Landscape Pollution Source Dynamics Highlight Priority Locations for Basin‐Scale Interventions to Protect Water Quality Under Hydroclimatic Variability

Extreme weather is associated with a variety of water quality issues that can pose harm to humans and aquatic ecosystems. Under dry conditions, contaminants become concentrated in streams with a greater potential for harmful algal blooms, while wet conditions can cause flooding and broadcast pollution. Developing interventions to improve water quality in a changing climate requires a better understanding of how hydroclimatic variability affects watershed processes, and which places are most vulnerable. We developed a Soil and Water Assessment Tool model of the Cape Fear River Basin (CFRB) in North Carolina, USA, representing contemporary land use, point and non‐point sources, and weather conditions from 1979 to 2019. The CFRB is a large, complex river basin undergoing urbanization and agricultural intensification, with a history of droughts and floods. To identify intervention priorities, we developed a Water Quality Risk Index (WQRI) using the average and variability of contaminant loads across dry, normal and wet conditions. The landscape generally contributed the majority of pollutants (e.g., via erosion, fertilizer and manure applications), including 90.1% of sediment, 83.2% of total nitrogen, and 52.4% of total phosphorus at the City of Wilmington's drinking water intake, yet point sources were influential during dry periods. Approximately 16% of the watershed contributed most of the pollutants across conditions—these represent priority locations for interventions such as restoration, urban, or agricultural best management practices. The WQRI approach considering risks to water quality across conditions can help identify locations where interventions are more likely to improve water quality under climate change.

depend on these resources for some part of their life cycle (Abramovitz & Peterson, 1996;Balian et al., 2010;Dudgeon et al., 2006).Billions of people rely on freshwater for basic needs, and also for fisheries, agriculture, energy production, industry and other uses (Gardner & Finlayson, 2018;Lynch et al., 2016;Pascual et al., 2017).Unfortunately, wetlands are being lost at three times the rate of forests (Gardner & Finlayson, 2018) and freshwater biota are declining more rapidly than other taxa (Reid et al., 2018).The number of stressors on freshwater environments has increased and some threats have intensified, including direct loss and hydrologic alteration, in addition to invasive species, infectious diseases, salinization, emerging contaminants, and climate change (Reid et al., 2018).Climate change has already altered 23 of 31 ecological processes that support key freshwater functions with impacts at genetic, community, and ecosystem levels (Scheffers et al., 2016).
Extreme events are associated with a variety of risks for water supply and water quality.Increasing extreme precipitation is intensifying erosion, and loads of nitrogen and phosphorus (Sinha et al., 2017;Z. Tan et al., 2021).Flooding from hurricanes and extreme rainfall heightens the risks of pollutant transport from vulnerable infrastructure and non-point sources over large areas, with consequences for both surface water and shallow groundwater, as well as estuarine water quality (Du et al., 2020;Paerl et al., 2018;Schaffer-Smith, 2020).Under extremely dry conditions (i.e., seasonal low flow periods or extended droughts), contaminants can become more concentrated in streams, increasing the potential for harmful algal blooms (Mosley, 2015).These distinct water quality issues can both cause harm to aquatic systems, including low dissolved oxygen levels, fish kills, and more (Ascott et al., 2016;Blaszczak et al., 2018;Golladay & Battle, 2002;Lake, 2003;Mallin et al., 2006;Mosley, 2015).Some watersheds have persistent water quality issues under normal conditions-while these long-term "press" disturbances may not always represent acute problems, their effect on environmental degradation and public health cannot be discounted (Frei et al., 2021;Lake, 2003).
Extreme events are becoming more frequent and severe under climate change (IPCC, 2022).Among recent natural disasters, 74% have been related to water, with at least 1 billion people impacted by droughts and floods from 2001-2018 (UNESCO & UN-Water, 2020).Droughts have become more frequent and intense, impacting larger areas for longer durations due to human activities (Chiang et al., 2021).Tropical cyclone driven precipitation events over the U.S. East Coast have increased by 2-4 mm/decade over the last three centuries, with most of the increase taking place over just the past 60 years according to a recent tree ring study (Maxwell et al., 2021).Climate change is also expected to increase the prevalence of harmful algal blooms (Chapra et al., 2017;Paerl & Paul, 2012).These climate-induced impacts to freshwater systems will disproportionately impact the lives and livelihoods of vulnerable communities, particularly in coastal zones (IPCC, 2022).Anthropogenic land use, land management, and appropriation of water resources can further exacerbate the impacts of extreme events on people and ecosystems.Ongoing urbanization and agricultural expansion, as well as intensification of these land uses, have had profound impacts on water and nutrient cycling (Shi et al., 2017;Tong & Chen, 2002).Conversion of floodplains and wetlands to other land uses reduces the capacity of the landscape to buffer extreme conditions (Johnson et al., 2020;Narayan et al., 2017).Dams and water extraction activities are associated with increased hydrologic drought (Wada et al., 2013).Urbanization and population growth increase water use and loadings of contaminants to streams (Foley, 2005;McDonald et al., 2011;Paul & Meyer, 2001).In spite of urban growth, agriculture is often the dominant water consumer, accounting for as much as 92% of the human water footprint (Foley, 2005;Hoekstra & Mekonnen, 2012;Power, 2010).Nutrients, sediment, bacteria, heavy metals and other contaminants in agricultural runoff can substantially reduce water quality (Foley, 2005;Gordon et al., 2010;Koneswaran & Nierenberg, 2008;Power, 2010).These compounding modifications to the water cycle may impose greater stress on water resources (Haddeland et al., 2014).
Hardened infrastructure and technological solutions may be attractive solutions for mitigating the effect of hydroclimate variability on water resources, given well-documented costs and effectiveness, yet these alone may not reduce vulnerability (Haghighatafshar et al., 2020;Walker et al., 2022).For example, built infrastructure for flood protection can cause a "levee effect" that paradoxically increases flood risk (Di Baldassarre et al., 2009).Development of seemingly "safe" areas can produce a bigger catastrophe when a storm exceeds the infrastructure's defense capabilities (Di Baldassarre et al., 2009).Most current water distribution and treatment infrastructure, sewage, and stormwater management systems in the U.S. were designed using event intensity, duration, frequency information that did not consider climate and land use change (Wright et al., 2019).For rural areas, hardened infrastructure solutions may be less desirable given the high costs of engineering and design, permitting, implementation over large land areas, and long-term maintenance (Alves et al., 2018;Browder et al., 2019;Hovis et al., 2021;Suttles et al., 2021).
Nature-based solutions, such as conservation of wetlands and forests, restoration of degraded natural habitats, agricultural field measures or "best management practices" (BMPs), and managed retreat can all play a role in improving the resilience of watersheds (Antolini et al., 2020;Johnson et al., 2020;Keesstra et al., 2018;Suttles et al., 2021).These solutions may be less costly and faster to implement than hardened infrastructure solutions, and can also provide additional co-benefits like improved access to greenspace and recreation, economic opportunities, as well as benefits for fish and wildlife habitat and biodiversity (Bassi et al., 2021;Chausson et al., 2020;DeLong et al., 2021;Keesstra et al., 2018).Among nature-based solutions, floodplain restoration is expected to have the greatest benefits for both water quality and flood-risk reduction (Suttles et al., 2021).
Formulating appropriate interventions that will deliver durable benefits (e.g., infrastructure adaptations or nature-based solutions) requires understanding how both dry and wet conditions can affect water quality.Watershed models, such as the Soil and Water Assessment Tool (SWAT), can provide insight into how interactions between, topography, soils, land use and weather interact and predict in-stream flow and water quality across watersheds (J.G. Arnold et al., 2012;Gassman et al., 2014).SWAT is one of the most widely used watershed models, and it has been previously applied to examine future changes in watershed processes by incorporating climate projections to evaluate resulting impacts on water quantity (Tan et al., 2021;Xu et al., 2019), with fewer studies examining water quality (e.g., Ouyang et al., 2018).A number of studies have explored contemporary extreme events with SWAT, including a sub-daily model of flash flooding for ungaged watersheds in Spain (Jodar-Abellan et al., 2019), examinations of streamflow response to climate variability and land use (Li & DeLiberty, 2020;Zhang et al., 2017), exploration of how more extreme rainfall has affected erosion and nutrient runoff into the Gulf of Mexico (Z.Tan et al., 2021), and assessment of impacts from frequent hurricane activity on water quality (Ouyang et al., 2022).While it is a well-established tool to guide placement of BMPs (e.g., Abimbola et al., 2020;Admas et al., 2022;Chiang et al., 2021), SWAT has not been used previously to identify priority locations for interventions to improve watershed resilience with explicit consideration of both dry and wet hydroclimatic conditions.
Given that many watersheds are already experiencing more frequent extreme weather patterns, retrospective analysis of hydroclimatic variability can help to highlight places where additional attention and mitigation strategies may be warranted.The Cape Fear River Basin (CFRB) in North Carolina (NC), USA, represents an ideal study location given its dynamic hydrology, with a history of both droughts and floods, including multiple 500-year storm events since 2016.A variety of interventions have been proposed to help moderate water quantity and quality in the watershed, including both human-managed infrastructure and nature-based solutions.To understand how hydroclimatic variability influences flow and water quality across the basin, we developed a SWAT water quantity and quality model for the CFRB, representing contemporary land use and management under weather conditions spanning 1979-2019.To identify strategic locations where landscape-based interventions could improve water quality and enhance the resilience of freshwater systems, we created a Water Quality Risk Index (WQRI) quantifying hotspot dynamics across conditions.

Study Area
The CFRB is the largest river basin fully contained within NC, at >2.35 million ha (or more than 9,100 mi 2 ) (Figure 1).The CFRB is divided into two major physiographic regions.The upper basin is in the clay-rich Piedmont plateau east of the Southern Appalachian Mountains, with rolling topography from 450-100 m elevation.Below the confluence of the Deep and the Haw Rivers, the Piedmont drops into the lower basin on the Atlantic Coastal Plain, with sandy soils that slope gently to meet the Cape Fear River Estuary and the Atlantic Ocean.The CFRB is characterized by a humid subtropical climate, with average temperatures ranging from −1°C during the winter to 31.7°C in the summer.Snow is rare, with most precipitation falling as rain in the Piedmont (1,120-1,220 mm/year) and Coastal Plain (1,120-1,420 mm/year).
The basin features outstanding aquatic biodiversity (NatureServe, 2022;NC Wildlife Resources Commission, 2015), and is also the most populous basin in NC, home to growing cities such as Greensboro, Durham, Chapel Hill, Fayetteville, and Wilmington.Millions of people depend directly on the river for drinking water; however, 26% of NC residents rely on privately owned shallow groundwater wells, which are vulnerable to contamination (MacDonald Gibson & Pieper, 2017;Naman & Gibson, 2015).The U.S. Army Corps manages B. Everett Jordan Dam (hereafter Jordan Lake) for flood control and drinking water provision, as well as operating three lock and dam structures on the lower Cape Fear River.Lock and Dam #1 is the location of the City of Wilmington's drinking water intake, and below this point the river is subject to tidal influence.There are additional reservoirs in the basin, including Randleman Lake, which provides drinking water, and Harris Lake, associated with operations at Shearon Harris Nuclear Plant.NC has a statewide chlorophyll-a standard specifying a limit of no more than 40 μg/L of for all surface waters (Fresh Surface Water Quality Standards for Class C Waters, 1976).
Water quality and quantity are highly variable in the CFRB.A severe drought in 2007 resulted in widespread water supply concerns across the state-it was reported that 79% of water customers faced restrictions and ∼600 wildfires occurred in August alone (Davis, 2015).In contrast, elevated hurricane activity has punctuated recent years.NC experienced multiple damaging storms since 2016, including Hurricane Matthew (2016), Hurricane Florence (2018), Tropical Storm Michael (2018), Hurricane Dorian (2019), Hurricane Isaias (2020), Tropical Storm Fred (2021), and Hurricane Ian (2022).More frequent extreme rainfall events have impacted the state over the same period, on top of large hurricanes.In addition to fluctuations in water supply, the CFRB has a long history of water quality issues, due in part to excess nutrient pollution from both point and non-point sources (DeMeester et al., 2019;NC Department of Environment & Natural Resources, 2005), including the highest density of concentrated animal feeding operations (CAFOs) in the U.S (Brown et al., 2020).The CFRB is a nutrient hotspot within the U.S. Southeast region (Hoos & Roland, 2019), and considering rivers nationwide, with loads of nitrogen and phosphorus rivaling those observed in the Corn Belt in the Midwest (Robertson & Saad, 2019;Shen et al., 2020).

Model Setup
To better understand the dynamics of hydrology and water quality of the CFRB, we developed a SWAT model representing contemporary land use, soil and slope, and historical weather conditions from 1979 to 2019 (SWAT version 2012, revision 681).SWAT is a semi-distributed hydrologic model that simulates a variety of watershed processes including the water balance, plant growth, and sediment and nutrient transport across the landscape and in-stream (Arnold et al., 2012).SWAT has been widely used in hydrologic studies and is well-suited to studies of agricultural landscapes (Gassman et al., 2014).We modified a SWAT model originally developed by the U.S. Geological Survey (USGS) South Atlantic Water Science Center as part of a Coastal Carolinas Focus Area Study of water availability and water use under population growth, land use change, and climate change, hereafter referred to as the CCFAS SWAT model (Gurley, García, & Pfeifle, 2023;Gurley, García, Pfeifle, et al., 2023).USGS delineated 2,928 subbasins comprised by 13,596 hydrologic response units (HRUs) and calibrated the model to represent unimpaired flow from 2000 to 2014.
Building on the USGS CCFAS SWAT model, we developed a new water quantity and quality model incorporating additional elements to capture water storage capacity and water quality in the basin.We updated the climate record using 1-km gridded weather data 1979-2019, spanning multiple drought periods and large storm events (Thornton et al., 2017).We included reservoir operations at B. Everett Jordan Lake, which is managed for flood control, and captured additional waterbodies and wetlands which store water and process nutrients based on the National Wetland Inventory (U.S. Geological Survey, National Geospatial Program, 2018).Contributions of flow, sediment, nitrogen and phosphorus from wastewater treatment plants, and other permitted emitters, were incorporated in the model using measured data 1994-2019 (NC Department of Environmental Quality, Division of Water Resources, 2019), and monthly averages for the period preceding recordkeeping.We also incorporated annual average atmospheric nitrogen deposition (National Atmospheric Deposition Program [NRSP-3], 2020).Nutrient and sediment loads from non-point sources were represented principally through land management practices, including cropping patterns and rotations, tillage, fertilizer and manure applications on crops, pastures, pine plantations, and lawns.We used a mass balance approach to parameterize fertilizer and manure applications considering fertilizer sales data (John & Gronberg, 2017), manure generated by grazing livestock (USDA-NASS, 2018), and by animals in concentrated animal feeding operations (Environmental Working Group & Waterkeeper Alliance, 2016;North Carolina Department of Environmental Quality, 2019;North Carolina State University, 2019).
To the best of our ability, we sought to minimize uncertainty with respect to input data and parameterization (Fu et al., 2020;Moges et al., 2020).Fundamental model inputs included spatial gridded representations of topography, soil, land use and climate variation across the watershed.Wherever possible, we used empirical data to inform parameterization, but in some cases we relied on literature values to inform parameter values.Given differences in the physiography and land use in the Piedmont and Coastal Plain, we parameterized these regions separately.More detail regarding model development is provided in Supporting Information S1.

Model Calibration and Validation
We calibrated and validated the model at Lock and Dam #1 (Figure 1) for the period 2000-2019 using a MATLAB routine integrated with SWAT; the Cape Fear River is subject to tidal influence below this point.Daily observed streamflow and water quality monitoring records collected from 2010 to 2019 were used for calibration, while we retained observations from 2000 to 2009 for validation.The calibration and validation periods represented a range of hydrologic flow conditions, as well as high and low loads of sediment and nutrients.
Observed flow and water quality data were retrieved via the Water Quality Portal for three monitoring stations co-located near Lock and Dam #1 (National Water Quality Monitoring Council, 2021;Read et al., 2017).Daily streamflow data spanning 2000-2019 were available at USGS gage #02105769.Loads of water quality parameters were calculated using streamflow and in-stream concentrations measured through March 2020.Sediment data was available from NC Division of Water Resources' monitoring station #B8349000, while total nitrogen (TN) and total phosphorus (TP) were collected from the NC Department of Water Quality's monitoring station #B8350000.Observations of TN in most cases were aggregated from individual measurements of total Kjeldahl nitrogen and inorganic nitrogen (nitrite and nitrate) recorded on the same day.For days with missing observations, we estimated daily constituent loads using the LOADEST model (regression model #0, Runkel et al., 2004); there were 256 true measurements of daily sediment (3.32%), 388 true measurements of daily TN (9.38%), and 308 true measurements of daily TP (5.13%) available.We used all available data to generate sediment and nutrient load estimates and retained the load estimates 2000-2019 for calibration and validation of the model.Beginning with flow, followed by sediment, TP, and TN, calibration was performed iteratively, changing one model parameter at a time.Sensitive parameters were altered to first achieve satisfactory hydrologic calibration, and then water quality calibration according to best practices for model evaluation (Arnold et al., 2012;Moriasi et al., 2007;Scavia et al., 2017).We relied on metric-based approaches for calibration and validation against streamflow and load estimates, including using the coefficient of determination (R 2 ), ratio of the root mean square error to the standard deviation of measured data (RSR), Nash-Sutcliffe efficiency (NSE) and percent bias (PBIAS; Arnold et al., 2012;Moriasi et al., 2007).We also employed graphical approaches to ensure that SWAT predictions generally captured the trends of true observations.Additional details are included in Supporting Information S1.

Simulations
To assess hydrology and water quality dynamics across many conditions, we ran a daily simulation with weather conditions from 1979 to 2019, with the first three years serving as a warmup period.To evaluate the relative importance of point versus non-point sources of water quality contaminants, we also ran the model without point sources for 2010-2019.

Assessing the Importance of Point and Non-Point Sources
We examined the relative importance of point and non-point sources in terms of the average and standard deviation (SD) of the load from each source by month for 2010-2019.We also separately examined a dry year (2011) and a wet year (2016).These two extreme years were characterized by consistent departures from normal flows in the Cape Fear River relative to the entire period of record (National Water Quality Monitoring Council, 2021;Read et al., 2017).

Tracking Landscape Source Hotspots Across Conditions
To better understand landscape source dynamics, we examined the spatial distribution of landscape-derived sediment and nutrient hotspots under dry, normal, and wet conditions, respectively.Watershed-scale, nature-based solutions implemented on the landscape are expected to help improve water quality under both dry and wet conditions, and also have benefits for moderating water quantity; therefore, we focused the bulk of our analysis on landscape-derived sediment and nutrient source hotspots across conditions.Landscape sources include non-point source pollution, as well as applications of manure from permitted CAFOs, and atmospheric deposition, but do not include point-source dischargers like wastewater treatment plants and industrial emitters.
We defined hydroclimatic conditions for each subbasin, separately, based on runoff generated over the full simulation period, considering "dry" conditions as the lower 25% of runoff volumes, "normal" conditions as the middle 50%, and "wet" conditions as the upper 25% of runoff (Helsel et al., 2020).For each subbasin, we then calculated the mean and SD of the load for each parameter under each climate condition.To facilitate comparisons across parameters and conditions, we standardized each measure, generating a z-score (Equation 1) with a mean at zero and SD equal to 1, capped at 3.5 SD to avoid undue influence from outliers.Z-scores are widely used to compare measurements with different scales to one another (Dixon, 1960), and can be used to create composite scores incorporating multiple factors (Song et al., 2013).where: avg(sed)dry = z-score of the average load under dry conditions.  = observed value of the average load under dry conditions.  = population mean of the average load under dry conditions.  = population SD of the average load under dry conditions.
By Equation 1, the z-score of the SD sediment load under dry conditions would be calculated as: where: SD(sed)dry = z-score of the SD load under dry conditions.  = observed value of the SD load under dry conditions.  = population mean of the SD load under dry conditions.σ = population standard deviation of the SD load under dry conditions.

Identifying Intervention Priorities With a Water Quality Risk Index
Ideally, intervention strategies such as nature-based solutions would be implemented at locations where they improve water quality under a range of conditions, representing no regrets investments.Conservation of remaining high-quality forests, floodplains, and wetlands is important for avoiding further loss of natural capacity to purify water and buffer communities downstream from droughts and floods.Restoration, through changing land cover and land use, stream rehabilitation, and floodplain reconnection, can also add or enhance natural capacity.
To identify priority locations for interventions, we developed a Water Quality Risk Index (WQRI) considering the relative amount, or "intensity" and variability of sediment, TN, and TP loads under dry, normal, and wet conditions for all subbasins (Figure 2).We considered the intensity (derived from the average load) and the variability (derived from the load SD) to be distinct aspects useful for characterizing the relative level of disturbance from contaminants across the watershed.First, for each subbasin and each parameter we generated an intensity score by summing the z-scores of the average load across conditions (Equation 2).We generated a variability score for each subbasin and each parameter by summing the z-scores of the load SD across conditions (Equation 3).
Next, we generated a composite intensity score for each subbasin by summing the z-scores of intensity across parameters (Equation 4), and a composite variability score in the same fashion based on z-scores of variability (Equation 5).Finally, for each subbasin we calculated an overall WQRI as the simple average of the z-score of composite intensity and the z-score of composite variability (Equation 6).At each step where a z-score was calculated, the value was capped at a maximum of 3.5 SD in order to limit undue influence from outliers.
where: I = intensity score.The approach we employed to generate the WQRI is similar to other assessments aimed at highlighting outliers and spatial priorities considering multiple factors.For example, The Nature Conservancy identified locations expected to be resilient to climate change that will support high biodiversity into the future based on a variety of biophysical and condition metrics using a z-score based approach (Anderson et al., 2014;Benner et al., 2014).The Center for Disease Control's social vulnerability index (SVI) is another example aimed at measuring communities' ability to respond and recover after a natural disaster (Flanagan et al., 2011(Flanagan et al., , 2018)).The SVI uses percentile ranking to put 15 socioeconomic metrics on the same scale, and gives equal weighting to each when aggregating them into four themes, finally integrating the theme scores into an overall composite index (Flanagan et al., 2011(Flanagan et al., , 2018)).

Model Calibration and Validation Results
The final calibrated model demonstrated very good daily performance at Lock and Dam #1 for hydrology, and good to very good monthly performance for water quality parameters over the calibration period (Table 1; Moriasi et al., 2007).Weaker performance during the validation period is not surprising given that we set up the model with contemporary land use and management, and many changes have occurred in the watershed over 20 years.Within the U.S., the southeast has experienced the most rapid recent land use change, particularly forest loss to suburban sprawl (Gaines et al., 2022;Homer et al., 2020;Sanchez et al., 2020;Sleeter et al., 2018).NC, and particularly the CFRB, has some of the highest urban and suburban growth rates in the country (U.S. Census Bureau, 2020) and is undergoing agricultural intensification, notably via expansion of swine CAFOs from the 1980s through the early 1990s and ongoing growth of poultry CAFOs (Environmental Working Group & Waterkeeper Alliance, 2016; Miralha et al., 2021;Montefiore et al., 2022).
We reported calibration statistics for the period January 2010 through December 2018 (Table 1, Figures S24-S27 in Supporting Information S1).After Hurricane Florence in September 2018, wet weather persisted through the spring of 2019 with extended high flow from Lillington down to the locks and dams.The locks and dams on the lower Cape Fear River may back water up behind them for extended periods of time-Lock and Dam #3 in particular is considered to be a dampening structure that causes backwater effects that may not be captured by SWAT (DeMeester et al., 2019).It is also possible that operations at Lake Harris, and the associated Shearon Harris nuclear facility, affected flows.Additional calibration and validation details, including calibrated parameters and plots used in graphical model evaluation, are provided in Supporting Information S1 (Text S11).
SCHAFFER-SMITH ET AL.

Relative Importance of Point Source Discharge and Landscape Sources
Analysis of the sources of in-stream flow and contaminant loads at Lock and Dam #1 revealed that the landscape represented the major source of flow and contaminant contributions from 2010 to 2019 (Table 2).Across all weather conditions, effluent from permitted wastewater treatment plants and industrial dischargers accounted for an average of 9.66% of the cumulative monthly flow at Lock and Dam #1, while the landscape accounted for an average of 90.34% over 10 years.Over the long-term we did not observe notable seasonal variation in the contributions of landscape sources and permitted discharge into rivers, yet their relative importance changed during very dry or wet conditions (Table 2).During a dry year in 2011, point sources contributed as much as 54.57% of the cumulative flow, 80.05% of the sediment, 84.50% of TN, and 75.55% of TP in a single month.During a wet year in 2016, landscape sources contributed as much as 99.30% of the cumulative flow, 98.89% of sediment, 97.69% of TN, and 81.21% of TP in a single month, representing a more prominent source of contaminants in spring and summer months.More detail about spatial and temporal patterns of in-stream loads is provided in Supporting Information S1.

Landscape Water Quality Hotspot Dynamics
Landscape hotspots differed spatially by pollutant when examining long-term average loads generated under weather conditions from 1982 to 2019 (Figure 3).Sediment was most often generated in urban areas, particularly in the Piedmont (upper basin), while nutrients were most often sourced from agricultural land, particularly in the Coastal Plain (mid-lower basin).Phosphorus loads were generally high both in cultivated crop areas and urban areas (Figure 3).
Examination of relative landscape contributions under dry, normal, and wet conditions (Figures 4 and 5) revealed distinct patterns across pollutants compared to long-term average loads (Figure 3).For example, important  Note.Loads for water quality parameters were estimated using LOADEST."Very good" performance is indicated by R 2 and NSE values from 0.75 to 1.00, RSR values from 0.00 to 0.50, with PBIAS < ± 10 for streamflow, PBIAS< ±15 for sediment and PBIAS < ±25 for nutrients (Moriasi et al., 2007)."Good" performance is indicated by R 2 and NSE from 0.65 to 0.75, RSR from 0.5 to 0.6, with PBIAS from ±10 to ±15 for streamflow, PBIAS from ±15 to ±30 for sediment, and PBIAS from ±25 to ±40 for nutrients."Unsatisfactory" performance is indicated by R 2 and NSE ≤ 0.5, RSR > 0.7, and PBIAS > ±25 for streamflow, PBIAS > ±55 for sediment, and PBIAS > ±70 for nutrients (Moriasi et al., 2007).Coefficient of determination (R 2 ), Nash-Sutcliffe efficiency (NSE), ratio of the root mean square error to the standard deviation of measured data (RSR), percent bias (PBIAS).sediment source areas in terms of the relative average load were quite widespread under normal conditions, and more spatially concentrated around urban centers, and in the Northeast Cape Fear under dry and wet conditions (Figure 4).The patterns of importance in terms of relative sediment load variability were similar (Figure 5).While the Piedmont landscape generated relatively low nutrient loads overall in terms of absolute measurement (Figure 3), relative contributions of nitrogen from the Piedmont were more important under dry conditions (Figure 4), though less variable than the contributions from the Coastal Plain (Figure 5).Under normal conditions, the subbasins contributing relatively large amounts of phosphorus were broadly distributed throughout the basin, while spatially clustered hotspots emerged within urban areas, the lower Cape Fear River mainstem, and the Northeast Cape Fear under dry and wet conditions (Figure 4).Subbasins with high intensity based on average load typically also demonstrated greater variability based on load SD (Figures 4 and 5).
WQRI scores across the basin identified locations that merit attention based on their relatively high intensity and variability of sediment, nitrogen, and phosphorus contributions across conditions (Figure 6).Subbasins with a low WQRI likely represent high priorities for land protection to maintain functioning floodplains, water purification, and habitat that supports biodiversity, as well as high quality community water supplies (e.g., Figure 6a).Conversely, subbasins with a high WQRI represent high priorities for interventions, such as restoration, BMPs, or urban green and gray infrastructure strategies to improve water quality, depending on local land use and management conditions (e.g., Figure 6b).Many of these strategies could also yield benefits for flood-risk reduction and water provisioning during droughts (Chausson et al., 2020;DeLong et al., 2021;Griscom et al., 2017;Kousky et al., 2013).We found that the highest risk regions (WQRI >1) comprised 16.4% of the entire watershed.Sediment loads are reported in metric tons per hectare (1 metric ton = 1,000 kg).

Utility of the Water Quality Risk Index Relying on Watershed Modeling
We developed the first SWAT water quantity and quality model for the entirety of the CFRB.We examined risks to water quality from landscape sources, taking into account the intensity and variability of pollutant loads for multiple contaminants across dry, normal, and wet conditions 1982-2019, presenting a new application of SWAT model results.The WQRI revealed water quality risks that were not captured by long-term average estimated loads predicted by SWAT-notably in swaths of the upper and middle basin outside of urban centers (Figures 3  and 6).The overall WQRI and the underlying load intensity and variability scores for specific contaminants under distinct hydroclimatic conditions can aid in understanding the drivers of water quality issues, avoiding degradation of resilient subbasins, and selecting appropriate interventions to reduce water quality issues.It should be noted that regions below Lock and Dam #1 are subject to processes not captured by SWAT, such as tidal cycles, storm surge, and sea level rise, which could alter local contaminant loads and risks to water quality in those areas (Upadhyay et al., 2022).Results should also be interpreted considering spatial variation in model performance; smaller catchments at the edges of the watershed are more likely to be sensitive to input data uncertainty and parameter uncertainty.Our finding that the vast majority of contaminants in the CFRB were generated by the landscape is consistent with previous SWAT-based assessments in the basin.A previous study of the lower CFRB found that while the upper basin contributed 50% of the total nutrient load at Lock and Dam #1, land applications of fertilizers and manures below Jordan Lake and the Deep River accounted for 70% of locally generated nutrients and 35% of the total load, while just 15% of the total load was derived from point sources (RESPEC, 2015).Similarly, a previous analysis found that 70% of the TP load in the Northeast Cape Fear River was due to erosion (Narayan et al., 2017).
A prior investigation of the Jordan Lake Watershed found that overall nutrient loads decreased from 1997 to 2010 due to reductions in loads from point sources and rural land uses, yet urban landscape loads increased over the same period (Tetra Tech, 2014).
The spatial patterns of important landscape source areas we identified in CFRB also agree with other existing data.For example, a USGS Spatially Referenced Regression on Watershed Attributes (SPARROW) model identified higher sediment loads in the Piedmont, particularly urban areas and disturbed land, with higher nutrient loads in the lower basin (Gurley, García, Hopkins, et al., 2019;Gurley, García, Terziotti, et al., 2019).The high risk hotspots that we identified with the WQRI overlap spatially with known surface water impairments, including surface waters near urban centers throughout the basin, the Jordan Lake Watershed, and tributaries to the Northeast Cape Fear River (NC Department of Environmental Quality, Division of Water Resources, 2020).High risk hotspots also track with regions where groundwater nitrate likely exceeds the standard of 10 mg/L based on well monitoring data and modeling (Messier et al., 2014).
The CFRB SWAT model and our baseline model results provide vital information for ungaged, and poorly monitored areas of the CFRB, with important insights for public health and ecosystem health.Given strong alignment between nitrate exceedances and high-risk landscape hotspots we identified, our model can provide information for communities that lack groundwater monitoring data.Groundwater nitrate levels as low as 2.5 mg/L may cause significant health impacts (De Roos et al., 2003;Ward et al., 1996Ward et al., , 2005;;Weyer et al., 2001).Our results also can provide new information regarding many reaches which currently have "insufficient information to make a determination" about impairment status (NC Department of Environmental Quality, Division of Water Resources, 2020).Model results could be used to target future surface water monitoring efforts by state and federal agencies, as well as volunteer groups.Notably, stream gages and other surface water monitoring data  Program, 2016;National Water Quality Monitoring Council, 2021), which are more likely to be impacted by extreme events including flooding (Schaffer-Smith et al., 2020).

Transferability
Our approach using watershed modeling and the WQRI can be applied in other watersheds to identify regions that present water quality issues across conditions, which may merit further exploration and interventions.The use of standardized z-scores to compare among distinct water quality risks and calculate an overall WQRI is transferrable to any watershed's local context and weather conditions.We used simple cutoffs for the lower and upper percentiles of runoff to separate dry and wet from normal conditions, but these criteria could be customized using locally relevant thresholds for water management or ecological concerns.We weighted all contaminants and all climate conditions equally, but the WQRI could easily be adjusted to incorporate weights if specific conditions, or specific contaminants, are of greater concern in a given region.For example, The Nature Conservancy's resilient and connected network assessment assigned higher weights to some variables when creating composite scores (Anderson et al., 2014).To date a small number of studies have examined water quality under extremes with SWAT, but given the proliferation of watershed modeling, our analysis can be replicated for other basins with existing models.

Solutions to Address Water Quality Issues and Improve Resilience to Hydroclimatic Variability
Following on recent years of volatile weather conditions, including multiple 500-year storm events since 2016, NC is exploring a variety of options to improve resilience across the entire state.Large investments planned for modeling studies and increases in funding for conservation and restoration programs aimed at reducing flood-risk represent a golden opportunity to select interventions that also improve the health and resilience of watersheds more holistically.Nature-based solutions (e.g., wetland and forest restoration, field measures that improve soil quality) as demonstrated by Keesstra et al. (2018) could provide substantial benefits including buffering communities from flooding (Acreman & Holden, 2013;Antolini et al., 2020;Sutton-Grier et al., 2015), augmenting water supply during droughts (Acreman & Holden, 2013), carbon sequestration, providing plant and wildlife habitat (Fargione et al., 2018;Griscom et al., 2017), recreation opportunities (Chausson et al., 2020), and more.
The results from this study can inform policies and programs to implement nature-based solutions in the CFRB.Protections on riparian buffers are a widely used strategy to protect surface water quality (Cole et al., 2020;Lovell & Sullivan, 2006).Aside from areas around Jordan Lake and Randleman Lake, the CFRB currently lacks regulations to protect riparian buffers around surface water features (North Carolina Conservation Network, 2016;Surface Water and Wetland Standards, 2022).Buffer protections could be an important strategy to avoid compromising remaining floodplains, particularly given high rates of population growth and land use change (Homer et al., 2020;Sanchez et al., 2020;U.S. Census Bureau, 2020).The WQRI that we developed could be included in criteria for allocating funding toward conservation, restoration, and voluntary strategies through a variety of state programs (e.g., the NC Land and Water Fund) and federal programs (e.g., the U.S. Department of Agriculture Conservation Reserve Program).Water quality issues in urban areas may be more successfully addressed with watershed-scale interventions rather than projects targeting individual stream segments or neighborhoods (Walsh et al., 2005).Our approach can support watershed planning and financing schemes for larger projects with cost-sharing and benefits for multiple jurisdictions.There is precedent in the neighboring Neuse River Basin for creative strategies, including a nutrient credit and trade system that financed headwater protection with additional flood storage benefits (Phthisic et al., 2018;Walls & Kuwayama, 2019), partnerships between local governments and conservation groups (Upper Neuse River Basin Association, 2021), and upper basin protection financed through a fee levied in the City of Raleigh (Patterson et al., 2012).
Additional landscape-based strategies can also be considered to improve water quality.Land applications of manure are subject to nutrient management plans, yet evidence suggests that these are not always followed in practice due to a variety of constraints (Cabot & Nowak, 2005;Osmond et al., 2015;Tao et al., 2014), and application above plant nutrient requirements can occur even while following nutrient management plan protocols (Long et al., 2018).Typically, agronomic rate limits are based on nitrogen, but some states have implemented nutrient limits based on phosphorus (Bradford et al., 2008;Sharpley et al., 2012).Phosphorus-based limits could SCHAFFER-SMITH ET AL.
be appropriate intervention, given high existing legacy phosphorus concentrations (Wegmann et al., 2013); over 50% of statewide soil samples have "very high" phosphorus (Mehlich-3 soil test extractant) and additional applications would not increase yields for 84% of fields tested (Gatiboni et al., 2020).Incentive programs can complement regulations to help reduce losses of sediment and nutrients.Reverse auctions are an innovative approach that can more rapidly scale payment for services programs (Valcu-Lisman et al., 2017).
Our focus in this study was on landscape sources of contaminants, yet point sources are also an important source of phosphorus, and under very dry conditions they can be the dominant contaminant source at Lock and Dam #1.

Future Work
Advances in watershed model development, calibration and validation methods are ongoing, offering refinements that could improve the use of SWAT for studying watershed resilience to climate change.A recent assessment found that underlying equations used by most hydrological models are pushed to their limits for contemporary extreme precipitation conditions (La Follette et al., 2021).Wellen et al. (2014) determined that implementing state-specific parameters improved predictions under extreme high flows for two watersheds near Lake Ontario.Dong et al. (2019) employed a season-specific multi-site calibration to model the Hamilton Harbor Watershed in southern Ontario, Canada.A recent review found that just 3% of SWAT studies have examined coastal river basins (Upadhyay et al., 2022).It is also possible to couple SWAT with models that better capture coastal dynamics (Upadhyay et al., 2022).This study of the CFRB is part of a growing literature applying SWAT to explore the effect of hydroclimatic variability on water quantity and quality.As interest in this topic grows, so too will guidance for appropriate model development and analysis methods.
Provided additional data becomes available, it would be possible to improve the representation of watershed processes in the CFRB.The inclusion of extractive water use and water transfers would improve modeled hydrology.Currently reporting is only required for water withdrawals exceeding 3,785 m 3 per day (1 million gallons per day) for agricultural uses, or 379 m 3 per day (100,000 gallons per day) for other uses and for water transfers between basins (Water Use Registration and Allocation., 2022).Representation of additional structures that modify hydrology and trap sediment and nutrients could also improve the model (Jalowska & Yuan, 2019); in addition to major reservoirs and the locks and dams, there are numerous small barriers (e.g., defunct mill dams) which are not mapped by existing authoritative spatial datasets.Similarly, additional nutrient sources could be incorporated in the future, including septic tanks and biosolids application areas that are not well documented in public datasets.Finally, there is limited knowledge of the specific location, and timing of land management activities including applications of fertilizer and manure (Rosov et al., 2020;Shea et al., 2022), which vary not only spatially but also year-to-year, given changing constraints and incentives for individual operators, as well as social and psychological factors (O'Connell & Osmond, 2022).
To evaluate the effectiveness of possible strategies to improve water quality, and to determine how much intervention may be needed, additional scenario modeling can be performed.Scenarios simulating interventions will demonstrate how each strategy could alter flow and nutrient loads for each subbasin under a range of weather conditions.We expect this will highlight trade-offs among strategies and help to identify the places where the greatest potential exists to improve water quality, also offering quantitative estimates for moderation of droughts and floods.Furthermore, there is a need to consider the impacts of future changes in climate and land-use.Urbanization is likely to impact water availability in addition to altering contaminant loads in the CFRB (Sanchez et al., 2018).For the neighboring Neuse Basin, climate and land use change may result in a 30% increase in nitrogen loads by 2070 (Gabriel et al., 2018).The implications of these future changes can be evaluated through additional SWAT model scenarios simulating development and climate change.

Conclusion
Taking distinct hydroclimatic conditions into account in watershed modeling can help highlight priority places to improve the resilience of watersheds in terms of both water quantity and quality.Conservation and restoration are SCHAFFER-SMITH ET AL.
key that may help to ensure resilient, high quality water supplies into the future to support both human and natural communities.In the CFRB, the landscape consistently contributes a large amount of contaminants, but ∼16% of subbasins are the most important contributors across dry, normal and wet conditions.These regions merit further attention for actions to improve water quality, and hopefully, other aspects of watershed condition.
Regions with low WQRI scores that currently lack formal protection should be strongly considered for future conservation investment.Our straightforward WQRI approach to identify watershed-scale intervention priorities is directly translatable to any watershed seeking to increase the resilience of community water resources and aquatic ecosystems.The WQRI can easily be adapted based on locally specific concerns, including customized definitions of key weather and event thresholds, and consideration of relevant contaminants of interest.This publication was produced while DS was a NatureNet Science Fellow, funded by The Nature Conservancy and Arizona State University Center for Biodiversity Outcomes.We thank The Nature Conservancy and Arizona State University Center for Biodiversity Outcomes, the Ira A. Fulton Schools of Engineering, the School of Geographical Sciences and Urban Planning, and the School of Sustainable Engineering and the Built Environment for supporting this project.Additional funding was provided by an Environmental Enhancement Grant administered by the North Carolina Attorney General's Office, and generous support from the WILEMAL Fund to The Nature Conservancy's North Carolina Chapter.We acknowledge Ana Garcia and Laura Gurley for their contributions as developers of the CCFAS SWAT water quantity model, which this study built upon.We also thank Barbara Doll, Jack Kurki-Fox, Sheila Saia, Deanna Osmond, Yuhang Wei, Ting Liu, Katie Martin, Tracy Baker, John Spivey, and staff at The Nature Conservancy in North Carolina for their advice and feedback on earlier stages of this work.
= z-score.  = observed value.  = population mean σ = population SD By Equation 1 the z-score of the average sediment load under dry conditions would be calculated as: avg(Sed)dry = (Sed)dry − (Sed)dry (Sed)dry SCHAFFER-SMITH ET AL.

Figure 2 .
Figure2.A water quality risk index (WQRI) was calculated for each subbasin in the Cape Fear River Basin using a series of z-score calculations and aggregations to account for distinct aspects of water quality risk for different parameters under different weather conditions (Equations 1-6).For each parameter, first a z-score (mean = 0, SD = 1, capped at 3.5 SD) was calculated for the load mean and SD for each condition for each parameter.Intensity and variability for each parameter were calculated by summing z-scores across conditions.Composite intensity and variability scores were calculated by summing intensity and variability z-scores, respectively, across parameters.Finally, a WQRI was generated for each subbasin by taking a simple average of the z-score of composite intensity and the z-score of composite variability."Dry" conditions were defined as the lower 25% of runoff, while "normal" constituted the middle 50%, and "wet" conditions were represented by the upper 25% based on weather 1982-2019.Abbreviations: total nitrogen (TN), total phosphorus (TP).

Figure 3 .
Figure 3.Long-term average daily runoff, sediment, total nitrogen (TN) and total phosphorus (TP) loads varied spatially across the Cape Fear River Basin based on contemporary land use and historical weather conditions from 1982 to 2019.Sediment loads are reported in metric tons per hectare (1 metric ton = 1,000 kg).

Figure 4 .
Figure 4.The relative contaminant load amounts across the Cape Fear River Basin varied by parameter under different hydroclimatic conditions 1982-2019, determined by calculating standardized z-scores of the average load for each parameter by condition, capped at a maximum of 3.5 SD (Equation 1, Figure 2).For each parameter, the z-scores for the average load across conditions were combined into an intensity score (Equation 2, Figure 2).Abbreviations: z-score for the average sediment load by condition [z avg (Sed) c ], z-score for the average total nitrogen load by condition [z avg (TN) c ], z-score for the average total phosphorus load by condition [z avg (TP) c ], where condition c = dry, normal, or wet.

Figure 5 .
Figure 5.The relative variability of contaminant loads across the Cape Fear River Basin varied by parameter under different hydroclimatic conditions 1982-2019, determined by calculating standardized z-scores of the load standard deviation for each, capped at a maximum of 3.5 SD (Equation 1, Figure 2).For each parameter, the z-scores for the standard deviation of the load across conditions were combined into a variability score (Equation 3, Figure 2).Abbreviations: z-score for the standard deviation of the sediment load by condition [z SD (Sed) c ], z-score for the standard deviation of the total nitrogen load by condition [z SD (TN) c ], z-score for the standard deviation of the total phosphorus load by condition [z SD (TP) c ], where condition c = dry, normal, or wet.
Limits on point sources have been recommended for specific waterbodies in previous basin-wide water quality plans, including portions of the Deep River (NC Department of Environment & Natural Resources, 2005), the Cape Fear River (NC Department of Environment & Natural Resources, 2000), and Jordan Lake (The Jordan Lake Nutrient Management Strategy., 2009).Updates to nutrient criteria and implementation of nutrient limits on point sources, especially during low flow periods, could help to improve water quality in the basin under anticipated population growth (U.S. Census Bureau, 2020).

Table 1
Cape Fear River Basin Water Quantity and Quality Model Performance for the Calibration Period (2010-2018) and the Validation Period (2000-2009) Was Evaluated Against Observations of Daily Streamflow and Monthly Water Quality at Three In-Stream Gages Co-Located Near Lock and Dam #1 SCHAFFER-SMITH ET AL.