Can multi-strategy management stabilize nitrate leaching under increasing rainfall?

The increased spring rainfall intensity and amounts observed recently in the US Midwest poses additional risk of nitrate (NO3) leaching from cropland, and contamination of surface and subsurface freshwater bodies. Several individual strategies can reduce NO3 loading to freshwater ecosystems (i.e. optimize N fertilizer applications, planting cover crops, retention of active cycling N), but the potential for synergistic interactions among N management practices has not been fully examined. We applied portfolio effect (PE) theory, a concept originally developed for financial asset management, to test whether implementing multiple N management practices simultaneously produces more stable NO3 leaching mitigation outcomes than what would be predicted from implementing each practice independently. We analyzed simulated data generated using a validated process-based cropping system model (APSIM) that covers a range of soils, weather conditions, and management practices. Results indicated that individual management practices alone explained little of the variation in drainage NO3 loads but were more influential in the amount of residual soil NO3 at crop harvest. Despite this, we observed a general stabilizing effect from adopting well-designed multi-strategy approaches for both NO3 loads and soil NO3 at harvest, which became more pronounced in years with high spring rainfall. We use the PE principle to design multi-strategy management to reduce and stabilize NO3 leaching, which resulted in 9.6% greater yields, 15% less NO3 load, and 61% less soil NO3 at harvest than the baseline typical management. Our results make the case for applying the PE to adapt NO3 leaching mitigation to increased climate variability and change, and guide policy action and on-the-ground implementation.


Introduction
Climate change in the United States (US) Midwest is leading to a trend of increasing rainfall, especially in spring and early summer [1][2][3][4][5], resulting in greater and more variable hydrological nitrogen (N) losses from agricultural soils [6]. Cropland in this region is already prone to nitrate (NO 3 ) leaching losses, owing to a combination of factors including the widespread use of N inputs, highly fertile soils that mineralize large amounts of N, long fallow periods with no plant N uptake, and the ubiquity of subsurface drainage systems [6][7][8]. Excessive rainfall not only provides the main physical driver for the loading of soil NO 3 into surface waters [9], but also has detrimental effects on crop plants. Even short-term waterlogging (saturated soil conditions) can damage roots, inhibit their growth and reduce overall nutrient uptake and productivity [10] with impacts comparable in magnitude to those produced by drought [11]. With greater rainfall, the frequency and severity of NO 3 leaching events is expected to increase [6,12,13].
Building resilience into crop production systems to increasing spring rainfall should not only focus on reducing the magnitude of NO 3 leaching losses, but also on producing stable (i.e. less variable) outcomes, as perceived efficacy can shape farmers' attitudes towards adopting nutrient loss mitigation practices [14]. Mitigation practices can be grouped into three general strategies, based on how they affect biophysical controls on plant-soil N cycling: (i) optimize crop N supply (e.g. adjust N fertilizer rate, timing and source; [15][16][17][18]); (ii) enhance the crops' N-sink capacity (e.g. earlier and longer growth cycle [19][20][21]), and (iii) improve retention of actively cycling soil N [22] (e.g. crop rotations [23,24] or cover crops [25]). Although field testing has shown that these strategies can reduce NO 3 leaching, their effectiveness varies substantially in time and space [26][27][28]. Because of the complexity of the crop-soil system, the impacts of weather perturbations are non-additive and may linger into subsequent growing seasons [12]. Hence, the effectiveness of a given practice depends on the system state (e.g. legacy soil NO 3 , crop residues) as much as the overall influence of weather and the interaction with the crop species. This is the reason often cited when explaining the large variability in effectiveness of NO 3 loss mitigation practices [22,24].
Well-established ecology and evolutionary theory postulates that greater species diversity leads to increases in the stability of biological systems [29][30][31]. This concept, often known as the portfolio effect (PE), has been ultimately borrowed from finance [31], where the diversification of financial assets is known to minimize risks and stabilize the variability of returns in investment portfolios [32]. More recently, PE theory has been applied in agriculture to examine the role of crop diversity in the adaptation of food production systems to climate change [33,34]. Similar to the stabilizing effects of asset, species or crop diversity, we hypothesize that multi-strategy management approaches lead to more stable NO 3 leaching mitigation outcomes under scenarios of increasing environmental variability. In other words, we expect that implementing multiple improved management practices simultaneously produces synergistic stabilizing benefits greater than implementing each practice independently. Similar hypotheses have been previously formulated [6,26], but to our knowledge, no direct evidence has been provided to support the PE theory in the context of adapting NO 3 loss mitigation strategies to the changing climate.
Portfolio effects are often evaluated by comparing the variability of individual entities (e.g. financial assets or species populations) to the variability of the aggregate portfolio-a construct derived from statistically averaging or summing its constituent features [35]. In our case, however, PE can be only tangentially examined through the influence of a given combination of management practices on the object of study. Concretely, the crop-soil system remains the same, and we only evaluate how the variability in outcomes (e.g. NO 3 leaching) change as we act upon it. This means that outcomes for both unimproved and improved management across a wide range of environmental variation must be quantified to estimate their differences in stability. This is difficult to do empirically, as conducting field experiments with enough treatments and environmental replication is expensive and impractical. Alternatively, outcomes can be approximated by mechanistically modeling the underlying processes that drive NO 3 losses. Cropping systems simulation models are routinely used in the Midwest to evaluate interactive effects of management, environment and genetics on crop yields and NO 3 leaching [7,12,[36][37][38][39].
Here, we use simulated data covering a range of soils, weather conditions, system states and management, to examine whether simultaneously implementing various NO 3 leaching reduction practices leads to changes in stability consistent with the PE. We then use the PE principle to design multi-strategy approaches that result in both low and stable NO 3 leaching under a gradient of increasing spring rainfall in the US Midwest.

Sites and data sources
Data from seven long-term experimental field sites in the Midwest (figure 1(a)), were used to configure, drive and test the APSIM cropping systems simulation model. The KELLEY and NASHUA sites have been described in detail in previous studies [7,38,39], while the remaining sites (DPAC, HICKS.B, GLIMORE, SERF and STJOHNS) were obtained from the Sustainable Corn CAP Research Database [40]. The sites fall within gradients of soil and climate characteristics, including: soil organic carbon (SOC; range 1.0%-2.7% averaged over 1 m depth; figure 1(b)), soil plantavailable water holding capacity (PAWC; range 70-132 mm integrated to 1 m depth; figure 1(b)), mean annual precipitation (range 711-1050 mm; figure 1(c)) and mean annual temperature (range 7.6°C-10.6°C; figure 1(c)).
Soil information for each site was retrieved from the SSURGO database [41]. Soils in these sites are deep, fertile, and artificially drained using subsurface drain tiles. Further details about the sites are included in the supplemental information (table S1 is available online at stacks.iop.org/ERL/14/124079/mmedia). Daily weather data (1987-2016) for the sites were retrieved from the Daymet dataset [42] using the single pixel extraction tool (downscaled to 1 km resolution). The obtained weather data at the sites reflected a trend of increasing April-June rainfall (figure 1(d)), which is consistent with literature reports for this region [2,43].

Simulation model and experiments
We used the Agricultural Production Systems sIMulator (APSIM; version 7.8) to conduct simulation experiments. APSIM is a cropping systems modeling platform that is composed by interconnected crop, hydrological, and nitrogen cycling process-based models. Using daily weather and user-defined soil and management information, the model calculates many soil-plant-atmosphere variables, including crop growth processes, soil water, soil temperature, and N and C cycling (for an in-depth description of APSIM see [44]). Before performing simulation experiments, we configured APSIM using soil, weather and management information of each site to replicate the longterm experiments that were performed at the sites. We compared these simulations against measured crop yield and NO 3 loads and concentrations in drainage measured at the sites to assess the robustness of the model outputs. Details on model configuration procedures and testing are provided by a previous study [45] and in the supplemental information (S1).
Following model testing, we designed a factorial simulation experiment to quantify the impact of five environmental factors and five management practices on NO 3 leaching and yield in corn (Zea mays L.) cropping systems (table 1). We characterized NO 3 leaching by examining two model output variables: annual (i.e. harvest-to-harvest) NO 3 loads in drainage water and the unused soil NO 3 at corn harvest. The former represents the main pathway of hydrological N losses from drained fields in the Midwest, while the latter is related to the risk of subsequent NO 3 leaching. The five environmental factors included (i) site (figure 1), (ii) weather-year (1987-2016), and soil-state initial conditions including: (iii) previous crop (residue type), (iv) soil NO 3 content in the fall, and (v) water table depth. Initial conditions were derived from spinup runs (see supplemental information S1). Soil conditions were reset to initial levels every 20-Oct (at harvest), which means that our annual cycle was harvestto-harvest. The five management practices evaluated included (i) genotype selection, (ii) planting date, (iii) fertilizer N rate, (iv) fertilizer N timing and (v) cover crop, each of which fall within one of the three general N reduction strategies. The combination of factor levels, sites, and weather-years resulted in~0.9 million simulated scenarios. Detailed descriptions of each simulation treatment and levels are provided in table 1.

Sensitivity analysis
We assessed the importance of the simulated factors on the model outputs with a variance-based sensitivity analysis approach [47]. This technique measures the sensitivity of model outputs to each input factor and interactions by approximating the proportion of factor or interaction sums of squares in relation to the total sums of squares. For full-factorial designs, factor variances are decomposed into orthogonal variance terms for main effects, two-way interactions, and so on, which are added to calculate the total variance. This is analogous to the classic analysis of variance (ANOVA) decomposition [48,49]. We computed first-order (i.e. main effect) sensitivity indices and attributed the residual variance to two-way or higher order interactions. These sensitivity indices were computed for all simulation factors (table 1) within each site.

Portfolio effect
We tested the existence of the PE under simultaneous implementation of improved practices by examining how their temporal (year-to-year) variation compared to the scenarios where only one improved practice was implemented (figures 2(a)-(b)). By single practice implementation, we mean that only one of the management factor levels was either 'good' or 'best' and all others were 'null' (see table 1). To do this across different levels of rainfall, we grouped weather-years into terciles of low, average and high April-June rainfall. Then, for the ith combination of rainfall class, site and management scenario, we computed the mean (μ i ) and variance (s i 2 ) of output variables. We assume that the variability of any theoretical single practice can be estimated using the log-log mean-variance linear relationship of single practices [35]. The PE theory predicts a stabilizing effect when the slope of this relationship (z) is less than 2 [34]. We extrapolated the mean-variance relationship to the observed mean of a multi-strategy scenario (figure 2(b)). Here, we expressed the PE of a multi-strategy scenario as the ratio of the observed standard deviation (s i ) to the expected standard deviation (ŝ ) derived from the extrapolated log-log mean-variance relationship. PE values less than 1 indicate that the multi-strategy scenario is less variable than what would be expected under a hypothetical single-practice scenario with the same mean. Details about the rationale and methodology of the mean-variance PE approach are provided by in the supplementary information (S2) and [35].

Scenario ranking
To identify multi-strategy approaches with low and stable NO 3 loads and soil NO 3 at harvest, we ranked scenarios according to their mean value and strength of the PE. This was done by computing a stability adjusted mean (m i PE ) as the product of m PE , i i ⁎ which was then averaged for each scenario and April-June rainfall class. Each output variable was ranked from best to worst (best being the smallest m PE ) separately, though grain yield was ranked only based on the mean value because no consistent z value (slope of the mean-variance relationship) was found for this variable (see supplementary information S2). A combined ranking for each management scenario was developed by computing the Euclidian distance from the origin in the three-dimensional space of each ranking ( 2 ). As before, the scenarios were ranked from best to worst (best being the smallest d i ).

Environmental factors are the main drivers of variation in crop yields, NO 3 loads and unused soil NO 3
The APSIM model was able to accurately reproduce the observed corn yields (RMSE=1.27 Mg ha −1 yr −1 ), and tile drainage NO 3 concentrations (RMSE=5.21 mg N L −1 yr −1 ) and loads (RMSE=10.9 kg ha −1 yr −1 ; supplemental figure S1.2). In figure 3(a), we show the first-order sensitivity indices (i.e. the share of the residual variance attributed to each simulation factor) and the aggregate of all higher-order interactions, which were derived from the factorial simulation experiment. This analysis indicated that simulated corn yields were highly sensitive to the weather-year, which accounted for 40%-82% of the total variation in yields ( figure 3(a)). Drainage NO 3 loads were also sensitive to weatheryear, although less so in the eastern sites where water table depth has a large effect. The reason for this is probably the low water holding capacity of these soils ( figure 1(b)) or the shallow depth of the drains at these sites (table S1). Such a large influence on these hydrological factors rendered NO 3 loads less sensitive to management than to the soil-state factors such as fall soil NO 3 content and previous crop, which reflect the soil NO 3 levels at the beginning of the simulation and conceptually represent the residual soil NO 3 from the preceding crop year. The previous crop also affected soil NO 3 content at harvest by changes in soil NO 3 immobilization-mineralization through differences in the quality and quantity of residues. In the simulation, the amount of unused soil NO 3 leftover corn harvest was very sensitive to N rate, and moderately to fall soil N and previous crop. Interactive effects account for less than a third of the variation in the simulated corn yield, drainage NO 3 loads and unused soil NO 3 across all sites ( figure 3(a)).
Springtime (April-June) rainfall was an important driver of NO 3 losses and unused soil NO 3 at corn harvest in the simulation ( figure 3(b)). Overall, NO 3 loads were generally positively correlated to spring rainfall, except for sites in Indiana and Ohio were the variance was dominated by initial water table level ( figure 3(a)). On the other hand, unused soil NO 3 was negatively correlated to spring rainfall across the board. The latter is explained by the greater NO 3 losses through drainage in years with high rainfall. Corn yields also tended to be positively correlated with spring rainfall, although the correlation was generally weak (i.e. <0.5) in all but a few sites-management scenarios ( figure 3(b)).

Portfolio effects of multi-strategy management are enhanced under increasing spring rainfall
Simulated NO 3 loads and soil NO 3 at harvest became less variable as more improved management practices are adopted (i.e. those labelled as 'good' or 'best' in table 1), which is consistent with the portfolio effect hypothesis ( figure 4). For instance, scenarios with both 30% less N applied in-season and overwintering cover crops are 5% and 33% less variable than their single practice equivalents (PE=0.95 and 0.67, respectively). Critically, the PE tended to become more pronounced under scenarios of higher spring rainfall. In contrast to NO 3 loads and soil NO 3 at harvest, we did not find support for the PE hypothesis on crop yields ( figure 4).
Across all scenarios with at least two improved practices and spring rainfall levels, PE ranged from 0.74-1.42 for NO 3 loads and 0.22-1.32 for soil NO 3 at harvest. The fact that PE is greater than 1 in some cases indicates that adopting any combination of improved management does not ensure that NO 3 losses become less variable. For example, adopting a winter-kill cover crop with high N demanding genotypes and early planting shows a destabilizing effect on soil NO 3 at harvest (PE=1.09), if these are implemented under unimproved N fertilizer management (i.e. 30% more N than the recommended rate applied in the fall). In some years, the winter-kill cover crop was able to retain some of the N applied, prevent it from leaching and release it in time for corn uptake. In other years, the release (N mineralization) occurred too early, resulting in high NO 3 loads and soil N at harvest.

Best multi-strategy management decreases NO 3 leaching variability with minimal effects on yields
From all 232 possible scenarios (see supplemental table S2), here we subset the 10 top-and bottomranked at each spring rainfall classification for comparison. The alluvial diagram in figure 5(a) offers a visual representation of the frequency of each management practice for the top-and bottom ranked scenarios. The top-ranked strategies are generally those that couple early planting, high N demand genotypes, low N rates applied at planting or in season, and cover crops. The top-ranked management strategies appear to change little across years with distinct rainfall ( figure 5(a)). The worst-performing multistrategy scenarios were those that combined normal N demand genotype, average planting, and winter-kill cover crops while applying high levels of N fertilizer in the fall ( figure 5(a)). This is probably caused by the destabilizing effect of the winter-kill cover crop following the fall N application.
On average, the 10 top-ranked scenarios resulted in 25% less NO 3 loads, 72% less soil NO 3 at harvest, and 17% greater corn yields than the 10 bottomranked scenarios ( figure 5(b)). When compared to a typical management situation (i.e. all 'good' practices in table 1, except with no cover crop), the 10-top ranked practices produced 9.6% greater yields, 15% less NO 3 load, and 61% less soil NO 3 at harvest. Finally, the top-10 scenarios reduced NO 3 loads (5.9%) and soil NO 3 at harvest (60%), but also yields (3.8%) when compared to the scenarios that produced the highest yields.

Discussion
Increasing springtime rainfall in the Midwest is accentuating the risk of severe NO 3 leaching losses and impacts to water quality. Climatological studies indicate that wetter springs are a result of more intense precipitation during each rainfall event, rather than an increase in the frequency of events [1,4]. These changes in rainfall patterns are leading to more pulses of water availability that concentrate NO 3 leaching events into so-called 'hot moments' [13], and potentially decreasing the effectiveness of in-field strategies to mitigate NO 3 leaching [6]. Indeed, our simulation experiments indicate that NO 3 loads are strongly sensitive to the combination of environmental (weather-year) and legacy soil-state (previous crop, fall soil NO 3 , water table) factors, whereas the influence of individual management practices is relatively small. Here, the soil-state factors are related to the amount of soil NO 3 and water with which each simulation was shaded area) indicate that the multi-strategy scenario is more variable (i.e. less stable) than a theoretical single practice with the same mean.
initialized, conceptually representing legacies from the preceding crop year. This may suggest that if high residual NO 3 remains in soils after harvest, there is little management can do to decrease NO 3 loads in the spring. After accounting for the initial soil-state, NO 3 loads are almost entirely dependent on year-to-year variation in weather in all but two sites ( figure 3(a)), which was well-described by the variation in cumulative springtime rainfall (April-June; figure 3(b)).
Conversely, the amount of unused soil NO 3 left over after corn harvest is more sensitive to management than to environmental and soil-state factors, indicating that management would have a larger influence in the potential NO 3 loads the following crop year. This dynamic may provide some basis for the oftenobserved temporal lags between management implementation and NO 3 load responses in annual cropping systems [51]. It may also suggest that measurements of deep-profile residual soil NO 3 after harvest rather than NO 3 loads may be a better indicator of the performance of N management practices on an annual scale. Thus, these types of data could prove useful in the development and validation of predictive tools for use in adaptive management [52].
However, it is worth noting that the amount of unused soil NO 3 at harvest is negatively correlated to springtime precipitation ( figure 3(b)), which may be related to higher NO 3 losses (leaching and gaseous denitrification) and crop yields in wet years. Thus, low residual soil NO 3 after harvest cannot be entirely attributed to better management without considering the wider effects on the system, given that the legacies from previous management and previous weather are inextricably intertwined [12].
Our results suggest that the magnitude and variability of NO 3 leaching outcomes can be reduced by adopting more complex management strategies. Building resilience into crop production is often framed in terms of biological diversification (either within or across crop species) as different strategies of distinct genotypes, on the aggregate, tend to produce more stable outcomes [31,33,34]. The same rationale applies in management. The different strategies here examined (i.e. optimize N supply, increase crop N uptake, retain actively cycling N) complement each other; when one strategy fails to reduce NO 3 leaching, another might succeed. The result is a general stabilizing effect of adopting multi-strategy approaches Figure 5. Designing multi-strategy management for low and stable NO 3 leaching and high yields. (a) Alluvial diagram indicates the frequency with which sets of practices are paired together in the 10 top-ranked (best case) and 10 bottom-ranked (worst case) multistrategy scenarios across three levels of April-June rainfall. Horizontal blocks represent each management practice, and the widths of the blocks and alluvia are proportional to the frequency of scenarios containing the practices. Further information on how to interpret alluvial diagrams is provided by [50]. (b) Relative changes in simulated yields, NO 3 loads and soil NO 3 at harvest when comparing the 10 top-ranked management scenarios to three baselines. Error bars represent the 90% range of the site-year data.
( figure 4). Importantly for climate change adaptation, while the best-performing combinations of strategies remain largely unchanged across different levels of rainfall ( figure 5(a)), their PE (i.e. stabilizing) effect is more pronounced in years with high spring rainfall (figure 4). Though this principle is well-established in other areas of crop management such as integrated pest management (IPM) [53], to our knowledge, this is the first study to provide evidence of the PE in this context.
Although PE theory appears to generally hold true for NO 3 leaching mitigation in our simulation experiment, we highlight a few cases where a certain combination of what we considered improved practices indeed resulted in a destabilizing effect (see section 3.2 and figure 4). This is important because practice implementation does not exist in isolation, and whether a practice contributes to improved outcomes depends on how it is integrated with other strategies. Because field studies often seek to elucidate the effects of one or two interventions, researchers must decide what other management factors to hold constant, which has implications for the interpretation of findings. For example, cover crops studies are often conducted assuming that other improved practices, such as applying N at optimum rates (MRTN) in the spring or in-season, are already in place. This detail rarely receives any attention and has been poorly communicated to farmers, practitioners and policy makers. The result is that virtually none of the current public or private cover crop cost-share programs include a requirement to also improve N fertilizer management [54,55]. As shown here, there are cases where adoption of cover crops without improving N fertilizer management (rate and timing) produces more variable, and even higher, NO 3 leaching (figure 2).
In contrast to NO 3 leaching mitigation, our results do not support a PE for crop management on corn yields, mainly because no consistent mean-variance relationship was found across the different sites, sets of soil-state factors and weather-year classes (see the wide range of z values in figure S2). One possible reason for this might be that the model is failing to capture some yield variation, likely because the negative effects of high rainfall on crop growth are being underestimated. The version of APSIM we used in this study included improvements that inhibit root growth and yield in the presence of shallow water tables [10]. Yet, factors that often accompany excessive rainfall, such as decreased plant populations, damage to aboveground tissues (e.g. wind, hail) and increased disease, are not currently considered, a known weakness of crop simulation models [11]. Because APSIM is a biophysical model, other potential biotic synergisms (e.g. reduced weeds and diseases and increase microbial function with cover crops) are also not considered. However, the most likely reason why we did not observe a PE on yield is because the structure of our simulation treatments, which focused on NO 3 leaching mitigation, is not necessarily the same as those that would tend to provide high and stable yields. For example, here we consider applying 30% less N fertilizer that MRTN as an improved practice in the context of NO 3 leaching mitigation, but in some years this could lead to N deficiency and reduced yields. Thus, we use mean yield instead of the PE-corrected mean yield to develop our multi-strategy rankings (see section 2.5). Compared to the highest-yielding scenarios, the 10 top-ranked multi-strategy management resulted in modest mitigation of NO 3 loads, large reduction in soil NO 3 at harvest, but also minor reductions in yield ( figure 5). This points to the tradeoffs that exist between yields and water quality.

Conclusion
The design of nutrient loss mitigation strategies has generally relied on static approaches, such as the average effectiveness of practices and crop yield tradeoffs across many research studies [16,21,24,26]. However, with climate change disrupting hydrological and biogeochemical cycling, there is large uncertainty on how these outcomes might change under increasing environmental variation. Rather than relying on individual practices with their own independent responses to weather, portfolio theory provides a framework for adapting mitigation NO 3 leaching by designing interventions that simultaneously manage multiple biophysical controls on N cycling in an attempt to distribute risk. In this study we found evidence of portfolio effects (PE) in well-designed multi-strategy management for reducing and stabilizing NO 3 leaching under increasing spring rainfall. The in-field management practices here examined are a small subset of possible NO 3 loading mitigation options available. Future PE analyses should include other strategies such as edge-of-field infrastructures or land-use change, though these tend to be less compatible with current production systems due to external socioeconomic influences (markets, policies, etc) [26]. Adopting multi-strategy mitigation approaches could also result in additional financial cost, though these determinations are outside the scope of this study. Therefore, next research steps should explore the costbenefit of multi-strategy approaches with more stable NO 3 mitigation outcomes, at the farm-level and society at large. Our results make the case for applying the portfolio theory to adapt NO 3 leaching mitigation to climate change, and guide policy action and on-theground implementation.
(Hatch projects: IOW03814, IOW04414), the Foundation for Food and Agriculture Research (project: Optimizing Agricultural Water Use), and the National Science Foundation (NSF Award No. 1842097). We thank Lori Abendroth (Sustainable Corn CAP database support) and the APSIM initiative (simulation model support). We also thank Ranae Dietzel and Mike Castellano for comments on earlier versions of this manuscript.

Data availability
The data that support the findings of this study are publicly available from the following sources: • The APSIM model can be downloaded at: https:// apsim.info/.
• The measured data for the Sustainable Corn CAP sites is available form National Agricultural Library: https://doi.org/10.15482/USDA.ADC/1411953.
• Computer code used to access the SSURGO database, process and write soil files with the APSIM format has been made available at the following repository: https://doi.org/10.5281/zenodo. 1467205 • Data from the simulation experiments has been made public at the following repository: https:// doi.org/10.5281/zenodo.3370029.