Riverine nitrogen footprint of agriculture in the Mississippi–Atchafalaya River Basin: do we trade water quality for crop production?

Increasing food and biofuel demands have led to the cascading effects from cropland expansions, raised fertilizer use, to increased riverine nitrogen (N) loads. However, little is known about the current trade-off between riverine N pollution and crop production due to the lack of predictive understanding of ecological processes across the land-aquatic continuum. Here, we propose a riverine N footprint (RNF) concept to quantify how N loads change along with per unit crop production gain. Using data synthesis and a well-calibrated hydro-ecological model, we find that the RNF within the Mississippi–Atchafalaya River Basin peaked at 1.95 g N kg−1 grain during the 1990s, and then shifted from an increasing to a decreasing trend, reaching 0.65 g N kg−1 grain in the 2010s. This implies decoupled responses of crop production and N loads to key agricultural activities approximately after 2000, but this pattern varies considerably among sub-basins. Our study highlights the importance of developing a food–energy–water nexus indicator to examine the region-specific trade-offs between crop production and land-to-aquatic N loads for achieving nutrient mitigation goals while sustaining economic gains.


Introduction
The hypoxic zone in the northern Gulf of Mexico is greatly influenced by the high nutrient (e.g.nitrogen and phosphorus) loads from the Mississippi Atchafalaya River Basin (MARB) (Turner and Rabalais 2003, Lohrenz et al 2008, Boesch et al 2009), which degrades biodiversity, impairs ecosystem function and threatens coastal economies (Altieri et al 2017).Various studies have examined the natural and anthropogenic drivers of streamflow and nutrient loading variations across the MARB.Climate variations, precipitation in particular, has been identified as an important contributor to the year-to-year variations in N loads through altering water discharge and biogeochemical processes (Battaglin et al 2010, Smits et al 2019, Lu et al 2020).Meanwhile, intensive agricultural production serves as the principal driver for the growing nutrient loads in riverine environments, primarily attributed to the expansion of agricultural lands (Lambin and Meyfroidt 2011) and widespread increases in nitrogen (N) fertilizer uses (Van Meter et al 2018, Oelsner and Stets 2019, Stets et al 2020).With cropland encompassing roughly 58% of the basin's land area, the MARB drains about 41% of the land area in the contiguous United States and yields an economic value exceeding $100 Billion (B), stemming from the farming and fishing sectors (Goolsby et al 1999).Collectively, these attributes make the MARB an ideal testbed to understand interplays within the food-energy-clean water nexus, and to quantitatively assess the relationship between water quality management and the attainment of sustainable food and biofuel production under climate changes (Gordon et al 2010).
Over the past several decades, cropland expansion required to meet the world's rising food and biofuel demand has increased the total N export (Seitzinger et al 2010).The increased cropland area is closely linked to the cumulative N input within agricultural systems (Lu et al 2019).It also determines the accumulation of available soil legacy N that is readily mobilized when crop production decreases (Lee et al 2016).In the MARB, both USDA census data and process-based ecosystem modeling (Dynamic Land Ecosystem Model (DLEM) in this case) show grain crop production has increased by 75% from the 1970s to 2010s, with increasing rates of 4.3-5 million metric tons (MMT) yr −1 (range of model estimate to survey average, p < 0.01, figure 1(a)).To meet the increases in crop production in the Basin, total N fertilizer input increased by about 75%, rising from 4.4 Tg N yr −1 (1 Tg = 10 12 g) in the 1970s to 7.7 Tg N yr −1 in the 2010s (figure 2).However, dissolved inorganic N (DIN) exports from the MARB to the Gulf, estimated by both water quality monitoring data (i.e. the USGS LOADEST, Weighted Regressions on Time Discharge and Season (WRTDS)) and the process-based modeling, demonstrate large interannual variations with a slightly decreasing trend during 1980-2017, except for a few peak years (e.g.2008 and 2017, figure 1(b)).We hypothesized that the decoupling between increasing agricultural N input and the 'near-flat' trend in riverine N load from the MARB would be caused by a looser connection between crop production and water quality.To test this hypothesis, in this study we used a process-based modeling approach and historical time-series datasets to quantify how key agricultural activities have regulated inorganic N loads while changing crop production (in the concept of riverine N footprint (RNF), i.e. the ratio of ∆N loads to ∆crop production) and how their impacts varied over space and time across the MARB.
To maintain the quality of the agroecosystems on which human society relies, modern agricultural activities seek the most sustainable synergy between increasing crop production and reducing N loads to the environment.Some of these activities are primarily used to increase crop production and crop N uptake.For example, genetic improvement of crop varieties is widely adopted to boost crop yield and maximize economic profits (Evenson andGollin 2003a, 2003b).Other activities, such as organic fertilizer application and crop rotation systems are more focused on reducing fertilizer input into the fields and lowering N loss to the environment (Ma et al 2007, Wei et al 2020).While previous studies have examined how basin-wide agricultural management affects N export to the northern Gulf (National Research Council 2009, Robertson and Saad 2013, 2021, Marshall et al 2018), a comprehensive analysis quantifying the trade-offs between N pollution and crop production, driven by key agricultural activities, has not emerged.This is partially caused by limited modeling capability in tracking the coupled carbon and nitrogen cycling in the plant-soil-water-river continuum, such as simulating crop production and riverine N loads at the same time.Additionally, it remains unclear how the agricultural activities driving the rise of crop production have contributed to riverine N loads among the sub-basins of the MARB.Insufficient understanding and quantification of food-energy-water conflicts hinders our ability to identify alternative practices and assess the related costs to arrive at these alternatives (Tian et al 2018).
The DLEM model we used in this study can simultaneously simulate plant physiology, biogeochemical and hydrological at a daily time step.DLEM can estimate the fluxes and pool sizes of carbon, N, and water in terrestrial ecosystems.At the same time, DLEM can estimate the transfer and delivery of carbon and nitrogen from terrestrial ecosystems to aquatic systems (e.g.streams, rivers, and lakes).In this version of DLEM, we emphasize cropspecific management practices.Agricultural management practices considered in this study include chemical N fertilizer use, manure N application, irrigation, tillage, tile drainage, crop rotation, and technology improvement on crop yield at various time steps (Cao et al 2018, Yu and Lu 2018, Yu et al 2018).Along with the management practice data, we use the prescribed input data, such as data on daily climate (average, minimum, and maximum temperature, precipitation, and shortwave radiation), annual land use patterns, monthly concentration of atmospheric CO 2 , and annual N deposition, to characterize the key environmental changes and to drive the model simulations.We also parameterize the model to capture the magnitude and historical trajectories of yield for multiple major crops, including corn, soybean, wheat, rice, cotton, barley, sorghum, and so on.DLEM is shown to capably capture the long-term trend and inter-annual variations in crop production and N loading in the MARB during 1970-2017 (figures 1 and S1-S7) (Liu et al 2013, Lu et al 2018, 2020, 2021, Yu et al 2019, Tian et al 2020, Zhang et al 2022).
By using a series of simulation experiments, we attributed the relationship between riverine N loading and crop production to four key agricultural activities from 1970 to 2019.The impacts of other drivers and point-source contributions have been excluded in this study as we keep them consistent among the simulations so they cancel out in calculating the impacts of drivers of interest.The agricultural activities examined in this study include synthetic N fertilizer application (referring to N fertilizer), manure N application, land use/land cover change (LUCC, including cropland expansion, abandonment, and inter-annual crop rotation), and crop technology improvement (referring to the crop genetic improvement in enhancing crop community photosynthesis rates and N uptake capability).More details can be found in Lu et al (2018).Crop technology improvement was incorporated into DLEM through two mechanisms: (1) an increase in the harvested amount of a crop based on a crop-specific time-series harvest index, and (2) improved productivity achieved by enhanced crop N uptake capacity.The key parameter values in regulating the model-estimated long-term trends of crop yield under changing climate and management practices were calibrated against the national crop yield records for each crop type from the USDA National Agricultural Statistics Service (NASS) crop databases (www.nass.usda.gov/index.asp).It is noteworthy that the impact of N fertilizer use examined in this study specifically refers to how fertilizer use rate change for each crop type has affected crop production and N loading across the MARB.The paired model simulations (comparing with and without fertilizer use change) maintain consistency in cropland area dynamics, and their differences solely reflect the influence of historical fertilizer use rate changes (table S1 in supplementary material).In our simulation design, we disable management practice alterations in the paired experiments involving scenarios with and without LUCC.This allows us to isolate the effect of LUCC and exclude the influence of management practices on the newly cultivated or abandoned croplands.Moreover, as we incorporate annual crop type maps into our LUCC input, the estimation of LUCC impacts on crop production and N load encompasses N fixation alterations caused by area change of legume crops.In determining the N fixation of crops such as soybean, alfalfa, and other nitrogenfixing varieties, we simulate their annual dynamic productivity across various climate, soil, and management conditions and utilize crop-specific N-fixing parameters.
We depicted the trade-off using a concept of RNF of grain crop production (i.e. the ratio between the change of N load and the change of total grain crop production associated with the aforementioned agricultural activities, where the unit is g N kg −1 grain).Using the RNF as an N footprint indicator, we aim to investigate how agricultural activities in the MARB have contributed to altering water quality and grain crop production since 1970.It is noteworthy that these agricultural activities also play an important role in affecting the production of 'other non-grain crops,' which are excluded from the RNF estimation in this study due to their relatively small shares of cropland area and anthropogenic N input.Positive RNF values mean that crop production and riverine N load are changing 'concurrently' , while negative values indicate 'counter-current' changes between crop production and riverine N load.Under the circumstance of concurrent changes, water quality deterioration (i.e.positive N load change) occurs while crop production reaches a net gain (i.e.positive production change), and vice versa (i.e.negative changes in both riverine N load and crop production).We also quantify the economic revenues from major crop types driven by these activities, linking monetary gains with environmental costs.The results reflected by our study contribute to quantitative insights into food-biofuel-clean water conflicts that may be applicable in other agriculture-dominated river basins across the globe.

Model simulations
We used the DLEM model to quantify N leaching to local waters at each simulation grid as N yield (in a unit of g N m −2 d −1 ) and accumulated N delivered to rivers and coastal areas at the river outlet grids as N load (in a unit of Tg N d −1 ).Daily estimates were aggregated to monthly and annual total for analysis and comparison purposes.The N considered in DLEM includes dissolved organic N (DON), dissolved inorganic N (DIN: NO 3 − N and NH 4 + N), and particulate organic N (PON).We only quantified DIN loading in this study, which accounts for about 70% of the total N load to the Gulf of Mexico (Dominguez-Faus et al 2009).Annual DIN load at the river outlet to the Gulf was also validated by comparing DLEM estimates with the USGS LOADEST (the software is made available by the USGS for estimating constituent loads in streams and rivers based on a regression model given a time-series of streamflow, additional data variables, and constituent concentration, https:// water.usgs.gov/software/loadest/)and WRTDS simulation results from Stackpoole et al (2021).For both LOADEST and WRTDS, we combined NO 2 -NO 3 and NH 3 from the datasets to obtain annual DIN loading.N loading data from WRTDS and LOADEST are strongly correlated at an annual scale (R 2 = 0.94).The Nash-Sutcliffe model efficiency coefficient and Percent bias (PBIAS) between DLEM estimates of annual DIN loading and the WRTDS data are 0.61 and 4.5%, respectively (figure 1(b)).The calculations were performed using the 'hydroGOF' R package (Zambrano-Bigiarini 2017).We set up a series of counterfactual model simulation experiments by diminishing the change of driving factors one at a time, and compare them with the 'best-estimate run' in which all the driving forces vary over time (table S1).Their differences are used to quantify the changes in crop production and riverine DIN loading in response to changes in anthropogenic agricultural drivers in the MARB.Details of the DLEM model, model drivers, model calibration and validation, and simulation design can be found in supplementary material.

Examining the relationship between N balance and N loads
Based upon the existing data, we quantified annual anthropogenic N inputs, outputs, and N balance across the MARB during 1970-2017, and investigated the relationship between N balance and flow-normalized N loads during the same period.Anthropogenic N inputs considered in this study include atmospheric N deposition, synthetic N fertilizer use, manure N application, and N fixation by legume crop cultivation.The total N fixation amount is estimated based on the DLEM-estimated yield of legume crops (e.g.soybean, alfalfa, etc) that has been validated against the USDA NASS database.The remaining N input data are derived from the timeseries gridded database we developed in previous work for characterizing how human activities have altered N cycling and forcing the ecosystem modeling (e.g.Cao et al 2018, Bian et al 2021).N outputs in this study include crop-harvested N and gaseous N emissions.The annual amount of harvested N is obtained from the NuGIS (Fixen et al 2012), which comprises 21 crops (alfalfa, apples, barley, dry beans, canola, corn for grain, corn for silage, cotton, other hay, oranges, peanuts, potatoes, rice, sorghum, soybeans, sugar beets, sugarcane, sunflower, sweet corn, tobacco, and wheat) during the period 1987-2017.We calculated the quantity of harvested N from 9 major crops grown in the MARB since 1970 using the methodologies detailed in Zhang et al (2021), relying on county-level crop-specific yield and harvested acreage data as reported by NASS.Subsequently, we utilize the ratio of the harvested N by the 9 major crops to that of the 21 major crops reported by the NuGIS that is available during 1987-2017 to estimate the harvested N from the 21 major crops for the period preceding 1987.Nitrogen emissions in this study are determined by three pathways, namely direct N 2 O emissions (1% of total anthropogenic N inputs), N 2 emissions resulting from denitrification (1.7 times the direct N 2 O emission, Sabo et al 2019), and indirect N 2 O emissions due to N leaching (1.1% of the total regional N leaching estimated by DLEM in this study, Hergoualc'h et al 2019).All the N input and output variables are processed within the MARB region.The flow-normalized TN (total N) loads in the MARB are obtained from the WRTDS model (Stackpoole et al 2021) to reflect the anthropogenic trend in riverine N loads with random streamflow changes removed.The detailed data can be found in Lu et al (2023).

Definition of partial RNF of grain crop production
Based on the above experiments, we quantified how agricultural activities (including land use changes and extensive management practices) have simultaneously altered riverine N loading and production of eight major grain crops in the MARB since 1970.We examined the relative contributions of key agricultural activities to changes in N loading and grain crop production across the Basin.Due to the lack of detailed crop type, distribution, agronomic and management information, the production of non-grain and other crops were excluded in this analysis, although their production contributed to N yield and loading to the Gulf as well.Therefore, the RNF could partially reflect the water quality cost of per unit grain production gain under historical agricultural resource (e.g.land, nutrient) allocation since 1970.We also summarized the sub-basin RNF on a decadal scale for comparison purposes.To identify the N reduction potential of each agricultural activity, we defined the partial RNF as: where N footprint,i is the RNF driven by land use or management practice i, ∆N loading,i is N loading change driven by each activity, and ∑ i ∆Production i is the net change of crop production driven by all four activities considered in this study.The concurrent changes in RNF indicate that either water quality is sacrificed while pursuing a crop production gain or is improved at the cost of losing crop production.Under counter-current changes, however, water quality improvement (i.e.negative N load change) can be accompanied by crop production gain (i.e.positive crop production change), such as the impacts of crop technology improvement.The other possible scenario under this circumstance is that water quality deterioration (i.e.positive N load change) is accompanied by crop production loss (i.e.negative crop production change), which likely occurs under extreme climate conditions.In our analyses, a net crop production gain is seen in most cases.However, changes in crop production during the 2000s in the Lower Mississippi sub-basin (LM) is close to zero, which results in an extraordinarily large value of N footprint.Therefore, to avoid providing a misleading interpretation of RNF, we excluded the calculations for the LM in the 2000s.Furthermore, our analysis did not consider the drainage area changes over the past decades in the sub-basin within the MARB.

Estimation of crop-specific revenue changes
We calculated the revenue change of the eight major grain crops within the MARB, driven by the four agricultural activities.First, we obtained the statelevel crop-specific price data from the USDA NASS survey database.To make sure the price in one year is comparable to the price in another year, we deflated the price data to remove the inflation factor using the Gross Domestic Product Implicit Price Deflator (https://fred.stlouisfed.org/series/GDPDEF)with 2010 treated as the base year.We then multiplied the model-estimated annual crop production change associated with the four agricultural activities, by the deflated price data for each crop in each state.The product was treated as the annual crop-specific revenue change given the assumption that all the crop products were traded in the market.Then we summarized the sub-basin total revenue change on a decadal scale for comparison purposes.

Historical patterns of agricultural activities in the MARB
In the MARB, total cropland area averaged 90.3 ± 3.7 Mha (mean ± SD, SD is quantified to reflect year-to-year variations) with little interannual variations and a relatively stable pattern (trend of 0.99 Mha decade −1 ) during 1970-2021.Among them, the eight major grain crops accounted for 73% ± 5% of the total cropland areas in the region, exhibiting a significant increasing trend of 3.6 Mha decade −1 (p < 0.05) over the past five decades.In addition, total synthetic N fertilizer input averaged 6.2 ± 1.2 Tg N yr −1 , among which 81% ± 5% were applied for the production of these eight major crops.The amount of synthetic N fertilizer received by all the crops and these eight major crops demonstrates a similar increasing trend (0.73 vs 0.61 Tg N decade −1 , p < 0.05).The total area of the eight major crops increased by 16 Mha (27% of the 1970s average) over the past five decades.Five out of the eight major crop types presented an increase in their planting areas (figure 2(a)), with corn and soybean each increasing by ∼10 Mha, followed by spring wheat (0.58 Mha), rice (0.35 Mha), and durum wheat (0.16 Mha).In contrast, the planting areas of sorghum and barley decreased by 3 Mha and 1.2 Mha, respectively, followed by winter wheat (-0.5 Mha).Overall, the land and nutrient resources invested to these major crop types grew substantially during the past five decades although the total cropland area in this region remained stable.
We found that among the six sub-basins, the expanded croplands are primarily found in the Missouri (MO, 8.5 Mha) and Upper-mid Mississippi (UM, 5 Mha) River Basins.Specifically, the corn area in the MO and the UM increased by 6.2 Mha and 3.8 Mha, respectively (figure 2(a)).Nitrogen fertilizer input increased by 1.0 Tg and 0.7 Tg in the MO and the UM, respectively, which was mainly driven by the corn area expansion (figure 2(b)).Soybean  fertilizer use, and manure N application in croplands, have increased from 9.5 Tg N yr −1 in the 1970s to 14.4 Tg N yr −1 in the 2010s, with an increasing trend of 0.12 Tg N yr −2 (R 2 = 0.89).On the other hand, the total N outputs, including crop harvested N and N emissions, have increased from 8.2 Tg N yr −1 to 13.1 Tg N yr −1 during the same period, exhibiting a similar increasing trend but larger inter-annual variations (0.13 Tg N yr −2 , R 2 = 0.79).Despite large year-by-year fluctuations in the annual N balance, the smoothed N balance shows a clear trend with an initial increase since 1970, followed by a continued decline after the mid-1980s.We find this trend corresponds well with the flownormalized TN load (figure 3(b)), an indicator that removes the random effects of streamflow change, and therefore, can better represent the anthropogenic trend of the land N pollution to water bodies.However, there is also evidence of decoupling between them: the terrestrial N balance reaches its peak a few years later than the flow-normalized TN load (1988TN load ( vs 1983)); furthermore, while the N balance continues to decline thereafter, the flow-normalized TN load remains stable and has even increased slightly in the most recent decade.This implies that the rise in land N balance consistently contributed to deteriorating water pollution in the Gulf prior to the mid-1980s, and the subsequent decrease in N balance correlates with the enhancement of water quality during the mid-1980s to mid-1990s.Nevertheless, beyond that time frame, the decline in N balance, driven by augmented crop uptake, has had a diminished impact on further improving water quality.This finding is consistent with the relationship between N balance and flow-normalized TN loads reported by Stackpoole et al (2021).

Impacts of agricultural activities on N loading, grain crop production, and RNF
Our model attribution indicates that the major agricultural activities considered in this study together led to a net crop production increase, ranging from 43 MMT yr −1 in the 1970s to 107 MMT yr −1 in the 2010s across the Basin (figure 4(a)).The impacts of LUCC on crop production changes varied a lot during the study period, with its contribution ranging from 12% in the 1990s to 86% in the 1970s.However, apart from 1970s, crop technology improvement was identified as a primary contributor, leading to 29%-56% of the net change in crop production.Meanwhile, the growing usage of N fertilizer contributed to 18%-36% of the increase in this region but manure N application played a trivial role.
Similar to the trend in flow-normalized N load (eliminating random streamflow effects, figure 3(b)), the changes in model-estimated riverine N load due to these agricultural activities (human impacts) also demonstrate a pattern of initial increase followed by leveling off, spanning the last 50 years (figure 4(b)).We find that agricultural activities together drove a net increase of 0.07 Tg N yr −1 in N load in the 2010s.Among these activities, N fertilizer use plays a dominant role in leading to N load increase ranging from 0.02 Tg N yr −1 in the 1970s to 0.16 Tg N yr −1 in the 2010s, while LUCC grows to become an important contributor next to fertilizer input since the 1980s, with its contribution growing by over threefold from 0.03 Tg N yr −1 in the 1980s to 0.1 Tg N yr −1 in the 2010s.Crop technology improvement, however, decreased N load, and its 'mitigation' effects on N load increased by over tenfold from −0.018 Tg N yr −1 in the 1970s to −0.19 Tg N yr −1 in the 2010s.Our simulation indicates that manure N application in the croplands had a negligible impact on N loading changes in the Basin, compared with other driving factors.This is possibly caused by the fact that we only consider the part of manure N applied in croplands, rather than the total manure N produced by the livestock sector.Globally, in 2017, the estimated ratio of manure N application to production stands at 0.18 (Bian et al 2021).
The model estimation shows that the RNF (the ratio ∆N loading /∆Production; see methods for details) of grain crop production in the MARB has increased from −0.28 g N kg −1 grain in the 1970s to its peak, 1.95 g N kg −1 grain, in the 1990s, followed by a two-decade decline to 0.65 g N kg −1 grain in the 2010s (figure 4(c)).This implies that the increases in Basin-wide crop production were not necessarily associated with a proportionate degradation in water quality during recent two decades.On the other hand, the four agricultural activities show varied impacts on the RNF.The N fertilizer-induced RNF has increased six-fold, from 0.49 g N kg −1 grain in the 1970s to 2.94 g N kg −1 grain in the 1990s (figure 4(a)).It has declined and then remained stable at approximately 1.5 g N kg −1 grain in the most recent two decades.This could be linked to the observation that the rate of increase in synthetic N fertilizer use for each crop has decelerated in the last 20 years (Cao et al 2018), and the rise in N loading due to fertilizers has been less than the overall increase in crop production since the 1990s (figure 4).Notably, the LUCC impacts on the RNF has reversed from −0.33 g N kg −1 grain in the 1970s to 0.67 g N kg −1 grain in the 1980s.Likewise, the LUCC-induced RNF also peaked in the 1990s (2.01 g N kg −1 grain), and then declined to 0.93 g N kg −1 grain in the 2010s.The impacts of crop technology improvement on reducing RNF demonstrates a similar 'V-shape' trend with its peak in the 1990s (−2.89 g N kg −1 grain), followed by a decline to −1.78 g N kg −1 grain in the 2010s.This indicates its efficiency on improving water quality diminished during the most recent decades.Together, these shifts imply a disconnection between the rising rates of crop production and N loading caused by the four agricultural activities since the 1990s.

Sub-basin impacts of agricultural activities on RNF
We find that the RNF in the UM River Basin has increased from a negative value (−1.3 g N kg −1 grain, represented by the black point in figure 5) in the 1970s to 2.6 g N kg −1 grain in the 1990s, and then decreased to 0.54-0.66g N kg −1 grain in the past two decades (figure 5).In this sub-basin, N fertilizer input has consistently increased the N footprint, with the largest contribution found in the 1980s (5.2 g N kg −1 grain), followed by a substantial decline thereafter.Crop technology improvement has constantly reduced the N footprint, with its peak impact in the 1980s.In contrast, LUCC impacts on the RNF shifted from a negative value in the 1970s to positive values thereafter, implying a complicated land use changes at the sub-basin scale over time.However, starting from the 1990s, LUCC and N fertilizer inputs have played a comparable role in augmenting the RNF, and their impacts have decreased in tandem.Crop production increase in the UM River Basin accounts for approximately one-third to half of the MARB total in the past half century (figure S8).The decreasing RNF in this region since the 1990s indicates that enhancements in crop production have resulted in a diminishing impact on water pollution.
The Missouri River Basin (MO) demonstrated a similar pattern of RNF as in the UM, increasing from a negative RNF (−1.3 g N kg −1 grain) in the 1970s to its peak (2.4 g N kg −1 grain) in the 1990s, followed with a decline to 1.5-1.6 g N kg −1 grain in the last two decades.Similar to the UM, in the MO River Basin since the 1990s, LUCC and N fertilizer inputs have played a comparable role in increasing the N footprint.Nevertheless, the reduction in N footprint due to crop technology improvement has remained relatively consistent over time in this region.Crop production increase in the MO accounts for 23%-47% of total rise in the MARB during the study period (figure S8).The declining N footprint, in this context, indicates a loose connection between grain production gain and water pollution.The RNF in the Arkansas (AR) River Basin has demonstrated a similar inverted V-shaped pattern, increasing first and then declining.However, in this region, the RNF dynamics have been predominantly affected by LUCC and fertilizer inputs, while the effects of crop technology improvements have been much smaller.Despite the incomplete trajectory of RNF estimation in the Lower Mississippi River Basin, we can still identify a decreasing trend in RNF following the initial rise.The similar RNF patterns in these sub-basins suggest that although boosting crop production, the expansion of cropland and increased N fertilizer use have escalated the RNF over the first three decades.However, this trend has waned over the most recent two decades, partially due to the fact that the response of N load to these agricultural activities has become less rapid than the response of crop production.
An encouraging pattern is found in the Ohio River Basin (OH, figure 5).Although the RNF started at a high level (i.e. on average 7.5 g N kg −1 grain in the 1970s), it decreased to a near-zero value (0.04 g N kg −1 grain) in the 2010s.This notable decrease was primarily propelled by the reduced pollution sources arising from both LUCC and N fertilizer.Collectively, the impacts of LUCC and N fertilizer on RNF declined from approximately 15 g N kg −1 grain in the 1970s to around 5 g N kg −1 grain in the 2010s.We find that LUCC and N fertilizer have increased N load by only 0.09 Tg N yr −1 and 0.06 Tg N yr −1 over decades, while they have raised crop production by 11.2 MMT yr −1 and 8.0 MMT yr −1 , respectively (figure S8).Meanwhile, crop technology improvements have exerted significant impacts on reducing the RNF, ranging from −4.8 g N kg −1 grain in the 2010s to −9.6 g N kg −1 grain in the 1990s.Even though the rise in crop production within the Ohio River Basin comprises only 10%-17% of the overall increase in the MARB, the steady reduction in its RNF suggests a promising path toward sustainable agriculture and a diminishing environmental footprint.
Unlike other sub-basins, the Red River Basin demonstrates a growing RNF, with its peak of 14.1 g N kg −1 grain in the 2010s.Even though the Red River Basin contributes a smaller proportion to the overall increase in crop production (<5% of the MARB total in the recent three decades), the increasing RNF serves as a warning that even modest crop production gains might come at the expense of deteriorating water quality in certain regions.Although the changing trend in the RNF varied substantially across the MARB (central map in figure 5), it aligns well with the sub-basin patterns (centered map in figure 5).

Interpretation of modeling limitations and uncertainties
In this study, the model-estimated concurrent or counter-current change in crop production and N loads influenced by agricultural activities and the consequent RNF is different from the latent process that was investigated and reported to be important in reducing TN load since the mid-1980s (Stackpoole et al 2021).Their study argues that increased N retention in watershed, as well as slowed increase rate of N balance, have reduced TN loads in recent decades.This is consistent with what we find through crop-specific N balance data analysis and counterfactual modeling assessment (e.g. the declining N balance since 1988 in figure 3, and decreasing RNF since the 1990s in figure 4).From the perspective of the crop N budget that was entirely estimated by the USDA census data, our earlier research (Zhang et al 2021) also revealed that the crop N surplus at the national level (referring to the total N input not utilized by field crops) exhibited a relatively consistent pattern from the 1980s to the 2000s within US cropping systems.Subsequently, a reduction occurred during the 2010s, bringing the surplus to a level akin to that of the early 1970s.Despite concentrating solely on crop N budgets, Zhang et al (2021) mirrors a trend comparable to our study, underscoring the stabilization or reduction in N pollution to the environment within the principal cropping systems of the US over the recent decades.
Mounting evidence indicates that effective land management and sustainable agricultural practices in upstream watersheds can significantly improve downstream water quality (Vitousek et al 1997, Blann et al 2009, Yang et al 2016).Upstream conservation programs in the MARB have targeted reductions in N moving to surface waters through adopting best management practices (BMPs) (e.g.cover crops and side-dressing N application) and increasing crop N use efficiency.But our study aims to investigate how crop production and N loads simultaneously respond to a given agricultural activity change, in which the impacts of BMPs such as prairie strips, planting of cover crops, are not examined.For example, we include historical tillage records and tile drainage as model inputs to drive the DLEM for historical simulations (details can be found in the appendix and Lu et al 2022).However, the impacts of these practices are not teased out and quantified in this study.In addition, the model performance in estimating water discharge and DIN loads vary among sub-basins (figures S6 and S7).This variability can be attributed in part to the following factors: (1) limited availability of input data: the accuracy of our model relies on the quality of input data used for driving it, such as crop area and distribution maps, N fertilizer utilization, and manure N application maps.However, these data sources are constrained by the extent of state-level crop-specific surveys.Consequently, there might be discrepancies between the data and the actual land and nutrient management practices within the boundaries of individual sub-basins.( 2 Future climate change and more frequent extreme climatic events are projected to increase N export from the basin, even with hypothetically unchanged N fertilizer inputs and cropland expansion in the scenario testing (Zhang et al 2022).While this study does not explicitly quantify the effects of historical climate variations and changes in atmospheric CO 2 concentrations, their interplay with agricultural activities has been considered.These factors influence modelbased estimations of crop growth and nitrogen loading by modulating the availability of resources such as radiation, light, water, and CO 2 , as well as influencing biogeochemical processes.We acknowledge that models used for estimation and attribution have various sensitivities to environmental changes such as rising temperature, increasing atmospheric CO 2 concentration, N fertilization, irrigation.The modeling approach has limitations in accurately quantifying the impacts of individual factors due to limited knowledge on the complex process, interactions, and confounding factors (Ruehr et al 2023).In this case, the DLEM estimates of crop production and N loads have been calibrated and validated against long-term monitoring, measurement, and survey data across the MARB and the country (figures 1 and S1-S7).The model-estimated responses to LUCC, agricultural N inputs, and crop technology improvement have been cross-validated with other studies in our previous work (Lu et al 2018, 2020, 2021, Yu et al 2019) and averaged at the decadal level in this study to reduce the influence of inter-annual variations over a halfcentury study period.To constrain modeling results, in terms of factorial contribution assessment, more studies are needed to provide long-term measurement data under paired experiments (e.g.control and treatment) across various climate and soil conditions to inform and evaluate modeling work.

Implications for future nutrient management across the MARB and beyond
Agricultural practices associated with corn cultivation are the major cause of N load in the U.S. agricultural systems (Broussard and Turner 2009).As a result, the expansion of corn would significantly increase the total N load in the river systems (Ni et al 2021).At recommended N fertilizer use rates, corn-soybean rotation systems would lead to 14%-36% lower NO 3 -N concentrations than in continuous corn systems (Helmers et al 2012).However, soybean cultivation may lead to a higher phosphorus (P) yield than corn cultivation (Ni et al 2021).Therefore, future research efforts are needed to evaluate trade-offs between N loads and P loads in rotation systems.
The results of this study provide evidence that LUCC, crop choices, and follow-up nutrient management play a key role in determining if a region improves water quality or not while maintaining or increasing crop production.In the MO and the UM, although cropland expansions, combined with the widespread use of fertilizer, increased crop production, they came at the cost of water quality degradation (like the trend discussed in Ramankutty et al 2018).Nevertheless, in the OH, increased crop production was accompanied by a reducing riverine N load during the study period (figure S8).These divergences were due to grain crop production increases caused by different crops and different drivers among sub-basins.For example, corn production change in the MO, rising from 0.1 MMT yr −1 in the 1970s to 27 MMT yr −1 in the 2010s, accounted for 80% of the total crop production increase over the study period (figure 6).Corn production increase in the UM went up to 25 MMT yr −1 in the 2010s, accounting for 83% of the total change.However, soybean showed secondary increases over decades in these two sub-basins, with a net increase of 11 MMT yr −1 (28%) and 8 MMT yr −1 (23%) in the MO and the UM in the 2010s, respectively.These results indicated that corn production consistently dominated and had a large share of the grain crop production increase in the Midwest regions, which were associated with water quality deterioration in the U.S. Midwestern basins in the first three decades.However, primarily driven by crop technology improvement, the rise in crop production during the most recent two decades was not accompanied by a proportionate N loading increase in these two basins (figure S8).
The changing trends we find in other sub-basins indicate that a reduction in RNF is achievable through properly managing agricultural lands where high N fertilizer-consuming crops (e.g.corn) used to dominate the crop shares.For instance, grain-crop production increase in the OH rose from 3 MMT yr −1 in the 1970s to 18 MMT yr −1 in the 2010s (figure 5).Among crops, while corn's contribution to the overall production change increased from 54% in the 1970s to 60% in the 2010s, the proportion of soybean also rose significantly from 29% to 43%, with negligible or negative contribution from the remaining crops.The significant rise in soybean production share also elucidates the relatively stable synthetic fertilizer inputs in the OH since the 1980s, despite the ongoing increase in total crop production (figures 2 and S8).Based on the NASS database, our previous study found that the total harvested area of corn in Ohio, Indiana, Tennessee, and West Virginia, which consist of the majority of the Ohio River Basin, increased by only 1% in the 2010s, compared with the 1970s (Lu et al 2019).Such changes in crop choice and N input align well with the detected RNF decrease in the 2000s and the 2010s in this sub-basin (figure 5).Furthermore, we noticed that the cropping system in the LM experienced an alteration from corn-dominated to soybean-and ricedominated production over decades (figure 6).These changes in crop production indicate that, although corn-planting-induced N loss was likely replaced by rice-planting due to their high N demands, the elevated share of soybean production still reduced N input and led to a 'near-neutral' N loading change (figure S8).
Economically, corn contributed 53% (US$ 5.4 B yr −1 ) and 64% (US$ 6 B yr −1 ) of the total revenue changes of the eight major grain crops in the MO and the UM in the 2010s, respectively (figure 6(b)), which are lower than its shares of grain production.In contrast, soybean-induced revenue change was about 44% (US$ 4.4 B yr −1 ) and 38% (US$ 3.5 B yr −1 ) in the MO and the UM, respectively, exceeding its share in grain production in the same decade.In the OH, soybean-planting induced revenue change in the 2010s (i.e.US$3.1 B yr −1 , 63%) has already exceeded the contributions of corn (US$ 2.0 B yr −1 , 39%), suggesting that soybean has emerged as the primary crop for increasing economic income.Soybean's price is higher than corn, although its yield per unit area is lower.Also, soybean requires less cost (e.g. the need for a smaller amount of fertilizer).Currently, boosted by biofuel prices, the share of continuous corn is increasing over time in the U.S. Midwest (Wang and Ortiz-Bobea 2019).Therefore, soybean has the potential to serve as a key cash crop in the future to maintain the trade-off between water pollution and farmers' economic profits.Beyond managing crop choices, we may target adopting other restoration practices, such as restoring less productive agricultural lands to natural lands (Wu et al 2013, Cheng et al 2020), using slow-releasing fertilizers in fields, replacing monoculture with polyculture (Crews et al 2018), and postponing fertilizer input to meet plant nutrient demands (Lu et al 2020).

Outlook
Our results indicate that the temporal changes in crop production and N loads in response to the major agricultural activities became decoupled after 2000.This transition led to a reversal of the trend in RNF, shifting from an increasing to a decreasing pattern (figure 4(c)).The findings suggest that cropland expansion, increasing share of corn and soybean, and elevated anthropogenic N input before 2000 have consistently sacrificed water quality for higher crop production to meet food and biofuel demand, which is supported by earlier work (Foley et al 2005, Daniel et al 2010, Secchi et al 2011, Lark et al 2022).However, after 2000, crop technology improvement has become more effective in reducing N load as well as increasing crop production than LUCC and N fertilizer input, which on average led to a reducing RNF over the past two decades (figure 4(c)).Our result implies that N loading has changed at a slower pace under the agricultural resource use and management efforts in pursuing the same unit grain production in the most recent two decades, which is an encouraging signal for water quality improvement.
Local stakeholders should raise awareness of environmental degradation while managing agricultural land to achieve long-term environmental sustainability (Handmaker et al 2021).Farmers' choices over what to plant will be driven by markets and policies.However, influenced by the frequent extreme weather conditions in the largest crop-producing countries (e.g.Brazil: the 1st soybean-producing and 3rd largest corn-producing country), the market effects of trade wars, the lingering supply issues induced by coronavirus pandemic, and food crisis due to the Russia-Ukraine (the 6th largest cornproducing and the 8th largest soybean producing country) conflict, crop prices may become more unpredictable in the years to come.Therefore, it is difficult to predict how the choices of crops will change regionally and globally.These uncertainties underscore the significance of comprehending the intricate trade-offs between food and biofuel production and environmental quality, given agricultural management and economic event-induced market effects.Our conclusions drawn here have important implications regarding land and nutrient resource management for achieving long-standing policy goals for agricultural and environmental sustainability.

Figure 1 .
Figure 1.Annual total grain crop production and its trend as estimated by the Dynamic Land Ecosystem Model (DLEM) and USDA survey (a), and annual DIN load estimated by DLEM, the USGS LOADEST, and Weighted Regressions on Time, Discharge, and Season (WRTDS) (b) in the MARB during 1970-2017.Note that the LOADEST N load data starts in 1980 and WRTDS starts in 1976 due to the limited data availability.The post-1980 estimates of DIN loads show a near-flat long-term trend with large inter-annual variations (slope of DLEM estimates = −0.0005,p = 0.90; slope of LOADEST = −0.0026,p = 0.46; slope of WRTDS = −0.0024,p = 0.48).

Figure 2 .
Figure 2. Annual areas of eight major grain crops and total cropland (a), and total synthetic N fertilizer and manure N application (b) from 1970 to 2017 in the sub-basins of the MARB.

Figure 3 .
Figure 3. Annual time-series N inputs, outputs, and N balance (a) and comparison between N balance and flow-normalized N load (b) in the MARB during 1970-2017.Nfix, Ndep, and Nfer in figure (a) stand for N fixation, N deposition, and synthetic N fertilizer use, respectively.N manure refers to manure N application in the croplands.Smoothed Nbal refers to smoothed N balance derived from 6-year moving average.

3. 2 .
Anthropogenic N balance on land and riverine N load To further investigate how terrestrial anthropogenic N balance affects N loads from the MARB, we compiled the time-series data of N inputs, outputs, N balance from model input database, modeling estimates, and survey-based crop N budget data (Fixen et al 2012, Zhang et al 2021), as well as the flownormalized N load derived from Stackpoole et al (2021).Our data analysis (figure 3(a)) indicates that the anthropogenic N inputs received by the MARB land surface, including N fixation by legume crop cultivation, atmospheric N deposition, synthetic N

Figure 4 .
Figure 4. Model-estimated factorial contributions to changes in crop production (a), N load (b), and riverine N footprint (c) from the 1970s to the 2010s.Major drivers considered here include crop technology improvement (Crop Improvement), synthetic N fertilizer use (Fertilizer), land use and land cover change (LUCC), and manure N application (Manure N).The black dot curve is the net change driven by all these four agricultural activities.

Figure 5 .
Figure 5. Modeled factorial contributions to changes in riverine N footprint of crop production in the six sub-basins from the 1970s to the 2010s.The central map demonstrates the significant trends of N footprint, calculated as the ratio of the change in N yield (i.e.N leaching fom both surface and sub-surface runoff) to the change of crop yield, across the MARB.(Note: due to the close-to-zero values in crop production changes in the Lower Mississippi River Basin (LM) in the 2000s, 0.9 MMT yr −1 ), the N footprint would be extraordinarily large, which is less informative in comparison with the values in other sub-basins.Therefore, the 2000s ′ N footprint in the LM was not plotted in the figure.We also excluded the N footprint in the grids where the absolute value of annual crop yield trend is less than 0.01 kg m −2 ).
) Representation of flood control and dam operations: the DLEM employed in our study does not adequately capture changes in flood control and dam operations, potentially resulting in an overestimation of N loading in sub-basins characterized by dense reservoir networks during flooding years (Lu et al 2020).It is important to highlight that our analysis concentrates exclusively on anthropogenic influences.Notably, we have excluded the effects of climate on interannual fluctuations in N loads by computing the disparity between two simulation experiments.Research has demonstrated that climate variability and extreme events substantially influence historical riverine N loading originating from the MARB (Battaglin et al 2010, Smits et al 2019, Lu et al 2020).

Figure 6 .
Figure 6.Decadal average production change (a) and revenue change (b) of major crops resulting from the four agricultural activities in the sub-basins of the MARB, as estimated by DLEM.