walleye, yellow perch, and white perch (DuFour et al., 2015;Farmer et al., 2015;Jones et al., 2003;Ludsin et al., 2001Ludsin et al., , 2011Mion et al., 1998;ODW, 2017). A major driver of the biologically productive west basin is nutrient inputs from the WLEB catchment and specifically, the Maumee River, which drains the largest catchment in the entire Great Lakes basin (Bolsenga & Herdendorf, 1993). The Maumee River catchment is composed of primarily agricultural land (>70%, USDA NRCS, 2011) and delivers substantial nutrient and sediment inputs to the west basin (Baker & Richards, 2002;Keitzer et al., 2016;Scavia et al., 2014). Excessive phosphorus loading from agricultural runoff in the Maumee River catchment has been identified as the primary driver of Lake Erie's recent re-eutrophication (Scavia et al., 2014;Watson et al., 2016), resulting in efforts to reduce loading via the implementation of ACPs (Ohio EPA, 2013). Given that the Maumee River contributes a substantial portion of the TP load from the WLEB into the west basin of Lake Erie (Maccoux, Dove, Backus, & Dolan, 2016;Scavia et al., 2014), the need exists to understand how ACP implementation in this catchment will impact the resident fish community, especially amidst a changing climate .

| Study species
The three focal species in this study are: (1) walleye, which supports Lake Erie's largest recreational fishery and second largest commercial fishery (Kayle, Oldenburg, Murray, Francis, & Markham, 2015;Markham & Knight, 2017;ODW, 2017ODW, , 2018; (2) yellow perch, which supports Lake Erie's largest commercial fishery and second largest recreational fishery (Belore et al., 2014;ODW, 2018); and (3) white perch, an invasive species that is of minor commercial importance but has become the most abundant prey-fish in Lake Erie (FTG, 2013;ODW, 2017) and is a known predator of walleye and yellow perch during their early life stages (Carreon-Martinez et al., 2014;Ludsin et al., 2011;Schaeffer & Margraf, 1987). In addition to their economic and ecological roles in Lake Erie, walleye, yellow perch, and white perch are common and widespread across North America, and support recreational fisheries across their ranges. These three species also span a gradient of thermal F I G U R E 2 Conceptual modelling framework for forecasting Lake Erie walleye, yellow perch, and white perch recruitment under two greenhouse gas emission scenarios (RCP4.5 and RCP8.5) crossed by four levels of agricultural conservation practice (ACP) implementation in the western Lake Erie basin (WLEB) catchment. Predictive models of fish recruitment were built using generalised additive models (GAMs). Predictor variables included winter severity, spring warming rate, and Maumee River total phosphorus (TP) from Heidelberg University's National Center for Water Quality Research (NCWQR). Age-0 (juvenile) abundance data from bottom trawl surveys by the Ohio Department of Natural Resources-Division of Wildlife (ODNR-DOW) and the Ontario Ministry of Natural Resources and Forestry (OMNRF, 1987(OMNRF, -2015 was our recruitment proxy. Future recruitment was projected using ensemble climate forecasts and linked Soil and Water Assessment Tool (SWAT) and Agricultural Policy/Environmental eXtender (APEX) models guilds. Although walleye and yellow perch are both considered coolwater species, walleye has a lower optimal temperature range for both spawning and embryo hatching compared to yellow perch (Hokanson, 1977). In Lake Erie, walleye typically spawn during March-mid-May at temperatures of 3-12°C (May, 2015;Roseman et al., 1996) and yellow perch typically spawn during mid-April to May at temperatures of 8-14°C (Belore et al., 2014;Collingsworth & Marschall, 2011;Farmer et al., 2015). Both walleye and yellow perch age-0 (juvenile) abundance indices during August (c. 3-4 months post-hatch) are excellent predictors of age-2 abundance when recruitment to the fishery occurs (Belore et al., 2014;Farmer et al., 2015;Kayle et al., 2015), highlighting the importance of early life processes in determining fishery production. Because white perch prefer warmer temperatures compared to walleye and yellow perch, the west basin of Lake Erie is a favorable recruitment environment for this species. White perch spawn in its tributaries (e.g. Maumee River) and shallow (<1.5 m) waters during spring at temperatures of 11-15°C (Boileau, 1985;Hartman, 1972).

| Modelling overview
Our modelling approach involved two primary steps ( Figure 2).
First, we built species-specific predictive recruitment models using historical age-0 (juvenile) abundance data and indices of winter severity, spring warming rate, and Maumee River TP loads. Second, we projected future (2020-2065) recruitment under different combinations of anticipated climate change and four levels of ACP implementation in the WLEB catchment, using our predictive models and forecasted values of winter severity, spring warming rate, and Maumee River TP loads that were generated from linked climate, catchment-hydrology, and agricultural-practice-simulation models.

| Fish recruitment indices
We used annual indices of age-0 abundance generated from bot- large boom-bust cycles) during which environmental drivers were probably not the main drivers of population dynamics (Simberloff & Gibbons, 2004;Williamson, 1996). Surveys were conducted during the last 2 weeks of August, sometimes extending into early September.
The agencies used a stratified-random design to sample 56-70 stations annually across the west basin of Lake Erie. To correct for catchability differences, vessel-specific and species-specific fishing power corrections were applied to standardise trawl catches (Tyson, Johnson, Knight, & Bur, 2006). Catches were averaged within each year to generate a basin-wide mean. Herein, we report annual catch per unit effort (CPUE) as the number of individuals per minute of trawling.

| Abiotic predictors
We assessed winter severity, spring warming rate, and total springtime (March-May) Maumee River TP loads as potential environmental predictors of walleye, yellow perch, and white perch recruitment.
We chose these metrics based on previous research indicating their influence on recruitment for at least one of the three species (Busch et al., 1975;Carreon-Martinez et al., 2014;Clady, 1976;Farmer et al., 2015;Hokanson, 1977;Johnson & Evans, 1990;Jones et al., 2006;Ludsin et al., 2011;Mion et al., 1998;Reichert et al., 2010). We ). Our use of air temperature is justified, as it has often been used as a proxy for water-related thermal metrics (Sharma, Jackson, Minns, & Shuter, 2007;Van Zuiden et al., 2016) and has been used to successfully predict the effects of climate change on fisheries (Van Zuiden et al., 2016). Furthermore, previous research has shown that local air temperature is correlated to Lake Erie water temperature in western Lake Erie (Farmer, 2013). We defined winter severity as the total number of days that the mean daily maximum temperature across the WLEB catchment was ≤6°C during January-April of each year. Our inclusion of winter severity as a recruitment predictor was based on previous research identifying optimal spawning conditions for percids, which showed that at colder water temperatures, maturation is more likely to result in a spawning event as compared to warmer water temperatures (Hokanson, 1977). Our winter index is also strongly correlated with ice cover indices used in previous studies (e.g. Farmer et al., 2015) that have been shown to explain variation in Lake Erie recruitment success, and it is easily projected using existing climate change models. To calculate spring warming rate (°C/ day), we fit a linear regression model to the basin-wide mean daily maximum temperature during April-May (the approximate larval production period for all three species) for each year and defined the annual spring warming rate (°C/day) as the slope of the least-squares regression line (Busch et al., 1975

| Predictive models of fish recruitment
Because environmental driver-biological response relationships are often nonlinear (Hunsicker et al., 2016), we used generalised additive models (GAMs, Hastie & Tibshirani, 1987) to examine the relationship between fish recruitment and winter severity, spring warming rate, and Maumee River TP loads. Generalised additive models are an ideal tool because of their flexibility in fitting nonlinear relationships that need not be defined a priori, and they have previously been used to model fish recruitment (Cardinale & Arrhenius, 2000;Daskalov, 1999). We built species-specific GAMs using thin-plate regression splines (Adams, Leaf, Wu, & Hernandez, 2018) and a γ-distribution with a log-link function. The γ-distribution is a flexible, continuous distribution, appropriate for skewed data, which has commonly been used in fisheries applications (Maunder & Punt, 2004). Smoothness parameters were estimated with generalised cross validation. Because fewer years of data were available for white perch, as compared to walleye and yellow perch, we restricted the basis dimension, k (controls the degree of smoothness in the model), to 6 in all white perch models to avoid overfitting (Decker et al., 2013;Quiñones et al., 2015). To avoid the confounding effects of multicollinearity, prior to model construction, we used pairwise correlations to confirm that no substantial (r > 0.6, Zuur, Ieno, Walker, Saveliev, & Smith, 2009) multicollinearity between predictor variables was present.
Once we determined that the predictor variables were not strongly correlated, a global model with the form: was constructed and fit for each species where s represents the smoothing function. Model fit and temporal autocorrelation were assessed using standard diagnostics (Anderson & Burnham, 2002).
To determine the most supported predictive model for each species, we constructed model formulations with all possible combinations of predictor variables (Burnham & Anderson, 2003) and considered the most supported model for each species as the one with the lowest sample size-corrected Akaike information criteria (AICc) value. If models were equally supported (ΔAICc < 2), we chose the model with the greatest predictive ability. All GAM analyses were conducted using the mgcv package in R 3.3.0 (Wood & Wood, 2015; R Core Team, 2019).
We used a resampling technique to visualise the partial effects of each predictor variable on the response variable (recruitment).
First, we generated a separate, simulated uniform sequence (n = 250) of each predictor variable using its observed range as bounds. Next, using the original data, we resampled the other two predictor variables (with replacement) to generate a new resampled dataset. We repeated this process with each predictor to generate three new resampled datasets, each with a simulated sequence of one variable and a resampled sequence for the other(s). Finally, we predicted recruitment values using the most supported candidate model for each species for each new dataset.
We then plotted the predicted recruitment values as a function of the simulated predictor sequence for each of the predictors included in the most supported model (maximum of three possible predictors).

| Climate change scenarios: thermal metrics
We projected future (2020-2065) winter severity and spring warming rate using daily maximum air temperature values from global circulation models (GCMs) used in the IPCC's Fifth Assessment Report (IPCC, 2014). We included two greenhouse gas emission (representative concentration pathway, RCP) scenarios: a moderate-reductions scenario (RCP4.5, 18 GCMs) and a business-as-usual scenario (i.e. status quo, RCP8.5, 17 GCMs). For each GCM, multiple ensembles (slightly different versions of a GCM model) were run to generate a range of future climate conditions. (RCP4.5, n = 38; RCP8.5, n = 37, Table S1; Ohio Supercomputer Center, 1987). The use of multiple GCMs and ensembles is a common way to incorporate uncertainty in future climate conditions and to better reflect the range of possible future outcomes (Sharma, Vander Zanden, Magnuson, & Lyons, 2011;Van Zuiden et al., 2016). The GCM data spatially overlapped the 29 stations that were used to calculate historical winter severity and spring warming rate.
Although the GCM outputs were bias-corrected and downscaled (both temporally and spatially) using standard approaches (Bureau of Reclamation, 2013;Maurer, Brekke, Pruitt, & Duffy, 2007), we further adjusted the forecasted data to account for any remaining bias, given that the climate data were generated at different temporal and spatial scales than those used in this study. For our winter severity and spring warming metrics, we employed a multi-step approach.
First, we built scenario-specific linear models between hindcasted  temperature data from each climate scenario and historical observed temperature data during the same period. We then applied the linear coefficients to future forecasted temperature data. Next, we used the bias-corrected temperature data to calculate winter severity and spring warming rate with the same methods described above. We then calculated the magnitude of change between the forecasted value and hindcasted  median value for that specific scenario. Finally, we added this difference to the historical median. We used the resulting forecasted values to predict recruitment, which mitigated potential bias in our modelling efforts.
We removed a small subset (n = 11) of scenarios in which the projected spring warming rate was ≤0 for the remainder of our analyses.
These scenarios only represented c. 0.3% of our total projections.

| Agricultural conservation practice scenarios
To generate future TP loads that resulted from different levels of implementation of nutrient and sediment reduction strategies in the catchment, we applied a set of previously developed catchment-hydrology and agricultural-simulation models (Arnold, Srinivasan, Muttiah, & Williams, 1998;Gassman et al., 2009;USDA NRCS, 2011;Wang et al., 2011;Daggupati et al., 2015;Yen et al., 2016 which reliably simulates TP input into the west basin of Lake Erie via the Maumee River (Daggupati et al., 2015;Yen et al., 2016).
The SWAT is a commonly used catchment-hydrology model and has been used in multiple WLEB studies to explore the impact of climate change and ACP implementation on river flow and agricultural runoff (e.g. Bosch et al., 2013;Keitzer et al., 2016). In our study, SWAT inputs were nutrient loads from cultivated fields simulated by APEX (under each level of ACP implementation) and meteorological data from each climate change model. We used an approach similar to that used for our thermal metrics to bias-correct the TP forecasts from the SWAT, with some notable differences. In this case, we built scenario-specific and month-specific (March-May) linear models because the observed biases were month-specific. After forecasting springtime TP loads, we calculated the proportional change between the forecasted value and the hindcasted (1987-2015) median value for that specific scenario. Regardless of the level of ACP implementation, we used the baseline ACP median in each emission scenario, which simulated historical nutrient and erosion control practices.
Finally, we multiplied the proportional change by the historical median to obtain the forecasts that were used to predict recruitment.
We chose to use proportional change for TP to account for precipitation-driven changes in TP loading in the climate models, as simply adding or subtracting load differences would not make sense. The results of this modelling effort were projections of Maumee River TP loads for each combination of climate change and ACP implementation scenario, analogous to the historical Maumee River TP loads in regards to timing (spring) and spatial scale.

| Forecasting fish recruitment
We used our predictive recruitment models to forecast walleye, yellow perch, and white perch recruitment under both RCP scenarios and the four ACP implementation scenarios. Under each combination of climate and ACP implementation scenario (total n = 8), species-specific recruitment values were projected from winter severity, spring warming rate, and/or Maumee River TP loads annually and subsequently summarised by decade.
We evaluated the effects of future climate change and ACP implementation on fish recruitment at the decadal scale using two approaches. First, we compared the median decadal recruitment trends, relative to the past . Second, because walleye and yellow perch fisheries in Lake Erie are supported by sporadic strong recruitment events (i.e. year-classes; Farmer et al., 2015;Vandergoot, Cook, Thomas, Einhouse, & Murray, 2010), we calculated the frequency of annual forecasts in a decade that would constitute a strong recruitment event, defined as greater than or equal to the historical 75th percentile. The 75th percentile has commonly been used as a metric to define a strong recruitment event in Lake Erie (Vandergoot et al., 2010). We calculated the median proportion of strong recruitment events in each decade under each climate and ACP implementation scenario. If no differences in the frequency of strong recruitment events in the future existed relative to the past, we would expect the proportion of strong recruitment events within a decade to be centered on 25%, indicating that strong recruitment events occurred at the same frequency compared to the past.

| Fish recruitment indices
Lake Erie walleye, yellow perch, and white perch recruitment, as in-

| Abiotic predictors
Winter severity, spring warming rate, and Maumee River TP loads also varied throughout the historical period (Figure 3d

| Predictive models of fish recruitment
Based on standard, qualitative diagnostics (Anderson & Burnham, 2002) all three final models of fish recruitment displayed good fit with no obvious patterns in the residuals ( Figure S1)  significant temporal autocorrelation based on autocorrelation function plots ( Figure S2, Zuur et al., 2009). Based on our defined AICc and predictive ability criteria, the most supported predictive model of walleye recruitment included only winter severity (Table S2). Our resampling analysis revealed that, as winter severity increased, so did walleye recruitment, with exponentially higher walleye recruitment occurring after severe winters ( Figure   S3). The final walleye model explained 31.3% overall deviance in walleye recruitment (Table 1). The most supported predictive model of yellow perch recruitment included winter severity and Maumee River TP as predictors (Table S2). The partial response of yellow perch recruitment to winter severity was similar to the walleye partial response; as winter severity increased, so did yellow perch recruitment ( Figure S4). The partial response of yellow perch recruitment to TP was dome-shaped (unimodal) with the greatest positive effect occurring at intermediate TP loads ( Figure S4). The final yellow perch predictive model explained 51.1% of the overall deviance in yellow perch recruitment (Table 1). The most supported predictive model of white perch recruitment included spring warming rate and Maumee River TP as predictor variables (Table S2). In general, white perch recruitment increased with increasing spring warming rate ( Figure   S5). Although the partial response of white perch recruitment to Maumee River TP was nonlinear, as TP increased white perch recruitment generally decreased ( Figure S5), a finding that ran counter to our expectations. The final white perch recruitment model explained 55.4% of the variation in observed recruitment (Table 1).

| Forecasts of winter severity and spring warming rate
As expected, winter severity decreased through time in both the RCP4.5 and RCP8.5 emission scenarios, although the declines varied in their magnitude, especially during later decades (Figure 4a). in the RCP8.5 emission scenario, representing a 21% and 29% reduction, respectively, relative to the historical period. In contrast to winter severity, no obvious temporal trends in projected spring warming rate were apparent, with differences between emission scenarios also being negligible (Figure 4b). Median spring warming rates under the RCP4.5 emission scenario were slightly higher than the historical median, although the projected rates were variable ( Figure 4b).

| Forecasts of TP loading
Three general patterns emerged from our projected Maumee River TP loads ( Figure 5) Figure 5). However, in the ACP implementation scenarios where current levels of implementation were carried into the future or only acres in critical need were treated, TP loads were forecast to increase above the historical median levels ( Figure 5).

Walleye
The final walleye recruitment model only included winter severity as a predictor. Thus, walleye recruitment was not projected under different levels of ACP implementation, only under different greenhouse gas emission (RCP) scenarios. In general, median annual projections of recruitment decreased through time in both the moderate reductions (RCP4.5) and business-as-usual (RCP8.5) scenarios ( Figure 6). Interestingly, the projected median annual recruitment values during earlier decades (2020s-2040s for RCP4.5, 2020s and 2030s for RCP8.5) were 3-50% higher than the historical median, owing to more variable forecasts that resulted in projected severe winters during earlier decades. However, during subsequent decades (2040s-2060s), both emission scenarios had projected median annual recruitment values lower than the historical median ( Figure 6). During the 2060s under the RCP8.5 emission scenario, which represents the worst-case scenario in our projections, median annual recruitment decreased by 38% relative to the historical median. By contrast, the frequency of a projected strong (≥ historical 75th percentile) annual walleye recruitment event (year-class) during a decade was lower than 25% (the expected frequency) during all decades, under both emission scenarios (Figure 7), except under the RCP8.5 emission scenario during the 2020s. Under the RCP4.5 emission scenario, the median frequency of a projected annual strong recruitment event decreased from 18% during the 2020s to 5% during the 2060s. The decline was more severe under the RCP8.5 emission scenario during which the median value was only 5 and 4% in the 2050s and 2060s, respectively. The median frequency of strong recruitment events was slightly higher early in the projected period under the RCP8.5 emission scenario compared to the RCP4.5 one, owing to a slightly higher occurrence of projected severe winters during early decades. Overall, under both future greenhouse gas emission scenarios, the projected frequency of strong walleye recruitment events decreased substantially compared to the past.

Yellow perch
Our analysis of yellow perch recruitment yielded four major findings.
First, median annual recruitment was projected to be lower than the historical median across all decades under all climate change × ACP implementation scenarios, except for one (RCP8.5, ENMC, 2020s;  (Figure 9). Projected strong yellow perch recruitment events decreased through time, decreased with increasing levels of ACP implementation, and were lower than expected (a frequency of 25%, based on the historical frequency) across all decades during all future scenarios (Figure 9). In contrast to the projected median levels of yellow perch recruitment, the highest projected frequency of strong recruitment events (23%) occurred under the RCP8.5-ENMC scenario, in part, owing to a greater proportion of projected severe winters relative to the RCP4.5 emissions scenario (Figure 9).

White perch
Unlike walleye and yellow perch, recruitment of invasive white perch was projected to be near or above the historical median across all climate and ACP implementation scenarios ( Figure 10).

The projected increases in the median annual recruitment values
were typically greatest during the near-term (2020s and 2030s) at the two highest levels of ACP implementation (ENMA and ENM; Figure 10). Although temporal trends in median white perch recruitment were less apparent relative to yellow perch and walleye, white perch recruitment was projected to be slightly higher

| D ISCUSS I ON
Predicting the effects of anthropogenic stressors such as climate change and agricultural-derived NPS pollution has been identified as a critical research need that could benefit fisheries management in the face of future ecosystem change (Arvai et al., 2006;DeVanna Fussell et al., 2016;Pritt et al., 2014). The Great Lakes, and specifically Lake Erie, is an ideal study system for such work because it has experienced these anthropogenic stressors, supports valuable fisheries, and is data-rich (Farmer et al., 2015;Ludsin et al., 2014;Pritt et al., 2014;Scavia et al., 2014). In this study, we forecasted how the recruitment of three ecologically and/or economically important western Lake Erie fish populations, which span a gradient of thermal preferences, might vary under future scenarios of climate change and ACP implementation in the WLEB catchment. Our modelling showed that, in general, walleye and yellow perch recruitment can be expected to decrease and that white perch recruitment can be expected to remain stable or increase during the next several decades, relative to the recent past. Interestingly, our modelling also revealed offsetting effects between climate change and ACP implementation, highlighting the potential for trade-offs between improving water quality, maintaining fisheries production, and controlling invasive species in the face of potential climate change. Although attaining a complete understanding of future recruitment dynamics is impossible (Schindler & Hilborn, 2015), and more research is encouraged to verify some of our suggested mechanistic linkages and recruitment projections, our study presents a useful modelling framework to forecast fish population dynamics, specifically recruitment, and provides a range of potential outcomes for resource management agencies and policy-makers that can help them develop adaptive and resilient management strategies in the face of continued ecosystem change (Heller & Zavaleta, 2009;Lynch et al., 2016;Paukert et al., 2016).
Our results support previous studies, which have predicted that climate warming will differentially affect species with varying thermal preferences. Similar to other studies (Chu, Mandrak, & Minns, 2005; stages (e.g. juvenile) and specific biological processes (e.g. reproduction and ovary development) that require cold temperatures, and therefore, are more likely to be affected by warming. Understanding the influence of warming is especially critical for populations such as Lake Erie walleye and yellow perch, the recruitment dynamics of which have been shown to be influenced by temperature, which in turn drives variability in the fishery (Farmer et al., 2015;Shuter & Koonce, 1977;WTG, 2017). In fact, species that are sensitive to winter conditions (such as walleye and yellow perch) may be the first to be affected by climate change (Shuter, Minns, & Lester, 2002).

| Walleye recruitment
Based on our findings, western Lake Erie walleye recruitment, especially episodically strong recruitment events that keep the F I G U R E 9 Boxplot of the proportion of projected (2020-2065) forecasts of annual western Lake Erie yellow perch recruitment events (year-classes) that were greater than or equal to the historical  75th percentile by decade and agricultural conservation practice implementation scenario. Projections were made for two greenhouse gas emission scenarios: RCP4.5 (grey) and RCP8.5 (white). The horizontal line represents the expected proportion of strong recruitment events, if the frequency were not to change in the future relative to the past. See the Figure 5 legend for a description of each agricultural conservation practice scenario recreational and commercial fisheries viable (Vandergoot et al., 2010), were predicted to decline, owing to a projected reduction in winter severity. Recruitment declines were greatest further into the 21st century and in the business-as-usual greenhouse gas emission scenario (RCP8.5), relative to the moderate-reduction (RCP4.5) scenario. Although we documented an overall downward trend in projected walleye recruitment in the future, median walleye recruitment was generally projected to be at or above the historical median during the 2020s-2040s. Even so, the projected frequency of strong walleye recruitment events was below the expected proportion (25%) during these and subsequent decades, under both greenhouse gas emission scenarios.
Our modelling results are consistent with other modelling stud- we fully expect western Lake Erie to continue to support walleye fisheries, our modelling suggests that the strong recruitment events (year-classes) that drive order of magnitude differences in the fishable population (Vandergoot et al., 2010) may decline with continued climate warming.
Much work on Lake Erie walleye recruitment has suggested the rate a three-dimensional hydrodynamic model to demonstrate that walleye recruitment was more strongly associated with wind speed and direction than it was with spring warming rate. Data such as wind speed, however, are not available from climate models and were therefore not considered in this study.
In addition to abiotic factors, biotic factors are also likely to affect walleye recruitment, which we did not consider. For example, prey abundance, specifically, age-0 gizzard shad (Dorosoma cepedianum) abundance has been correlated to Lake Erie walleye recruitment (Madenjian et al., 1996), but not with consistent, replicable results (Zhao et al., 2013). It is also possible that climate change could have indirect effects on walleye recruitment that is mediated by biotic factors such as zooplankton prey availability to larvae . Ultimately, our understanding of walleye recruitment in Lake Erie remains largely speculative and is based primarily on correlative work. While the use of winter severity to predict walleye recruitment is partially supported by mechanistic evidence, as walleye prefer cooler incubation and fertilisation temperatures (Koenst & Smith, 1976), exactly why the relationship between winter severity and walleye recruitment exists remains unknown (Fedor, 2008).
More research into this linkage is warranted, as well as into how climate change might affect walleye through other direct and indirect pathways.

| Yellow perch recruitment
Similar to walleye, our modelling projected that Lake Erie yellow perch recruitment will decline under future climate warming scenarios. Furthermore, it suggests that this decline would be exacerbated by efforts to reduce nutrient inputs (i.e. TP) into Lake Erie via F I G U R E 11 Boxplot of the proportion of projected  forecasts of annual western Lake Erie white perch recruitment events (yearclasses) that were greater than or equal to the historical  75th percentile by decade and agricultural conservation practice implementation scenario. Projections were made for two greenhouse gas emission scenarios: RCP4.5 (grey) and RCP8.5 (white). The horizontal line represents the expected proportion of strong recruitment events, if the frequency were not to change in the future relative to the past. See the Figure 5 legend for a description of each agricultural conservation practice scenario the implementation of ACPs in the WLEB catchment. The prominent driver of the projected yellow perch recruitment decline was reduced winter severity, with the level of ACP implementation having a secondary negative effect. Although yellow perch recruitment displayed a unimodal, dome-shaped response to Maumee River TP inputs, future anticipated ACP implementation led to TP loads that were lower than the optimum for strong recruitment events to occur. Previous studies of yellow perch recruitment have suggested that rapid warming during the spring can positively (Eshenroder, 1977) and negatively (Zhang, Reid, & Nudds, 2016) (Farmer et al., 2015;Hokanson, 1977). Additionally, our modelling suggests that ACP implementation efforts designed to improve water quality by reducing NPS nutrient loading could inadvertently reduce fisheries production, a notion that was espoused earlier when Lake Erie was undergoing oligotrophication (Ludsin, 2000;Ludsin et al., 2001). In this way, ACP implementation could potentially magnify the anticipated negative effects of climate warming on yellow perch recruitment.
The need to consider these kinds of trade-offs is paramount as they could help fishery managers and policy-makers identify nutrient mitigation strategies that improve water quality without compromising fisheries production.
Our yellow perch model explained 51.1% of the variation in recruitment and included both winter severity and Maumee River TP as environmental predictors. Similar to walleye, various measures of spawning-stock size have not consistently explained yellow perch recruitment variation (Henderson, 1985;Henderson & Nepszy, 1988;Zhang et al., 2016) and yellow perch recruitment synchrony throughout the Great Lakes region indicate that regional-scale environmental factors, such as those included in this study, are more likely than stock size to drive recruitment (Honsey et al., 2016). The two environmental drivers of yellow perch recruitment that we identified are consistent with findings from previous correlative and mechanistic studies (Carreon-Martinez et al., 2014;Farmer et al., 2015;Hall & Rudstam, 1999;Hokanson, 1977;Ludsin et al., 2011;Reichert et al., 2010). For example, a greater percentage of yellow perch successfully spawn at colder water temperatures and after long chill durations (Hokanson, 1977) compared to warmer temperatures, indicating a benefit of long, cold winters for yellow perch. This finding is supported by experimental research, which has demonstrated that short warm winters cause reduced egg hatching success and reduced egg and larvae size and quality (Farmer et al., 2015). Declines in yellow perch abundance have also previously been correlated with reduced TP availability (Hall & Rudstam, 1999).
The possibility exists, however, that Maumee River TP is only a proxy for a more complex suite of ecological responses associated with Maumee River discharge and nutrient and sediment loading, all of which are highly correlated (D.A.D., unpublished data).
Turbid, nutrient-rich river plumes, which are created by Maumee River inflow during the spring, have been shown to lead to greater yellow perch recruitment in western Lake Erie (Carreon-Martinez et al., 2014;Ludsin et al., 2011;Reichert et al., 2010). Survival of larvae inside the Maumee River plume has been shown to be greater than larval survival outside the plume (Carreon-Martinez et al., 2014;Reichert et al., 2010), and this difference appears due to reduced predation inside the plume (Carreon-Martinez et al., 2014;Ludsin et al., 2011;Reichert et al., 2010). Although the exact causal mechanism(s) remain incomplete, sediment and/or nutrient loading from the Maumee River seem(s) to have a positive effect on yellow perch survival to the age-0 stage. Thus, reduced nutrient loading in the future via ACP implementation could have negative effects on yellow perch recruitment. Given the high degree of covariation among TP loading, sediment loading, and Maumee River inflows, as well as a possible trade-off between water quality and yellow perch production with ACP implementation under a changing climate, we advocate for more research aimed at identifying the mechanism(s) underlying the unimodal relationship that we found between Maumee River TP loading and yellow perch recruitment.

| White perch recruitment
White perch recruitment (both median levels and the frequency of strong recruitment events) was forecasted to be close to or greater than the historical median across all scenarios during all decades, and increased levels of ACP implementation (i.e. reduced TP loading) resulted in generally higher white perch recruitment. Substantially less information exists on the drivers of white perch recruitment relative to walleye or yellow perch, especially in ecosystems where this species is invasive (e.g. Lake Erie). However, our finding that climate warming may lead to higher white perch recruitment is generally consistent with the literature that does exist. For example, Johnson and Evans (1990) speculated that climate warming would cause higher recruitment and ultimately expansion of white perch in the Great Lakes by reducing overwinter mortality. Although winter severity was not included in the final predictive model of white perch recruitment, climate warming could possibly result in a longer growing season and improve overwinter survival (Johnson & Evans, 1990). The generally negative (although nonlinear) relationship between Maumee River TP and white perch recruitment was the opposite of our expectation, as white perch generally prefer eutrophic over oligotrophic waters (Boileau, 1985).
Adult white perch abundance has also previously been shown to be positively associated with high turbidity and eutrophic conditions in other ecosystems (Hawes & Parrish, 2003), indicating a need for more research to understand our observed association.
Similar to yellow perch, the possibility exists that our TP metric is only a proxy for another correlated abiotic factor, such as river discharge, which could actually be the driver of our observed association. For example, because white perch spawn in west basin tributaries (Boileau, 1985;Schaeffer & Margraf, 1986, high river discharge events (strongly correlated with TP loading) could potentially dislodge or flush white perch eggs out of the Maumee River prematurely, thus reducing their survival potential. This hypothesis is consistent with previous research that found significant negative correlations between Maumee River discharge and age-0 white perch abundance in the west basin of Lake Erie (Briland, 2018). Maumee River TP could also be a proxy for a complex biotic mechanism. For example, another plausible hypothesis is that high levels of TP, which results in high yellow perch recruitment, could reduce white perch recruitment through interspecific competition.
Such interactions between age-0 white perch and yellow perch have been observed in other ecosystems (Prout, Mills, & Forney, 1990

| Study limitations
As with all forecasting studies, our approach has several limitations.
Although the proportions of variance in recruitment that our predictive models explained were similar to or better than those reported in the literature (see species-specific examples above), they certainly did not explain all, or in some cases the majority of recruitment variation. Still, such models can be useful in assessing the impacts of climate change (Guisan & Thuiller, 2005), although we strongly encourage managers to consider the breadth of information available when making future management and conservation decisions.
Furthermore, care should be taken when extrapolating recruitment responses to environmental conditions outside the range of observed historical conditions, as species-environment relationships may not be stationary (Schindler & Hilborn, 2015;Zhang, Reid, & Nudds, 2018). We also caution against interpreting the exact magnitude of our future recruitment values in a predict and prescribe approach (Schindler & Hilborn, 2015), given mechanistic uncertainties associated our predictive models. Even so, we feel comfortable interpreting the general trends and drivers apparent in our results and their use as guidance for future management.
Forecasting future population dynamics will always be incomplete, uncertain, and a simplification of reality, regardless of the ecosystem. However, studies like ours provide a range of possible outcomes that can be used as tools for resource managers (Schindler & Hilborn, 2015). While we do not know the true magnitude or extent of future warming, by modelling multiple greenhouse gas emission scenarios, using a suite of climate ensembles, we could propagate some of that future uncertainty into our recruitment forecasts (Hansen et al., 2017). This ensemble approach to forecasting future dynamics is ubiquitous and is an accepted way of acknowledging the uncertainty in future predictions (Bartolino et al., 2014;Hollowed et al., 2009;Lindegren et al., 2010), although we recognise this approach does not account for all of the uncertainty associated with forecasting future recruitment.
Another limitation to our study is that our methods were correlative and did not verify the underlying mechanisms by which winter severity, spring warming rate, and TP loads can alter recruitment dynamics. Thus, while we provided mechanistic support for their inclusion in our predictive models, which strengthens the confidence in our results (Hilborn, 2016), the possibility exists that the metrics included in this study encompass several underlying mechanisms (Hansen et al., 2017) or are actually proxies for other correlated environmental drivers. For example, because Maumee River TP loads into Lake Erie are highly correlated with Maumee River inflows and total suspended sediment loads, TP itself may not be the exact mechanistic driver of yellow perch or white perch recruitment.
Owing to the difficulty in implementing and designing rigorous, experimental approaches to determine causal relationships over large spatial and temporal extents (Hilborn, 2016), correlative studies such as ours remain the most reasonable approach to forecasting recruitment on large spatial and temporal scales (Guisan & Thuiller, 2005;Hansen et al., 2017).
Because of the inherent difficulty in forecasting future biotic conditions, we restricted our analysis to include only abiotic predictors. However, biotic factors such as competition and predation probably also contribute to current recruitment dynamics of these species, and will probably affect future recruitment dynamics (Forsythe, Doll, & Lauer, 2012;Guisan & Thuiller, 2005;Hall & Rudstam, 1999;Hartman & Margraf, 1993). Thus, we encourage continued investigations into the drivers of recruitment for all three species, especially those that consider other factors, use alternative modelling approaches, and occur at different spatiotemporal scales (Hilborn, 2016). This need is especially critical because the mechanisms underlying our observed correlations are unlikely to remain stationary in the future (Schindler & Hilborn, 2015).

| CON CLUS IONS
Our modelling allowed us to explore how anticipated climate change and ACP implementation designed to reduce NPS nutrient loading might interact to affect the recruitment dynamics of ecologically and economically important fish populations in Lake Erie. By including two emission scenarios, four levels of ACP implementation, and numerous GCMs and ensembles, we forecasted a range of future outcomes to better equip resource managers to make decisions that can promote sustainable and resilient fisheries in the future Paukert et al., 2016). Our findings highlight the importance of climate as a driver of fish recruitment dynamics and indicate that, in the future, native cool-water species such as walleye and yellow perch may be detrimentally affected by climate change, whereas nonnative warm-water species such as white perch might benefit.
Our modelling also suggests that reducing nutrient inputs to improve water quality (though ACP implementation) may lead to inadvertent trade-offs that could negatively affect the production of valued fisheries (Kao, Rogers, & Bunnell, 2018;Ludsin, 2000;Ludsin et al., 2001;Ney, 1996). For example, our modelling provided evidence to suggest that reduced nutrient (or possibly sediment) runoff from the WLEB catchment-resultant of ACP implementation-could exacerbate anticipated climate-driven reductions in western Lake Erie yellow perch recruitment. Simultaneously, these same conditions were projected to promote invasive white perch, which is a known predator on walleye and yellow perch early life stages (Carreon-Martinez et al., 2014;Ludsin et al., 2011;Schaeffer & Margraf, 1987).
In addition to identifying a need for more research into the mechanistic relationships among climate, catchment runoff, and yellow perch and white perch recruitment, we recommend that future studies seeking to quantify the independent and combined effects of human-driven perturbations (e.g. climate change, altered nutrient inputs) assess both the costs and benefits associated with changing conditions, in both upstream and downstream ecosystem services.
Such information would allow for the development of improved forecasting models, as well as allow resource management agencies and policy-makers to better anticipate trade-offs and avoid ecological surprises. For example, decision-makers could learn whether any likely combination of climate and land use conditions provide a winwin scenario (sensu Keitzer et al., 2016) for upstream (catchment) fish production, downstream (recipient ecosystem) water quality (e.g. reduced bottom hypoxia and harmful algal blooms), and downstream fisheries production. Armed with this knowledge, informed decisions can be made to keep fisheries productive and sustainable in the face of continued ecosystem change.

ACK N OWLED G EM ENTS
We would like to thank the ODNR-DOW and the OMNRF for providing data used in this manuscript. We would also like to thank USGS

CO N FLI C T O F I NTE R E S T
The authors report no conflicts of interest.

DATA AVA I L A B I L I T Y S TAT E M E N T
Data used to build the recruitment models are available as a separate file. The climate data are available at https://gdo-dcp.ucllnl.org/ downs caled_cmip_proje ction s/dcpIn terfa ce.html.