Monthly estimates of freshwater streamflow, inorganic nitrogen, and organic nitrogen at the non-tidal monitoring stations nearest to the heads of tide of the three largest tributaries to Chesapeake Bay (Susquehanna, Potomac, and James; Fig. 1a; Table S1 in the Supplement) were used to evaluate the performance of watershed models. Streamflow and nitrogen load estimates are derived from observations that are collected at U.S. Geological Survey (USGS) stream gages (U.S. Geological Survey, 2022) and comprise part of the USGS River Input Monitoring Program in the Chesapeake Bay watershed (Mason and Soroka, 2022). Estimates for the nitrogen species were calculated using a weighted statistical-regression process that accounts for the variability introduced by time, discharge, and season (Hirsch et al., 2010).

Main-stem bay observations collected over the period 1991–2000, accessible via a data repository maintained by the Chesapeake Bay Program (CBP; Olson, 2012; CBP DataHub, 2022), were used to assess estuarine model skill (see Sect. 2.2). Since 1984, numerous water quality data have been collected along the bay's main stem and throughout its tributaries at semi-monthly to monthly intervals as part of the Water Quality Monitoring Program. These data were collected at the surface, above and below the pycnocline, and at the bottom for chemical variables including nitrate and organic nitrogen, and throughout the entire water column at 1–2 m intervals for O2. Twenty CBP stations were selected for model comparison at the surface and bottom (Fig. 1b, Table S2), including those most frequently sampled and those located along the entirety of the bay's main channel, where hypoxia commonly occurs (Officer et al., 1984; Hagy et al., 2004). Estimates of annual hypoxic volume (AHV), defined as the volume of hypoxic water integrated over the year (with units of volume ⋅ time), were taken from the Bever et al. (2013, 2018, 2021) interpolation of O2 measurements at 56 CBP stations.

Model simulations are conducted with ChesROMS-ECB, a fully coupled, three-dimensional, hydrodynamic and estuarine carbon biogeochemistry (ECB) implementation of the Regional Ocean Modeling System (ROMS; Shchepetkin and McWilliams, 2005) developed for Chesapeake Bay (Xu et al., 2011) with 20 terrain-following vertical levels and an average horizontal resolution of approximately 1.8 km in the estuary's main stem (Feng et al., 2015; St-Laurent et al., 2020; Frankel et al., 2022). The following two parameter changes were recently made to improve the representation of modeled oxygen: (1) a decrease of the maximum growth rate of phytoplankton, which, following Irby et al. (2018), preserves the temperature-dependent linear Q10 described in Lomas et al. (2002); and (2) a decrease in the critical bottom shear stress from 0.010 to 0.007 Pa, which increases the resuspension of organic matter and is well within the range of observed shear stresses evaluated by Peterson (1999).

Estimates of watershed streamflow, nitrogen loading, and sediment loading to drive the estuarine model were obtained via two independently developed models of the Chesapeake Bay watershed: the Dynamic Land Ecosystem Model (DLEM; Yang et al., 2015; Yao et al., 2021) and the USEPA Chesapeake Bay Program's regulatory Phase 6 Watershed Model (Phase 6; Chesapeake Bay Program, 2020). Both models were applied to generate comparable reference runs over the average hydrology period of 1991–2000, chosen because it reflects the decade used by the Chesapeake Bay Program to calculate total maximum daily loads (USEPA, 2010) and to assess water quality improvements. Outputs from both watershed models were aggregated into 10 major river input locations (RIM in Fig. 1). Watershed outputs were mapped to estuarine variables as in Frankel et al. (2022), except that a more realistic partitioning of terrestrial organic nitrogen loading into labile and refractory pools was implemented such that the percent refractory organic nitrogen loading increases with streamflow at high flow volumes (Appendix A).

Atmospheric conditions, including temperature and winds, were obtained from the ERA5 reanalysis dataset (C3S, 2017) as in Hinson et al. (2021). Coastal boundary conditions were interpolated to match the nearest physical and nutrient observations, as in previous work (Da et al., 2021). In order to isolate the impacts of climate-driven changes in watershed inputs, atmospheric and coastal boundary conditions were kept the same in all model simulations under realistic 1991–2000 conditions for both reference runs (1991–2000) and all future scenarios (2046–2055).

Watershed and estuarine model skill was evaluated by comparing results from the two reference scenarios to available data (see Sect. 2.1). Nash–Sutcliffe efficiencies (Nash and Sutcliffe, 1970) were used to evaluate watershed model performance in terms of freshwater streamflow and nutrient loadings. Estuarine model skill was evaluated by comparing model outputs matching the spatio-temporal variability of observations at 20 main-stem stations over the 10-year reference period. Average bias (model output minus observed value) and root-mean-squared difference (RMSD) of annual O2, nitrate (NO3), and dissolved organic nitrogen (DON) concentrations were calculated at the surface and bottom of the water column. AHV estimates were calculated by summing the daily volume of model cells containing low-oxygen waters (O2 < 2 mg L-1) and are expressed in units of km3 d following Bever et al. (2013, 2018, 2021). Daily net primary production estimates were integrated over the entire water column and averaged across the bay and month before being compared to average bay-wide estimates from Harding et al. (2002).

Mid-21st century projected changes in atmospheric temperature and precipitation under a high-emissions scenario (RCP8.5; Cubasch et al., 2013) were obtained for multiple ESMs from the fifth Coupled Model Intercomparison Project (CMIP5) that were regionally downscaled via the following two statistical methodologies: Multivariate Adaptive Constructed Analogs (MACA; Abatzoglou and Brown, 2012; downloaded from MACAv2-METDATA, 2018) and bias-correction and spatial disaggregation (BCSD; Wood et al., 2004; downloaded from Reclamation, 2013). Note that downscaled CMIP5 ESM output was used because downscaled CMIP6 ESM output was not yet available when the research began. Downscaled MACA and BCSD projections have an average spatial resolution of approximately 0.042 and 0.125∘, respectively. A delta approach (Prudhomme et al., 2002; Anandhi et al., 2011) was used to estimate the absolute change in atmospheric temperature and the fractional change in precipitation over the Chesapeake Bay watershed. In this delta approach (also commonly referred to as a perturbation method or change-factor method), the difference in a given climate variable (i.e., air temperature or precipitation) is calculated by first subtracting monthly downscaled ESM estimates averaged over a hindcast period (in this case 1981–2010) from average monthly future projections (in this case 2036–2065). The resulting mean annual cycle (with monthly resolution) in the delta (i.e., the absolute change in temperature or the fractional change in precipitation) is then applied to reference atmospheric-forcing inputs (in this case for 1991–2000) to generate future watershed scenarios (in this case for 2046–2055, hereafter referred to as mid-century) and to limit the uncertainty introduced by interannual variability. An additional step to modify precipitation intensity is also included in all climate scenarios following the methodology outlined in Shenk et al. (2021b). Thirty-year averaging periods were used to limit potential biases introduced by multidecadal oscillations.

To reduce the computational load of applying the dozens of available ESMs to our combined watershed–estuarine modeling framework for a full factorial experiment, the Katsavounidis–Kuo–Zhang (KKZ; Katsavounidis et al., 1994) algorithm was applied to select a subset of five ESMs from both downscaled datasets. KKZ is an objective procedure for selecting a subset of members that best span the spread of the full ensemble in a multivariate space. Because changes to hypoxia must be computed after a subset of ESMs is selected, the downscaled results were classified in terms of changes to the two variables most likely to influence hypoxia, namely air temperature from May–October (i.e., the historic hypoxic season in Chesapeake Bay) and precipitation from November–June (corresponding to the highest set of correlation coefficients when regressed against historical AHV estimates; Fig. S1 in the Supplement). The KKZ algorithm first selected an ESM nearest to the center of the cluster of models in the two-parameter space, which is referred to hereafter as the center ESM, before iteratively selecting additional ESMs that were furthest from the center of the distribution and other previously selected ESMs (Fig. 2; Table S3 in the Supplement). The next four selected ESMs are referred to as hot/wet, cool/wet, hot/dry, and cool/dry ESMs to denote whether they are cooler, hotter, wetter, or drier relative to the center ESM. The specific ESMs selected based on MACA and BCSD differ slightly; however, three of the five models are the same (cool/dry, hot/dry, and cool/wet). The selection process incrementally adds members to those previously selected so that the entire ensemble is ordered and a subset of any size can be selected. This method has proven effective at covering the largest range of outcomes using the fewest ESMs in watersheds across the United States in previous research (Ross and Najjar, 2019). This ESM selection process allows for a more robust comparison of the distribution of ESMs from multiple downscaled datasets as opposed to individual ESM comparisons that may privilege one downscaling method over others. However, because inexact matches among ESMs can impact the quantification of relative uncertainty (Sect. 2.5), additional scenarios were simulated as needed for the center and hot/wet ESMs, which were different for the two downscaling techniques (Fig. 2, Table S3). Future change in temperature and precipitation between the two downscaling methods are shown for the center ESM (Fig. 3); changes for the other four ESMs are found in the Supplement (Fig. S2).

Three numerical experiments (sets of simulations) were conducted to evaluate the impacts of climate scenario factors, management conditions, and the use of a subset of ESMs on future AHV projections and uncertainty (Table 1). To isolate climate impacts on AHV from the watershed alone, direct atmospheric and oceanic forcings to the bay were held to be the same as in the reference simulations (see Sect. 2.3) for all experiments. The first experiment (multi-factor) evaluates the relative change in AHV (hereafter defined as ΔAHV) between the 1991–2000 and 2046–2055 time periods due to the following factors: ESM, downscaling method, and watershed model (Table 1, Fig. 4). Atmospheric deltas from 10 downscaled ESMs (five from MACA and five from BCSD) were applied directly to the two watershed models for a total of 20 simulations. A separate Phase 6 climate reference run is used to evaluate the impacts of climate alone by holding land use and nutrient applications constant. This differs slightly from the Phase 6 reference run that simulates realistic and interannually varying nutrient inputs and terrestrial conditions and is compared against observations (Sect. 2.2). Two additional simulations were conducted with Phase 6 to account for the fact that the ESMs selected by the KKZ method were not identical for MACA and BCSD (Table 1, Fig. 2).

The second experiment (management) applied the same deltas used for Phase 6 MACA scenarios in the multi-factor experiment (thereby varying runoff and nutrient loading) but also included the effect of changing environmental management conditions (affecting nutrient inputs to and export from the terrestrial environment) for a total of five additional simulations. These management simulations assume that reduction targets for nutrient and sediment runoff are met in accordance with established management goals (USEPA, 2010). One additional scenario was conducted in which management goals were imposed and climate change was not.

The third experiment (all ESMs) applied all 20 MACA downscaled ESM deltas to the DLEM scenarios without any changes to management conditions, thereby only modifying changes in runoff and nutrient export without intentional nutrient reductions, for a total of 20 additional simulations. Comparing the results of the first (multi-factor) and third (all ESMs) experiments highlights the strengths and limitations of using a subset of ESMs.

To analyze climate impacts on Chesapeake Bay hypoxia, changes in O2 and AHV were compared between the reference runs and the future simulations. Relative O2 impacts introduced by the three climate scenario factors (ESM, downscaling method, and watershed model) were determined by applying an analysis of variance (ANOVA) approach to average ΔAHV estimates for each climate scenario. This method has been previously applied to the quantification of uncertainty sources in climate and hydrological applications (Hawkins and Sutton, 2009; Yip et al., 2011; Bosshard et al., 2013; Ohn et al., 2021). To use this method in this study, an average annual metric is first calculated for an outcome of interest (i.e., change in streamflow, nitrogen loading, or hypoxic volume) within the multi-factor experiment. Then, the relative uncertainty is determined by calculating the sum of squares due to individual effects for each experimental factor (ESM, downscaling method, or watershed model). Following Ohn et al. (2021), the cumulative uncertainty is quantified for successive uncertainties introduced by each factor and their interactions, removing the unexplained interaction term described in Bosshard et al. (2013). The two additional ESM scenarios described previously (Tables 1, S3) were used due to the inexact matches between MACA and BCSD ESMs selected by KKZ. Despite five ESMs being used in combination with only two downscaling methods and two watershed models in this analysis, the approach outlined (Bosshard et al., 2013; Ohn et al., 2021) accounts for this factor imbalance (five vs. two) by repeatedly subsampling combinations of two ESM scenarios from the five available. An example of this methodological approach is described in Appendix B.

Relative frequency histograms and cumulative distributions were used to quantify the overall likelihoods of increasing or decreasing ΔAHV across the entire range of future scenarios. Average changes in the spatial distribution of O2 over the typical hypoxia season (May–September) were compared among all climate scenarios with no changes to management conditions. Results were considered significant if at least 80 % of the model scenarios tested agree on the direction of O2 change in the estuary, as in Tebaldi et al. (2011).

Increasing temperatures and changing precipitation throughout the bay watershed produce different streamflow responses for the two watershed models. On average, Phase 6 climate scenarios increase watershed runoff relative to the reference run by 4 %–6 %, while DLEM climate scenarios decrease average flow by 1 %–4 % (Table 4). The annual flow changes range from -12 % to +15 % among ESM scenarios, with wetter ESMs tending to increase annual watershed streamflow, while drier ESM scenarios generally decrease average watershed runoff, with a lesser impact due to atmospheric warming (Table 4; Fig. 6a). For both watershed models and downscaling methods, the cool/wet ESM produces the greatest increase in annual streamflow. Overall, the greatest variability in changes to streamflow estimates is due to the ESM, as MACA and BCSD scenarios increase or decrease annual streamflow by comparable amounts (Table 4; Fig. 6a).

Chesapeake Bay Phase 6 watershed model climate scenarios increase average annual total nitrogen (TN) by +30 % and +45 % for MACA and BCSD, respectively, but do not substantially change DLEM TN loads (+1 % and -2 % for MACA and BCSD, respectively; Fig. 7). Greater Phase 6 TN loadings are primarily due to extreme values in the cool/wet climate scenarios and are driven by increases in refractory DON (Fig. 7a). While DLEM scenarios show increases in the percentage of inorganic nitrogen and labile organic forms of total nitrogen loads, the contribution of particulate organic nitrogen (PON) decreases, resulting in little to no increase in overall TN loading (Fig. 7a). Phase 6 produces wetter climate scenarios that increase TN loading more than drier scenarios (Table 4; Fig. 6b), with this effect being most pronounced for the cool/wet ESM. Phase 6 also produces the greatest percent changes in the southern rivers (James, York, and Rappahannock), while DLEM produces similar percent changes in all rivers (Fig. 7b). Some Phase 6 climate scenarios substantially increase the average loading change in smaller watersheds like the Rappahannock and York, which increases TN between 77 %–172 % and 32 %–430 %, respectively, and are comparable to the absolute change in Susquehanna TN loading (Fig. 7b). In contrast with the multi-factor experiment results, climate scenarios that include management actions substantially reduce TN loading (-28 %; Fig. 7, Table 4). Like other Phase 6 climate scenarios that do not account for management actions, the proportion of refractory organic nitrogen increases for the climate scenarios with management (+49 %), but in these cases, the average labile inorganic and organic nitrogen loadings also substantially decrease (-40 %).

Climate change impacts on watershed streamflow and nitrogen loading substantially affect estuarine hypoxia, even when, as in this study, direct climate effects on the bay are not considered. On average, the multi-factor climate scenarios decrease average summer bottom O2 in the bay's main stem while also slightly increasing O2 at the surface in some mid-bay areas (Fig. 8). In the northern part of the main stem near the Susquehanna River outfall, model results show consistent decreases in both bottom and surface summer O2 (Fig. 8e, f). Further down the main stem in the mid-bay, surface O2 increases in wet years and experiences almost no change in dry years (Fig. 8b, c). In the same region, bottom O2 declines lessen during wet years and worsen during dry years (Fig. 8e, f). Increasing O2 levels are found in the shallow portions of the major tidal tributaries (i.e., Potomac and James) but are more pronounced in wet years than in dry years (Fig. 8b–c, e–f). Altogether, average summer surface O2 increases by 0.02±0.03 mg L-1 (average change and standard deviation), while bottom O2 decreases by 0.03±0.06 mg L-1.

There are some clear distinctions in the overall changes to future AHV when evaluating all multi-factor experiments. Climate effects on the watershed in DLEM increase AHV more than in Phase 6 (5.6 % vs. 3.1 %, respectively), but the overall standard deviation of DLEM ΔAHV results are greater than those for Phase 6 (Table 5). Similarly, using MACA vs. BCSD results in greater changes in ΔAHV (4.8 % vs. 3.9 %); albeit, this difference due to the choice of downscaling method is less than that due to the choice of watershed model. Depending on the choice of ESM, ΔAHV ranges between +0.9 % (for the cool/dry ESM) to +8.3 % (for the cool/wet ESM), with the center ESM producing intermediate results (+4.4 %). When comparing the impact of a particular ESM, wetter models tend to produce greater ΔAHV than drier scenarios (Fig. 6c), although interannual variability is still large. When climate scenarios are downscaled using different methodologies (either MACA or BCSD), average ΔAHVs have some notable differences; e.g., applying the cool/dry model to Phase 6 produces opposite average changes to hypoxia depending on the downscaling method (Fig. 6c). Considering all possible combinations of scenarios, ESM average annual projected AHV spans a range of 921–939 km3 d for Phase 6 and 1019–1049 km3 d for DLEM, and four out of five of the climate scenarios in the multi-factor experiment project increases in average AHV (Table 4).

When the full distribution of multi-factor scenarios is evaluated, the average and standard deviation of these annual ΔAHV results are estimated to be 37±64 km3 d (4.4 ± 7.4 %; Fig. 9). Wetter ESMs (blues in Fig. 9a) are more likely to increase hypoxia compared to drier ESMs, despite differences in downscaling method or watershed model. The likelihoods of the cool/dry or hot/dry ESMs increasing hypoxia are only 58 % or 50 %, respectively, but these chances are greater than 80 % for the center, hot/wet, and cool/wet ESMs (Fig. 9a). Altogether, the multi-factor experiment results in 72 % of the runs increasing AHV when considering climate change impacts on terrestrial runoff (Fig. 9b). Note, however, that this cannot technically be considered to be a statistical probability, as the KKZ selection process used to generate our sample of climate scenarios is neither random nor independent.

The all-ESMs experiment produces similar results to those obtained when only a subset of five ESMs is used. Specifically, ΔAHV increases by 6.3 ± 3.5 % using only 5 KKZ-selected ESMs and by 9.6 ± 1.7 % when using all 20 ESMs (Fig. 10a, b; model IDs are further defined in Table S3 in the Supplement). The use of 5 KKZ-selected ESMs covers approximately 69 % of the total range of all 20 ESMs (Fig. 10c). Despite more than 15 000 options to choose from when selecting 5 out of 20 ESMs, the subset selected in this work demonstrates an improved ability to outperform a random selection of 5 ESMs (Fig. 10c) and generates a useful range of hypoxia projections.

The results of the management experiment demonstrate the substantial impact of future nutrient reductions on hypoxia, decreasing average ΔAHV by 50 ± 7 % relative to the 1990s (ΔAHV = -438±47 km3 d; Table 4; Fig. 11). Because there is a linear relationship between ΔAHV computed with Phase 6 MACA scenarios including management actions (ΔAHVmgmt) and those without (ΔAHV = 0.56 ⋅ ΔAHVmgmt - 262; R2 = 0.59; Fig. S4), ΔAHVmgmt can be estimated for any scenario by applying this linear model to the non-management-scenario distribution. In effect, this linear relationship demonstrates a similar magnitude of relative nutrient export to and consequent hypoxia within the estuary. The result is a decrease of approximately 417 ± 67 km3 d among all scenarios, within the range of the management scenario subset obtained here by applying only MACA downscaled ESMs to Phase 6. As expected, hypoxia increases in the management experiment when climate impacts are also included relative to the reference management scenario, specifically by 17.1±34.8 km3 d or 3.8 ± 7.8 % (Table 4; Fig. 6c).

Applying an ANOVA approach (Ohn et al., 2021) to watershed streamflow, nutrient loadings, and ΔAHV within the multi-factor experiment reveals that the relative uncertainties introduced by the choice of ESM, downscaling method, and watershed model vary substantially (Fig. 12). The choice of ESM is the dominant factor affecting changes to watershed streamflow and nutrient loadings (Fig. 12a–c) and comprises 59 %–74 % of the total uncertainty. The choice of watershed model is the next largest source of uncertainty, making up 17 %–34 % of the total variance in watershed changes, while the downscaling method only contributes 3 %–14 %. Uncertainty in projected organic nitrogen loadings is particularly affected by the choice of watershed model, overwhelming the variability introduced by downscaling method and strongly affecting estimates of total nitrogen change. Unlike changes to watershed flow and loadings, the uncertainty of projected changes to hypoxia is much more evenly distributed among the three scenario factors, specifically 40 %, 25 %, and 35 % for the ESM, downscaling method, and watershed model, respectively (Fig. 12d).

Projected changes in watershed streamflow and nutrient delivery to Chesapeake Bay produce modest increases in estuarine hypoxia with medium confidence (Mastrandrea et al., 2010). Hypoxic volume has a high degree of interannual variability, and future hypoxia estimates are highly sensitive to the choice of ESM, downscaling method, and watershed model (Fig. 6c). Although certain factors (particularly ESM and greenhouse gas emission scenarios; Meier et al., 2021) have previously been extensively evaluated in coastal systems with regards to future hypoxia, the results presented here also demonstrate the importance of terrestrial forcings on estuarine oxygen levels.

In this study, future changes in watershed streamflow, nitrogen loadings, and estuarine hypoxia are found to be highly dependent on the selection of a specific ESM (Fig. 12), with this factor comprising the majority of the total uncertainty in watershed runoff and the greatest fraction of total uncertainty for O2 levels. When only the effect of ESM choice is considered (and downscaling and hydrological model options are not; Fig. 10), the average projected change in AHV using only three ESMs (often chosen to represent cool, median, and hot scenarios) has a greater standard error than the selection of five ESMs using KKZ in this study. Directly comparing results from the experiment that compared five ESMs, two downscaling methods, and two watershed models (multi-factor) versus that which only considered the impact of multiple ESMs (all ESMs) shows a substantial overlap in the range of projected ΔAHV. In addition, multiple ESMs downscaled with a single methodology and applied to one hydrological model produced meaningfully different estimates of ΔAHV than a more balanced approach (Fig. 11).

Inter-model variability among ESMs appears to contribute most substantially to differences in bay watershed inputs, but the choice of downscaling methodology can also affect these projections. The BCSD (Wood et al., 2004) and MACA (Abatzoglou and Brown, 2012) downscaling methodologies used here employ different approaches to reduce historical ESM biases, impacting the variability of spatio-temporal watershed hydrologic and water quality responses. The ability to statistically downscale ESMs accurately depends on the spatially coarser ESM's ability to simulate synoptic-scale (∼ 1000 km) patterns and may still underestimate the distributional tails of changes to temperature and precipitation. This increases the importance of properly selecting a subset of ESMs (Abatzoglou and Brown, 2012).

Watershed model variability is caused by differences in the representation of processes that affect streamflow, those controlling the fate and transport of nutrients from land and in rivers, and lag times of groundwater transport. The two watershed models used here project substantially different results in watershed streamflow and nitrogen delivery, even when the same changes to meteorological forcings are applied (Fig. 6). DLEM projects no change or decreases in streamflow for nearly all scenarios as opposed to greater average increases in streamflow for Phase 6 scenarios (Fig. 6a), likely driven by differences in the representation of evapotranspiration. Explicit soil biogeochemical processes within DLEM increase nitrification rates in warmer-climate scenarios, producing higher nitrate loadings than Phase 6 despite comparable streamflow changes (Fig. 6b). The greater total nitrogen loadings produced by Phase 6 are largely a consequence of its parameterizations for erosion and refractory nitrogen bound to sediment. Increases in bioavailable nitrate loadings, unlike refractory organic nitrogen that comprises the majority of DON loadings, produce greater levels of primary production and remineralization within the estuary. This largely explains the discrepancy between watershed model hypoxia estimates (Table 5).

Our findings demonstrate the importance of considering differences among these three factors (ESM, downscaling, and watershed model) that may contribute to a wider range of target water quality variables and living-resource responses in coastal marine ecosystems like Chesapeake Bay that are highly influenced by watershed processes. Hydrological model assumptions can have potentially significant impacts on estuarine hypoxia. For example, the relatively high organic nitrogen loadings in Phase 6 compared to DLEM's comparatively modest exports under the same future scenarios result in different levels of annual hypoxia. While dramatic increases in organic nitrogen loadings within bay tributaries are mostly limited to cool/wet Phase 6 scenarios, there is precedent for catastrophic erosion within the bay watershed driven by extreme precipitation events (Springer et al., 2001). The relative uncertainty introduced by individual factors is also not necessarily equivalent for streamflow, nitrogen loadings, and AHV (Fig. 12). The complex connections between terrestrial runoff and biogeochemical changes in the marine environment may expand further when higher-order trophic-level species are considered and even more so when direct atmospheric impacts on the bay are also included. It is unlikely that general conclusions regarding the relative impacts of different factors can be drawn for a marine ecosystem when only uncertainties in watershed streamflow and nutrient loadings are considered. Had our results only accounted for the impacts of these factors on watershed changes and not estuarine oxygen levels, the role of downscaling could be incorrectly assumed to contribute negligible variability to hypoxic volume (Fig. 12). It is the complex interactions of nitrogen species transformations within this estuarine model that are responsible for this somewhat unexpectedly large contribution of downscaling-method uncertainty that is less prominent in watershed changes.

Despite the relatively small magnitude of Chesapeake Bay watershed climate impacts on estuarine hypoxia compared to previous evaluations of other climate impacts, like atmospheric warming over the bay (Irby et al., 2018; Ni et al., 2019; Tian et al., 2021), the relative contributions of ESM and downscaling effects to the total uncertainty are large and are also likely to expand the range of outcomes for other climate sensitivity studies in this region. This suggests that, when attempting to determine a likely range of ecosystem outcomes, selecting additional downscaling techniques and hydrological model responses should be considered in addition to the more common practice of only selecting multiple ESMs.

The climate scenario projections evaluated in this study are in near-complete agreement that the Chesapeake Bay watershed will be warmer and will experience greater levels of precipitation by the mid-century, yet these results are not as straightforward to interpret, as they relate to changes in streamflow, nutrient loads, and estuarine hypoxia. Climate impacts on extreme river flows are currently evident at global scales (Gudmundsson et al., 2021), and projected increases in precipitation that could shape such events are aligned with estimates for this region derived from observational (Yang et al., 2021) and modeling (Huang et al., 2021) studies, as well as for other regions at similar latitudes (Bevacqua et al., 2021; Madakumbura et al., 2021). However, differences exist in the spatial distribution and timing of these precipitation increases, as well as in the temperature-affected rates of evapotranspiration. As a result, these estimates produce varied projections for future freshwater streamflow. These complex interactions make it difficult to directly predict future streamflow from projected precipitation changes and even more difficult to relate these to changes in nutrient loading. For example, in this study, half of the climate scenarios produce increasing streamflow on an annual basis, yet more than 75 % of these scenarios increase total nitrogen loading. Differences in the representation of soil and riverine nitrogen processes between watershed models also result in inconsistent simulated responses of nitrogen export to similar precipitation rates. Disparate export of nitrogen species (i.e., nitrate and organic nitrogen) between watershed models also directly affects future nutrient load projections. These hydrological model differences are evidenced by DLEM's higher NO3 outputs that offset lower organic nitrogen loadings (Fig. 7a).

Our analysis quantifies changes in hypoxia due to mid-century climate change impacts on the watershed and provides an estimate of the relative uncertainty in these estimates. Our experimental findings suggest that, in the absence of management actions, mid-century climate impacts on the Chesapeake Bay watershed will increase hypoxia, specifically annual hypoxic volume (AHV), by an average of 4 ± 7 %. This estimate is in good agreement with prior studies that examined the impacts of watershed actions alone. Irby et al. (2018) applied a sensitivity approach and projected increases in AHV of 5 %, while Wang et al. (2017) showed increases in annual anoxic volume of 9.7 %, nearly equivalent to the increase of 10 ± 16.5 % found here (Table 6). Results from this study also project that changes to bay O2 levels will vary spatially. Average bottom main-stem O2 levels from May–September are expected to decrease most in the southern half of the bay (south of 38.5∘ N), particularly in climatologically dry years (Fig. 8).

Importantly, the projected changes presented here only account for the effects of climate change on watershed response in isolation and do not include the additional direct impacts of the atmosphere and ocean. These additional changes have been estimated in other previous studies of 21st century impacts relative to observed conditions (Table 6). While numerous differing metrics have been reported for many of these studies, including shifting dissolved-oxygen concentrations and water quality regulatory criteria, this work can be compared against previous results by examining changes to annual hypoxic and anoxic volumes. The majority of these studies (Table 6) apply idealized changes to climate forcings and generally project increases in hypoxic conditions. Increases in mid-21st century annual hypoxic volume due to watershed forcings (+5 % and +4.4 ± 7.4 %) are smaller than the average impacts of increasing temperatures alone (+13 %), while the results of changing sea level are more mixed (Table 6). However, the variability in hypoxia due to watershed changes is likely greatest among these factors and may substantially modify the negative effects of warming on dissolved-oxygen concentrations. Our results and their uncertainties generally encompass the range of future hypoxia estimates found in previous research that has studied multiple climate impacts in isolation and in various combinations. Future work that accounts for the sources of uncertainty explored here by applying realistic climate change projections while also standardizing a metric for model results, like annual hypoxic volume, will help to narrow and better quantify definitive trends due to multiple factors that influence bay dissolved oxygen.

Our findings are focused on Chesapeake Bay hypoxia, but some lessons can also be drawn from other coastal ecosystems where changes in watershed streamflow and nutrient loadings are also projected. In the Baltic Sea, Meier et al. (2011b) reported that hypoxia was very likely to increase regardless of ESM or climate scenario, assuming targeted reductions in accordance with the Baltic Sea Action Plan (decrease of nitrogen loads by 23 ± 5 %) were not met. Extensive studies of projected oxygen change in the Baltic Sea have repeatedly demonstrated that climate impacts are likely to increase hypoxic area (BACC II Author Team, 2015, and references therein), but more recent reports (Saraiva et al., 2019a; Wåhlström et al., 2020; Meier et al., 2021, 2022) have reaffirmed that nutrient reductions in accordance with the Baltic Sea Plan are also highly likely to mitigate a substantial amount of those hypoxia increases. Repeated investigations into the impact of increased streamflow and higher temperatures in the Gulf of Mexico demonstrate a likely expansion of hypoxic area (Justić et al., 1996; Lehrter et al., 2017; Laurent et al., 2018) and that additional nutrient reductions would be required to mitigate these impacts (Justić et al., 2003). Finally, Whitney and Vlahos (2021) demonstrated a considerable erosion in oxygen gains in Long Island Sound due to nutrient reductions in the presence of climate effects, reducing projected mid-century improvements by 14 %, similar to the 9 % increase in hypoxic volume reported by Irby et al. (2018) for O2 levels <2 mg L-1. Although these studies include direct climate change impacts on coastal water bodies, most support the findings here, demonstrating that increases in streamflow and associated nutrient loadings are likely to increase Chesapeake Bay hypoxia. Overall, climate impacts on land have the potential to profoundly modify biogeochemical interactions in the coastal zone and to limit the efficacy of nutrient reductions.

Projections of changes to watershed streamflow and nutrient delivery can better inform regional environmental managers tasked with managing interactions among nutrient reduction strategies, climate change, and coastal hypoxia (Hood et al., 2021; BACC II Author Team, 2015; Fennel and Laurent, 2018). The Chesapeake Bay results provided in this analysis demonstrate that the management actions mandated to improve water quality (USEPA, 2010) will decrease hypoxia by roughly 50 %, approximately an order of magnitude more than projected increases due only to watershed climate change (Fig. 11). Therefore, nutrient reduction strategies are very likely to remain effective at reducing watershed nutrient loading and its contribution to eutrophication and hypoxia over a range of possible ESM scenarios (Mastrandrea et al., 2010). Should all management actions be implemented as outlined in the USEPA's Total Maximum Daily Load (USEPA, 2010), it is very likely that future climate impacts on bay watershed runoff will worsen bay hypoxia by a far smaller amount relative to 1990s reference conditions. These findings are consistent with those of Irby et al. (2018), who also examined the impacts of watershed climate on Chesapeake Bay hypoxia for the mid-21st century. When evaluating the effects of watershed climate impacts and management actions together, Irby et al. (2018) estimated an average AHV increase of 12.8 km3 d, which is well within the range of 17.1 ± 34.8 km3 d reported here (Table 6). Additionally, the combined impact of all climate stressors reported by Irby et al. (2018), i.e., atmosphere, ocean, and watershed, increased average AHV by 24.5 km3 d, which is also within the range of the results reported here. Because climate change impacts are likely to increase total nitrogen loads, implementing nutrient reductions that do not account for the detrimental effects of climate change will reduce the likelihood of attaining water quality targets. Further quantifying a range of future estimates of watershed streamflow and nitrogen loading using regional models is critical to understanding the possibilities and limitations of mitigating negative climate impacts via nutrient reductions.

Recent findings support the hypothesis that nutrient reductions will improve water quality despite projected climate impacts in both freshwater systems (Wade et al., 2022) and other coastal marine systems (Whitney and Vlahos, 2021; Saraiva et al., 2019a; Bartosova et al., 2019; Wåhlström et al., 2020; Pihlainen et al., 2020; Meier et al., 2021; Große et al., 2020; Jarvis et al., 2022). In Chesapeake Bay, reduced nutrient loading (Zhang et al., 2018; Murphy et al., 2022) has already helped mitigate growing climate change pressures (Frankel et al., 2022) despite rapidly increasing bay temperatures over the past 30 years (Hinson et al., 2021). Like these prior studies, our findings confirm that management actions will likely produce even greater benefits to O2 in coastal zones strongly affected by terrestrial runoff. While direct effects (e.g., air temperature) are expected to increase hypoxia more than watershed changes in Chesapeake Bay (Irby et al., 2018; Ni et al., 2019), the comparatively greater impacts of management actions reported here are also likely to substantially reduce the overall risk from a multitude of co-occurring climatic stressors.

Despite the plainly evident finding of nutrient reduction strategies improving water quality and counteracting negative climate change watershed impacts, a number of important caveats should temper this conclusion. First, the subset of scenarios that include management actions is limited to a set of five ESMs statistically downscaled with a single methodology and applied to one watershed model. As demonstrated in this work, this assumption may oversimplify the complex relationship between climate forcings and watershed model simulations, especially given that DLEM scenarios produce more change in nitrate and consequently more hypoxia than Phase 6 scenarios. Management actions implemented in Phase 6 nutrient reduction scenarios represent a multitude of possible methods to reduce point and nonpoint source pollution that are assumed to be fully implemented with a high operational efficacy by the mid-century, but the true performance of the best management practices operating under future hydroclimatic stressors remains largely unresolved (Hanson et al., 2022). Additionally, the importance of legacy nitrogen inputs to the bay may grow over time (Ator and Denver, 2015; Chang et al., 2021) and can only be properly accounted for via a long-term transient simulation that accounts for changing groundwater conditions.

A key strength of the delta method applied here is its ability to remove the influence of interannual variability, which is known to strongly influence hypoxia in Chesapeake Bay (Bever et al., 2013). However, the delta method is unable to account for the impacts of unanticipated extreme events or changing patterns of precipitation intensity, duration, and frequency that produce dramatic responses in sediment washoff, scour, and consequent watershed organic nitrogen export. Air temperature and precipitation were the only watershed model input variables adjusted in this analysis, allowing for a more equivalent comparison between downscaling approaches. Future representations of watershed change may also better account for changes in runoff through the inclusion of factors like ESM-estimated relative humidity that can help avoid possible unreasonable amplification of potential evapotranspiration that would decrease tributary streamflow (Milly and Dunne, 2011) and associated nutrient loads.

Although main-stem bay oxygen levels are the focus of this study, watershed impacts are also likely to influence water quality in smaller-scale tributaries. Differences in Chesapeake Bay temperatures introduced by the ESM and downscaling method have also been investigated by Muhling et al. (2018) and contribute to biogeochemical variability via the direct impacts of atmospheric temperature on bay warming. Incorporating different facets of these relative uncertainties into projections of coastal change has also been demonstrated to affect ecological outcomes like those surrounding fisheries (Reum et al., 2020; Bossier et al., 2021). Thus, the impacts of these uncertainties are also very likely to affect socio-economic systems tied to coastal resources. The analytical method applied here is well established within climatic and terrestrial settings, so the relative dearth of coastal applications (excluding Meier et al., 2021) may be more related to a consequence of computational demand or a greater focus on uncertain parameterizations of marine biogeochemical processes (Jarvis et al., 2022) that also play a large role in potential future hypoxia outcomes.